Old behavior would retry set and delete even if there's an error. This
can lead to the client returning an error for deleting twice, instead
of returning an error for an interdeterminate state.
Fixes#5832
Before
2016-07-01 14:57:50.927170 I | api: enabled capabilities for version 3.0.0
After
2016-07-01 14:57:50.927170 I | api: enabled capabilities for version 3.0
And 'maybeSnapshotAbort' does not 'unset'
the pendingSnapshot. 'resetState', which is called after this
metho, is the one that unsets pendingSnapshot. So this changes
the method name.
Current build script outputs its name to stdout because of its
checking argv[0].
$ ./build
./build
The line is a little bit mysterious so this commit removes it.
This commit adds a new flag --detail to etcdctl user get command. The
flag enables printing the detailed permission information of the user
like below example:
$ ETCDCTL_API=3 bin/etcdctl --user root:p user get u1
User: u1
Roles: r1 r2
$ ETCDCTL_API=3 bin/etcdctl --user root:p user get u1 --detail
User: u1
Role r1
KV Read:
[k1, k5)
KV Write:
[k1, k5)
Role r2
KV Read:
a
b
[k8, k9)
KV Write:
a
b
[k8, k9)
Currently the user can't list only the keys in a prefix search. In
order to support such operations the filtering will be done on the
server side to reduce the encoding and network transfer costs.
These legacy flags support are here only because we do not want
CoreOS updates to break people.
Now people will be aware of that they switch to etcd3. Do not need
to support 0.x flags any more.
This commit expands RPCs for getting user and role and support list up
all users and roles. etcdctl v3 is now support getting all users and
roles with the newly added option --all e.g. etcdctl user get --all
Currently auth tokens are generated in the replicated state machine
layer randomly. It means one auth token generated in node A cannot be
used for node B. It is problematic for load balancing and fail
over. This commit moves the token generation logic from the state
machine to API layer (before raft) and let all nodes share a single
token.
Log index of Raft is also added to a token for ensuring uniqueness of
the token and detecting activation of the token in the cluster (some
nodes can receive the token before generating and installing the token
in its state machine).
This commit also lets authStore have simple token related things. It
is required because of unit test. The test requires cleaning of the
state of the simple token things after one test (succeeding test can
create duplicated token and it causes panic).
This commit adds a feature for granting and revoking range of keys,
not a single key.
Example:
$ ETCDCTL_API=3 bin/etcdctl role grant r1 readwrite k1 k3
Role r1 updated
$ ETCDCTL_API=3 bin/etcdctl role get r1
Role r1
KV Read:
[a, b)
[k1, k3)
[k2, k4)
KV Write:
[a, b)
[k1, k3)
[k2, k4)
$ ETCDCTL_API=3 bin/etcdctl --user u1:p get k1 k4
k1
v1
$ ETCDCTL_API=3 bin/etcdctl --user u1:p get k1 k5
Error: etcdserver: permission denied
Fix https://github.com/coreos/etcd/issues/5598.
race conditions were detected in iterating the test table
because the go func closure doesn't receive the 'puts' index
in the argument. This can cause the test to run wrong put
operations.
Currently the auth mechanism doesn't support permissions of range
request. It just checks exact matching of key names even for range
queries. This commit adds a mechanism for setting permission to range
queries. Range queries are allowed if a range of the query is [begin1,
end1) and the user has a permission of reading [begin2, range2) and
[begin1, end2) is a subset of [begin2, range2). Range delete requests
will follow the same rule.
Release guide steps to artifacts signing defaults to binary
signatures, while producing .asc files.
This commit changes to armored signatures, also matching appc
requirements.
Fixes#5594
Tests need to disconnect the network connection for the client to check
reconnection paths but closing a grpc connection closes the logical connection.
To disconnect the client, instead have a bridge between the server and
the client which can monitor and reset connections.
Fix https://github.com/coreos/etcd/issues/5573.
Currently stresser starts at the same time as cluster start.
If the stresser got launched too fast/early, all stressers
exit from the error 'etcdserver: not capable', which
means the cluster is not ready yet. This adds additional
error checking, so stresser can retry.
Code is only there to handle an edge case where the tail wasn't preallocated
already (e.g., via old etcd version or a crash). It also triggers tmpfs
corruption, so remove it.
'MsgBeat' is an internal type to signal the leader, not the message type
that gets sent to its followers. 'MsgHeartbeat' is the type sent to followers.
Causes async reconnect to block while the client is dialing.
This was also causing problems with the Close error message, so
now Close() will return the last dial error (if any) instead of
clearing it out with a cancel().
Fixes#5416
This commit implements RoleGet() RPC of etcdserver and adds a new
subcommand "role get" to etcdctl v3. It will list up permissions that
are granted to a given role.
$ ETCDCTL_API=3 bin/etcdctl role get r1
Role r1
KV Read:
b
d
KV Write:
a
c
d
In the integration test, we sometimes stop/restart an etcd server.
Now our client has internal connection monitoring logic that might
set conn to nil when there is a connection failure and the redial
also fails.
Chaning randClient to always return a client with active connection
to make integration test reliable.
This commit adds a new subcommand "user get" to etcdctl v3. It will
list up roles that are granted to a given user.
Example:
$ ETCDCTL_API=3 bin/etcdctl user get u1
User: u1
Roles: r1 r2 r3
This commit also modifies the layout of InternalRaftRequest for
frequent update of auth related members.
Current permission checking mechanism doesn't return its error code
well. The internal error (code = 13) is returned to client and the
retry mechanism doesn't work well. This commit fixes the problem.
This commit lets etcdserver check permission during its log applying
phase. With this change, permission checking of operations is
supported.
Currently, put and range are supported. In addition, multi key
permission check of range isn't supported yet.
Current etcdserver tries to return result.resp even if result.err is
not nil. A situation of result.resp == nil and result.err != nil can
happen and it results an error like below:
18:49:57 etcd1 | interface conversion: proto.Message is nil, not *etcdserverpb.PutResponse
This commit lets the functions return result.err if it is not nil.
Makes w.cur into w.minrev, the minimum revision for the next update, and
retries cancelation if the watcher isn't found (because it's being processed
by moveVictims).
Fixes: #5459
When using refresh, etcd store v2 watch is broken. Although with refresh
store should not trigger current watchers, it should still add events into
the watchhub to make a complete history. Current store fails to add the event
into the watchhub, which causes issues.
Our v3 benchmark shows etcd v3 can do 40k writes per second.
1k throughput number is for etcd v2. Also adds YouTube's doorman
to example project lists.
Instead of holding the store lock while doing a lot of work like when syncung
unsynced watchers, the work from a blocked synced notify can be reused and
dispatched without holding the store lock for long.
Migrate command accepts a datadir and an optional user-provided
transformer function that transform v2 keys to v2 keys.
Migrate command then builds a v3 backend state based on the existing
v2 keys and the output of the transformer function.
doc: added missing (http)s to tls setup guide
This fixes a minor documentation omission, where the 1st initial-advertise-peer-url for tls setup appears to be http.
fixes documentation
Current genproto.sh doesn't pass -u option to go get. It is
problematic because the script depends on a specific version of
gogoproto. Actually it causes build error if a repository already have
an old version of gogoproto that doesn't have a specified commit
($SHA). This commit lets the script pass -u to go get for avoid the
error.
This commit adds a functionality of attaching an auth token to gRPC
connection as a per RPC credential.
For doing this, this commit lets clientv3.Client.Dial() create a
dedicated gRPC connection for doing authentication. With the dedicated
connection, the client calls Authenticate() RPC and obtain its
token. The token is attached to the main gRPC connection with
grpc.WithPerRPCCredentials().
This commit also adds a new option --username to etcdctl (v3). With
this option, etcdctl attaches its auth token to the main gRPC
connection (currently it is not used at all).
Client was reconnecting after establishing connections because the lease
and watch APIs were thrashing. Instead, wait a little before accepting
new reconnect requests.
There was a bug in protodoc.
This changes git SHA to use the latest protodoc.
And make the letter casing consistent with original
Protocol Buffer. Go capitalizes the member variables,
but the protocol buffer documentation should be same as
original proto files.
If the WAL is stored in a separate directory then the backup command
would need a --wal-dir option to pick the path to the WAL directory.
The user might also want to store the backup of data and wal separately
for which --backup-wal-dir option is provided.
Added a new command line option to etcd server to read in a YAML
based configuration file. I've also added an example configuration
file with comments and a set of test cases.
We recently changed the randomized election timeout from (et, 2*et-1] tp
[et, 2*et-2], where et is user set election timeout.
So 2*et might trigger two elections instead of one. We need to fix the test
code accordingly.
Thanks for Tikv guys for finding this issue. We probably need to randomize
etcd/raft test more.
The _total suffix is by convention for counters,
don't use it on a gauge. Clarify help string.
Tweak metric name so it'll sort with related metrics,
and be a little more understandable.
Remove open file descriptor metric, as Prometheus client_golang
provides that out of the box as process_open_fds which is also
more up to date. Both only support Linux, so there's no loss of
platform support.
Fixes#5229
Setup a travis test matrix on a new variable 'TARGET', which specifies the CI
target. Update the script section with a conditional that runs the needed
commands for each target.
Also, set go_import_path to make cloned repos work, enable the trusty VM, and
enable verbose builds when testing.
Signed-off-by: Geoff Levand <geoff@infradead.org>
Fix https://github.com/coreos/etcd/issues/5231.
The issue shows that slow CI can take more than 200ms
for purging. This increase the loop iteration to wait
up to 300ms in case the disk is being slow.
If the grpc connection is restored between an rpc network failure
and trying to reestablish the connection, the connection retry would
end up resetting good connections if many operations were
in-flight at the time of network failure.
If no leases are being kept alive, a connection reset would leak
the send routine since it would only test the stream when sending
keep alives.
Fixes#5200
With randomized keys, if the connection goes down, but the session remains,
the client would need complicated recovery logic to avoid deadlock.
Instead, bind the session's lease id to the lock entry; if a session tries
to reacquire the lock it will reassume its old place in the wait list.
KeepAliveOnce should have a per call ctx. Now we have a per
API ctx, but we might do rpc calls mutiple times in a for loop.
To avoid unnecessary routine leak, use per call ctx.
If switchRemoteAndStream fails, the whole lease API fails since
the internal routine exits. We should only fail the whole API when
there is a fatal error. For example, we should fail if we fail to
connection to all the endpoints user provided.
If we connect to an endpoint, but fail to create a stream, we should
retry instead of returning error to fail the entire API.
This commit implements Authenticate() API of the auth package. It does
authentication based on its authUsers bucket and generate a token for
succeeding RPCs.
Otherwise parsing get/txn output with json is somewhat complicated
because in some cases there's a json message and sometimes not.
Likewise, a get on an absent key has to return the current revision for
some algorithms to work.
A call file.Sync on OSX doesn't guarantee actual persistence on
physical drive media as the data can be cached in physical drive's
buffers. Hence calls to file.Sync need to be replaced with
fcntl(F_FULLFSYNC).
1. change file names
2. now if sub-command errors, the test will receive all
lines from stdout and stderr.
Expected output:
```
read /dev/ptmx: input/output error (expected key2, got ["key1\r\n" "val1\r\n" ""])
```
3. change how we check GRPC timeout (only bypass timeout error when we give 0
timeout)
- add another layer of abstraction in clientv3 for user, not expose internal storagepb ones
- provide commonly used routines IsCreate(), IsModify() on event
progressReportIntervalMilliseconds (old
ProgressReportIntervalMilliseconds) is accessed by multiple goroutines
and it is reported as race.
For avoiding this report, this commit wraps the variable with
functions. They access the variable with atomic operations so the race
won't be reported.
On ReadAll, WAL seeks to the end of the last record in the tail. If the tail did not
end with preallocated space, the decoder would report 0 as the last offset and begin
writing at offset 0 of the tail.
Fixes#4903
When compactKV request is halted before final acknowledgement,
it used to just continue on the next endpoint. But there could be
a case than compactKV is requested twice, and the first one is already
replicated and applied by the time the second request was to be
applied (returning compact revision error). This skips the case
by parsing the error message.
This commit adds a functionality of user deletion. It can be invoked
with the new user delete command.
Example usage:
$ ETCDCTL_API=3 etcdctl user delete usr1
Alarms are events that nodes can use to relay health information to
the rest of the cluster. A node may Activate an alarm and that alarm
will stay set until Deactivated.
This commit adds a new subcommand "user add" to etcdctlv3. With the
command users can create a user for the authentication.
Example of usage:
$ etcdctlv3 user add user1
Password of user1:
Type password of user1 again for confirmation:
GRPC and v2 client address share the same host(port)
but GRPC does not work with schema specified. This fixes
it by adding another member for GRPC without schema, as
we had before.
endpoint-health checks endpoint.
It can generate 3 outputs:
1. cannot connect to the member through endpoint
2. connected to the member, but member failed to commit any proposals
3. connected to the member, and member committed a proposal
This commit adds a new method Recovery() to auth.AuthStore for
recoverying auth state from backend during apply snapshot. It follows
a manner of the lessor.
File lock interface was more verbose than it needed to be while
simultaneously making it difficult to support systems (e.g., Windows)
that only permit locked writes on a single fd holding the lock.
leader/follower information is very helpful for debugging. User must get this information through http endpoint
before this patch, which is inconvenient.
The new leader needs to refresh with an extened TTL to gracefully handle
the potential concurrent leader issue. Clients might still send keep alive
to old leader until the old leader itself gives up leadership at most after
an election timeout.
This commit adds a new package auth. Its role is persisting auth
related metadata. This commit also connects its main interface
AuthStore and v3 server.
Add a missing blank line to one of the fenced code blocks in the
"Working with roles" section to conform with the formatting best
practice recommended in the GitHub guide "Creating and highlighting code
blocks"[1]:
> We recommend placing a blank line before and after code blocks to make
> the raw formatting easier to read.
The missing blank line prior to the code block causes incorrect HTML
rendering of this section within the CoreOS etcd Authentication
Guide[2]. This commit fixes the problem, but the underlying issue seems
to be a quirk of the markdown render used for the CoreOS documentation
pages, as the same markdown source does not exhibit the issue when
viewed on GitHub[3] or viewed through the python markdown module.
[1] https://help.github.com/articles/creating-and-highlighting-code-blocks/
[2] https://coreos.com/etcd/docs/2.2.5/authentication.html
[3] https://github.com/coreos/etcd/blob/v2.2.5/Documentation/authentication.md
Adding a new option to the mk command to create in-order keys
under a given directory, identified by <key> argument. Optionally
TTL can also be set on the newly created in-order key.
Fix https://github.com/coreos/etcd/issues/4730.
Previously we put keys async and there might be a race when
the watch triggers before the put receives the response. When that
happens, put might fails to get the response since we shutdown the server
when watch triggers.
Add a note about 2379/2380 vs 4001/7001. I believe this is as
far as we can go just yet, because we have not fully deprecated
the old ports, and still want to configure them in some instances.
Related #3480.
This makes the test endpoints same as we have in
goreman proc file, and adds 'IsProgressNotify' method
and the 'WatchProgressNotify' code example to godoc.
Lock would get the prior key on retry using WithRev(myRev - 1) instead of
using the latest revision; the Watch() would return immediately and Lock
devolves into polling.
Do() makes it possible to pass Ops around and apply them later.
Txn().Then(op).Commit() isn't enough because it will wrap the op
in a txn. Likewise, rewriting single op txns into single op rpc's
precludes deliberately submitting a single op transaction.
This sets Created and Cancel true in pb.WatchResponse
when it has received wrong start revision instead of
panic. So that clientv3 can set 'Canceled' in WatchResponse
as well.
Fix https://github.com/coreos/etcd/issues/4610.
If a current watcher didn't receive any events, a reconnect cycle would
advance its revision to the store's current revision. Instead, reconnect
using the watcher's creation header revision if the watcher hasn't received
any events.
Fixes#4502
Standard log package by default only prints out the second-scale
so the 3rd party log feeder mixes the order of the events, which makes
the debugging hard. This replaces it with capnslog and make them consistent
with all other etcd log formats.
Improved the overall performance more than 20% under heavyload
with little latency impact
heavy load
```
Requests/sec: ~23200 vs Requests/sec: ~31500
Latency distribution:
10% in 0.0883 secs.
25% in 0.1022 secs.
50% in 0.1207 secs.
75% in 0.1460 secs.
90% in 0.1647 secs.
95% in 0.1783 secs.
99% in 0.2223 secs.
vs
Latency distribution:
10% in 0.1119 secs.
25% in 0.1272 secs.
50% in 0.1469 secs.
75% in 0.1626 secs.
90% in 0.1765 secs.
95% in 0.1863 secs.
99% in 0.2276 secs.
```
Similar on light load too.
I manually print out the command outputs when the issue
was reproduced, and checked they are matching when compared as
Go string literals (UTF-8), but not when compared with regex.
Fixes https://github.com/coreos/etcd/issues/4480.
Fixes https://github.com/coreos/etcd/issues/4512.
When cluster fails before creation of log or data directory
the file does not exist and cannot be renamed. This skips such
error because there's no need to store empty logs in failure_archive.
The ConfChange fix crashes WAL replay because it assumed the node
always exists. Additionally, restart on a single node cluster
would deadlock because it had the wrong raft HardState.
To fix, replay path now goes through the regular raft loop.
Fixes#4474
when closed errors will be one of:
```
grpc.ErrorDesc(err) == context.Canceled.Error() ||
grpc.ErrorDesc(err) == context.DeadlineExceeded.Error() ||
grpc.ErrorDesc(err) == "transport is closing" ||
grpc.ErrorDesc(err) == "grpc: the client connection is closing"
```
Currently gRPC connection just gets recreated
for every Stress call. When Stress ends or gets
canceled, gRPC connection must also be closed.
For https://github.com/coreos/etcd/issues/4464.
Hash method returns either (nil, err) or (Hash, nil).
The current error checking is wrong. It only needs to check
the error is either nil or non-nil.
This causes panic in https://github.com/coreos/etcd/issues/4463
by allowing the case when resp is nil, but err is not nil.
./build takes long time. On my Core i5 box, it requires almost 25
seconds. Without -a flag, it takes almost 15 seconds. Therefore this
commit reduces the flag in default. ./test activates -a via a new env
var GO_BUILD_FLAGS.
Add a channel over which we can propose cluster config changes to
raft. In an upcoming commit we'll add an HTTP endpoint that sends config
changes over this channel.
Current etcdctlv3 doesn't handle options related to security when they
are not passed. Connections cannot be established and etcd produces
logs like this:
```
14:09:07 etcd1 | 2016/02/04 14:09:07 transport: http2Server.HandleStreams received bogus greeting from client: "\x16\x03\x01\x00\x9a\x01\x00\x00\x96\x03\x03\xf6\t\xda\x06QV\xb4\xdd\xc1gF\x1cC"
```
This commit fixes the problem. In addition, a case that empty strings
are passed to the options (e.g. --key="") are treated as error.
Perviously, we only use 8bits from member identification
in id generation. The conflict rate is A(256,3)/256^3, which
is around 1%. Now we use 16bites to reduce the rate to 0.005%.
We can attach the full member id into id generation if needed...
On the latest master branch, etcdserver/etcdserverpb/etcdserver.pb.go
is changed when scripts/genproto.sh is executed. The content only has
changes for comment. Therefore it is not important but the change is
annoying when we update the proto file.
1. proc.ExpectRegex returns the result of regexp.MatchReader,
which does not return error even if there is no match of regex.
This fixes it by checking the boolean value and if the boolean
value is false, it returns error.
2. Adds more tests and finishes coreos#4259.
3. When we do the regex match correctly, curl request through SSL
returns error. For the purpose of debugging, I changed it to log
without failing the tests. etcdctl with SSL works fine.
4. Add // TODO: 'watch' sometimes times out in Semaphore CI environment but
works fine in every other environments.
5. increase test time
This changes the behavior of KV's range and tx range to return
current revision rather than range revision. This makes populating
range response header easier.
For https://github.com/coreos/etcd/issues/4338.
And resp.Header.Revision should be from the one in storage
when we just do range, because there is no storage data
change.
Those three log statements in node.go have not been using the logger that was passed via `raft.Config`, but instead the default raft logger. This changes it to use the proper logger.
When raft broadcasts a Grant to all nodes, all nodes must
agree on the same lease ID. Otherwise, attaching a key to
a lease will fail since the lease ID is node-dependent.
Current V2 watch waits by encoding URL with wait=true.
When a client sets 'no-sync', it requests directly to
proxy and the proxy redirects it by cloning the request
object, which leads to cancel the original request when
it times out and the cloned request gets closed prematurely.
This fixes coreos#3894 by querying
the original client request in order to not use context timeout
when 'wait=true'.
This commit adds a document that provides tips of how to use the go
client library. Currently it describes how to use the
client.SelectionMode parameter that is added in
https://github.com/coreos/etcd/pull/4030.
removeNode reduces the required quorum size, so some pending entries may
be able to commit after it is applied.
Discovered in cockroachdb/cockroach#3642
This commit adds flags for profiling with runtime/pprof to storage
put:
- --cpuprofile: specify a path of CPU profiling result, if it is not
empty, profiling is activated
- --memprofile: specify a path of heap profiling result, if it is not
empty, profiling is activated
Of course, the flags should be added to RootCmd ideally. However,
adding common flags that shared by children command requires the
ongoing PR: https://github.com/spf13/cobra/pull/220 . Therefore this
commit adds the flags to storage put only.
Current syncWatchers method skips the events that have
prefixes that are being watched when the prefix is not
existent as a key. This fixes https://github.com/coreos/etcd/issues/4191
by adding prefix checking to not skip those events.
Unix commands interprets argument value of first character '-'
as a flag. And this won't be fixed in our upstream CLI libraries.
This adds help message to show users workarounds.
Addressing https://github.com/coreos/etcd/issues/2020.
This commit adds a new option for activating profiling based on pprof
in etcd process.
- -enable-pprof: boolean type option which activates profiling
For example, if a client URL is http://localhost:12379, users and
developers access heap profiler with this URL:
http://localhost:12379/debug/pprof/heap
Basic support for lease operations like create and revoke.
We still need to:
1. attach keys to leases in KV implmentation if lease field is set
2. leader periodically removes expired leases
3. leader serves keepAlive requests and follower forwards keepAlive
requests to leader.
go HTTP library uses type assertion to determine if a connection
is a TLS connection. If we wrapper TLS Listener with any customized
Listener that can create customized Conn, HTTPs will be broken.
This commit fixes the issue.
We want the KV to support recovering from backend to avoid
additional pointer swap. Or we have to do coordination between
etcdserver and API layer, since API layer might have access to
kv pointer and use a closed kv.
Given unix://<socketname>, NewListener will listen on unix socket <socketname>.
This is useful when binding to tcp ports is undesirable (e.g., testing).
Provides two implementations of Recorder-- one that is non-blocking
like the original version and one that provides a blocking channel
to avoid busy waiting or racing in tests when no other synchronization
is available.
storage/storagepb: remove watchID from Event
storage: add WatchResponse to watcher.go to wrap events, watchID
storage: watchableStore to use WatchResponse
storage: kv_test with WatchResponse
etcdserver/api/v3rpc: watch to receive storage.WatchResponse type
Watcher vs Watching in storage pkg is confusing. Watcher should be named
as watchStream since it contains a channel as stream to send out events.
Then we can rename watching to watcher, which actually watches on a key
and send watched events through watchStream.
This commits renames watcher to watchStram.
Current etcd repository has a test for benchmarking a storage backend
in storage/kvstore_bench_test.go. However, it is hard to test various
parameters (e.g. batch interval, a number of keys, etc) with the test.
This commit adds a new benchmarking subcommand "storage" to
tools/benchmark. It will encourage analysis of storage backends with
various parameter and complex workloads.
Exmaple usage:
$ ./benchmark storage put
total: 9.894173792s
average: 9.894173ms
minimum latency: 6.596991ms
maximum latency: 29.455695ms
newStore() uses constants for some important parameters
e.g. batchInerval. It is not suitable for other storage users
(e.g. tools/benchmark). This commit decouples the default parameters
from storage creation. etcd should use a newly added function
newDefaultStore() for creating a store with default parameters.
This commit fixes the issue of creating member dir before validating
the configuration. When member dir exists, it indicates the local etcd
process is a valid etcd member. So we should only create member dir
after we finish configuration validation, joining validation or
discovery validation.
As noted in #4075. There are numerous style issues with this document
but I don't want to go down that rabbit hole so this is an attempt at a
minimally invasive clarification.
As @jonboulle pointed out at
https://github.com/coreos/etcd/pull/4070/files#r48441847:
> math/rand is unrelated to crypto/rand; the latter reads from /dev/urandom and
> is relying on the kernel's PRNG. Just remove the seed entirely.
Current etcd client library chooses a default destination node from
every member of a cluster in a random manner. However, requests of
write and read (for consistent results) need to be forwarded to the
leader node as the nature of Raft algorithm. If the chosen node is a
follower, additional network traffic will be caused by the forwarding
from follower to leader.
Mainly for reducing the forward traffic, this commit adds a new
mechanism for various endpoint selection mode to the client library
which can be configured with client.Config.SelectionMode.
Currently, two modes are provided:
- EndpointSelectionRandom: default, same to existing behavior (pick
a node in a random manner)
- EndpointSelectionPrioritizeLeader: prioritize leader, for the above
purpose
I evaluated the effectiveness of the EndpointSelectionPrioritizeLeader
with 4 t1.micro instances of AWS (3 nodes for etcd cluster and 1 node
for etcd client). Client executes this simple benchmark
(https://github.com/mitake/etcd-things/tree/master/prioritize-leader-bench),
just writes 10000 keys. When SelectionMode == EndpointSelectionRandom
(default), the benchmark needed 1 min and 32.102 sec to finish. When
SelectionMode == EndpointSelectionPrioritizeLeader, the benchmark
needed 1 min 4.760 sec.
Reports depended on writing all results to a large buffered channel and
reading from that synchronously. Similarly, requests were buffered the
same way which can take significant memory on big request strings. Instead,
have reports stream in results as they're produced then print when the
results channel closes.
Database snapshot can be as large as 5GB. It is reasonable
to log before receiving it. Or the user might not know what
is happening and why etcd starts to use IO intensively.
The raft loop would block on the applier's done channel after
persisting the raft messages; the latency could cause dropped network
messages. Instead, asynchronously notify the applier with a buffered
channel when the raft writes complete.
raft's applyc writes block on the server loop's database IO since
the next applyc read must wait on the db operation to finish.
Instead, stream applyc to a run queue outside the server loop.
etcd might generate incomplete proxy config file after a power failure.
It is because we use ioutil.WriteFile. And iotuile.WriteFile does
not call Sync before closing the file.
deadlock if no leader; node selects on propc=nil and writes to Ready,
client blocks on propC in same select as Ready reader, and so progress
of raft state machine deadlocks.
For CoreOS users, they will get a updated version of etcdctl without updating
the etcd server version. And the users cannot really control this behavior.
We do not want to suddenly break them without enough communication.
So we still want the most basic opeartions like get, set, watch of etcdctl2 work
with etcd 0.4. This patches solve the incompability issue.
Mutex is a variable, which means there needs to be only one mutex
value per scope. We don't need a separate mutex inside store struct,
**if we assume that `TxnPut` and `TxnRange` are called ONLY ONCE
per transaction (between `TxnBegin` and `TxnEnd`)**, as documented.
Go 1.4 landed a new testing.M type [1][1] which allows for start-up and
shutdown hooks when running tests. The standard library now uses this
for checking for leaked goroutines in net/http [2][2].
This patch essentially re-ports the updated code from the net/http test
(we were using an older version of it) - in detail:
- updates the test to use `TestMain` instead of relying on
`TestGoroutinesRunning` to be implicitly run after all other tests
- adds a few new goroutines to the list of exceptions (the test itself,
as well as the golang/glog package and pkg/log.MergeLogger, both of
which spin off goroutines to handle log flushing/merging respectively)
- removes a couple of TODOs in the test for extra goroutines that's run
after individual tests (one of these re-enables the http package's
`.readLoop` and the other was an out-of-date TODO)
- re-enables the test
[1]: https://golang.org/pkg/testing/#M
[2]: https://github.com/golang/go/blob/release-branch.go1.4/src/net/http/main_test.go#L18
use defer statement to update `Stats` and report R/W Sucess/Failure, so
that the logic of Store's CURD operation and `Stats` update logic can be
separated.
CI is slow sometime. To make the test less flaky, we increases the
timeout. This does not affect the correctness, but the test might
take longer to finish or to fail.
Having the same variable name as 'index' type
can be confusing and shadowing other variables.
This gives different variable names to 'index' variables.
The issue is not caused by this code, but by reading snapshot
from disk. etcd assumes the snapshot of v3 kv should live in
memory. If not, etcd does not work well anyway.
Watch command run benchmark tests for watch releated operations:
1. watch keys 2. sending events to watchers.
To test 2, this test also puts keys into etcd cluster to trigger
event sending.
Besides the benchmark results showed at client side, the tester
can also monitor the server-side mem/cpu usage and also check
the metrics of slow watchers. If there are a lot of slow watchers,
it means etcd server is over-loaded by watchers.
This adds documentation on MessageType. Having clear explanation about
MessageType helps understand raft logic and debug etcd when there is a
message dropping. This is partially for coreos#3806.
This adds map operation tests for unsafeAddWatching, which
could have been failed without https://github.com/coreos/etcd/pull/3939.
It tests if unsafeAddWatching is correctly updating synced map.
In this case, we know we are waiting for an action happened on
storage. We can do a busy wait instead of calling waitSchedule.
The test previously failed on CI with no observed actions.
This is for coreos#3740. Some users find
`advertise-client-urls` and `initial-cluster` flags confusing to set up a
proxy. This clarifies why we need both flags. And add summary to proxy.md
If headerTimeout is not zero then two context are created but only one is released.
cancelCtx in this case is never released which leads to goroutine leak inside it.
This just makes the V3DemoProcfile flags consistent with the other Procfile.
(Makes the single dash to double dash)
Related to https://github.com/coreos/etcd/pull/3911.
This makes flags to etcdmain consistent with man page in help.go. The standard
Go flag package supports both '-' and '--', different than most of CLI
packages, but etcdctl with codegansta/cli uses double-dash('--'), which
sometimes gets confusing, especially to new users.
One watcher includes multiple watchings, and their events are
sent out through one channel. For the received event, user would like to
know which watching it belongs to.
Introduce a watch ID. When watching on some key, user will get a watch
ID. The watch ID is attached to all events that is observed by this
watch.
This is for coreos#3859 switching slice to map for synced watchings.
For a large amount of synced watchings, map implementation performs better.
When putting 1 million watchers on the same key and canceling them one by
one: original implementation takes 9m7.268221091s, while the one with map
takes only 430.531637ms.
When a watch stream closes, both of the watcher.Chan and closec
will be closed.
If watcher.Chan is closed, we should not send out the empty event.
Sending the empty is wrong and waste a lot of CPU resources.
Instead we should just return.
Based on the configuration doc, seems these two flags are missing
in the help. So add them and the descriptions are from config.go in
the same directory.
Signed-off-by: Yiqiao Pu <ypu@redhat.com>
The point is to decouple the key-value storage layer and the
event notification layer clearly. It gives the watchableKV the
flexibility to define whatever event structure it wants without
breaking the ondisk format at key-value storage layer.
Changes:
1. change the format of key and value stored in backend
Store KeyValue struct instead of Event struct in backend value for
better abstraction as xiang suggests. And record the corresponded
action in the backend key.
2. Remove word 'event' from functions
State outright that etcd is used in production and ready for more of same.
Supersedes #3884.
Adopt #3884 in spirit, but directly in README as jonboulle suggested.
Delete Documentation/production-ready.md.
We should wrap the blocking function with a closure. And first
creates a go routine to execute the function. Or the inner function
blocks before creating the go routine.
We should open real txn for applying txn requests. Or the intermediate
state might be observed by reader.
This also fixes#3803. Same consistent(raft) index per multiple indenpendent
operations confuses consistentStore.
We need to be able to force an election (on one node) after creating a
new group (cockroachdb/cockroach#1384), but it is difficult to ensure
that our call to Campaign does not race with an election that may be
started by raft itself. A redundant call to Campaign should be a no-op
instead of a panic. (But the panic in becomeCandidate remains, because
we don't want to update the term or change the committed index in this
case)
Rather than copying in .proto files, use the same symlink
trick we do for doing the actual etcd build.
Also check for exact version of protoc early on.
Extend the timeout from 1s to defaultRequestTimeout 5s.
The 1s may bring unwanted burden to the target member. If the member is
busy at recovering, it has limited bandwidth for client requests. A
short timeout at client side will retry quickly while keeping the
on-going connections. Thus, etcd will queue lots of requests and
connections and takes long time to clear them. This finally causes the
timeout of member health check.
This problem is a general one that how etcd handles amounts of requests
at the same time in a good way. We don't plan to address it at current
stage.
This adds build constraints in order to pass memory-map flags to bolt.Option.
If backend package passes syscall.MAP_POPULATE flag, the boltdb does read-ahead
which speeds up entire-database read, which then leads to faster storage
Restore. Benchmark result shows for 4GB, it opens and loads 6x faster in SSD,
and 12x faster in HDD. For 2GB, 1.6x faster with MAP_POPULATE in SSD.
The primary goal of this doc is to confirm the memory
consumption of watch is as expected. Each connection
consumes O(10kb) of memory. Each stream consumes O(10kb)
of memory. Each watching consumes < O(1kb) of memory.
Then when you have a large number of watching with small
number of connections and streams, the ave memory consumption
per watch will be O(1kb).
Edit to replace a relative link (won't work with that target) with
an absolute link.
Heading 1 Title Case.
Polish graf 3.
Fixes https://github.com/coreos/docs/issues/662
We have a structure called InternalRaftRequest. Making the function
shorter by calling it processInternalRaftReq seems to be random and
reduce the readability. So we just use the full name.
This fixes coreos#3647 by giving shorter proxy refresh interval whenever
there is no endpoints found. Deleted sleep command in Procfile and proxy
documentation accordingly.
This commit fixes an error log caused by the strict reconfig checking
option.
Before:
14:21:38 etcd2 | 2015-11-05 14:21:38.870356 E | etcdhttp: got unexpected response error (etcdserver: re-configuration failed due to not enough started members)
After:
log
13:27:33 etcd2 | 2015-11-05 13:27:33.089364 E | etcdhttp: etcdserver: re-configuration failed due to not enough started members
The error is not an unexpected thing therefore the old message is
incorrect.
This moves the code to create listener and roundTripper for raft communication
to the same place, and use explicit functions to build them. This prevents
possible development errors in the future.
rafthttp logs repeated messages when amounts of message-drop logs
happen, and it becomes log spamming.
Use MergeLogger to merge log lines in this case.
The logging mechanism is verbose, so it is removed from peerStatus.
We would like to see the status change
of connection with peers, and one error that leads to deactivation.
There is no need to print out all non-repeated errors.
When getting the watched events, it iterate all keys in putm and delm
to generate the events. If we don't delete the key from putm/delm,
it would range on the key that is not actually put or deleted. This is
incorrect.
Fix the panic that happens when single put/delete is watched.
This fixes#3647 by delaying proxy start by 3 seconds. Without this, proxy
starts at the same time as initial cluster and since the default proxy director
refresh interval is 30-second, if cluster is not ready at first trial, the
proxy misses to discover them and has to wait another 30-seconds, which delays
the proxying for first 30-second.
This pairs with remote timeout listeners.
etcd uses timeout listener, and times out the accepted connections
if there is no activity. So the idle connections may time out easily.
Becaus timeout transport doesn't reuse connections, it prevents using
timeouted connection.
This fixes the problem that etcd fail to get version of peers.
This attempts to decouple password-related functions, which previously
existed both in the Store and User structs, by splitting them out into a
separate interface, PasswordStore. This means that they can be more
easily swapped out during testing.
This also changes the relevant tests to use mock password functions
instead of the bcrypt-backed implementations; as a result, the tests are
much faster.
Before:
```
github.com/coreos/etcd/etcdserver/auth 31.495s
github.com/coreos/etcd/etcdserver/etcdhttp 91.205s
```
After:
```
github.com/coreos/etcd/etcdserver/auth 1.207s
github.com/coreos/etcd/etcdserver/etcdhttp 1.207s
```
In rafthttp, when making request to some endpoint, it may receive
response with unexpected status code and header. This indicates the endpoint
doesn't function correctly. It should mark the endpoint unreachable.
pipeline.handle is a long-living one, and should continue to receive
next message to send out when current message fails to send. So it
should `continue` instead of `return` here.
Travis has chronic problems successfully running the integration suite -
and we've successfully moved to Semaphore for that purpose - but can
still be useful as a fail-fast option for testing unit tests and formatting.
pipeline.handle is a long-living one, and should continue to receive
next message to send out when current message fails to send. So it
should `continue` instead of `return` here.
This fixes the failure met in semaphore CI:
```
--- FAIL: TestMultiNodeAdvance-2 (0.01s)
multinode_test.go:458: expect Ready after Advance, but there is
no Ready available
```
When a slow follower receives the snapshot sent from the leader, it
should rename the snapshot file to the default KV file path, and
restore KV snapshot.
Have tested it manually and it works pretty well.
consistentWatchableStore maintains an index that is always consistent
with the latest txn. The index could be used to indicate the progress
of the store so far when recovery.
This helps the test to pass safely in semaphore CI.
Based on my manual testing, it may take at most 500ms to return
error in semaphore CI, so I set 1s as a safe value.
We need to access the underlying store to use its RangeEvents function.
It is not good to use unnecessary type conversion.
The underlying store is also needed for further store upon
watchableStore.
When snapshot store requests raft snapshot from etcdserver apply loop,
it may block on the channel for some time, or wait some time for KV to
snapshot. This is unexpected because raft state machine should be unblocked.
Even worse, this block may lead to deadlock:
1. raft state machine waits on getting snapshot from raft memory storage
2. raft memory storage waits snapshot store to get snapshot
3. snapshot store requests raft snapshot from apply loop
4. apply loop is applying entries, and waits raftNode loop to finish
messages sending
5. raftNode loop waits peer loop in Transport to send out messages
6. peer loop in Transport waits for raft state machine to process message
Fix it by changing the logic of getSnap to be asynchronously creation.
streamTypeMsgApp is only used in etcd 2.0. etcd 2.3 should not talk to
etcd 2.0, either send or receive requests. So I deprecate streamTypeMsgApp
and its related stuffs from rafthttp package.
updating term is only used from streamTypeMsgApp, so it is removed too.
If I understand correctly, `progress` represents the states of follower. For
me, some comments weren't clear because it was missing the subjects of
`progress`. This adds more clarification on who is doing what. Please let me
know if I misunderstood anything. Thanks,
This attempts to fix#3676. `PrevNoExist` checks if the key previously exists
and if so, it returns an error, which is how `mk` command is supposed to work.
The previous code ignores the previous key and overwrites with the later value.
/cc @yichengq
Use snapshotSender to send v3 snapshot message. It puts raft snapshot
message and v3 snapshot into request body, then sends it to the target peer.
When it receives http.StatusNoContent, it knows the message has been
received and processed successfully.
As receiver, snapHandler saves v3 snapshot and then processes the raft snapshot
message, then respond with http.StatusNoContent.
hack: benchmarking to use Content Type: application/x-www-form-urlencoded
boom sends PUT request with Content-Type: text/html, and this cannot be parsed by etcd. It should use Content-Type: application/x-www-form-urlencoded. Post this fix the official etcd benchmarks should be updated as the keysize was not a factor as the value being set was always empty.
rafthttp has different requirements for connections created by the
transport for different usage, and this is hard to achieve when giving
one http.RoundTripper. Pass into pkg the data needed to build transport
now, and let rafthttp build its own transports.
I found some typos. Please let me know if you have any feedback.
Thanks,
Documentation: fix metrics.md typo
Documentation: trim blank lines in metrics.md
time.AfterFunc() creates its own goroutine and calls the callback
function in the goroutine. It can cause datarace like the problem
fixed in the commit de1a16e0f1 . This
commit also fixes the potential dataraces of tests in
etcdserver/server_test.go .
We use url.ParseQuery to parse names-to-urls string, but it has side
effect that unescape the string. If the initial-cluster string has ipv6
which contains `%25`, it will unescape it to `%` and make further url
parse failed.
Fix it by modifiying the parse process.
Go1.4 doesn't support literal IPv6 address w/ zone in
URI(https://github.com/golang/go/issues/6530), so we only enable tests
in Go1.5+.
This fixes the problem that proposal cannot be applied.
When start the etcdserver.run loop, it expects to get the latest
existing snapshot. It should not attempt to request one because the loop
is the entity to create the snapshot.
1. I found a little typo (easily -> easy)
2. If you go to https://coreos.com/etcd/docs/2.0.9/proxy.html,
the proxy flag command is out of width of the web-page. Can
we have line-breaks between flags to make the command easier
to read?
Thanks for the great documentation!
One txn is treated as atomic, and might contain multiple Put/Delete/Range
operations. For now, between these operations, we might call forecCommit
to sync the change to disk, or backend may commit it in background.
Thus the snapshot state might contains an unfinished multiple objects
transaction, which is dangerous if database is restored from the snapshot.
This PR makes KV txn hold batchTx lock during the process and avoids
commit to happen.
When using Snapshot function, it is expected:
1. know the size of snapshot before writing data
2. split snapshot-ready phase and write-data phase. so we could cut
snapshot first and write data later.
Update its interface to fit the requirement of etcdserver.
Before this PR, it always prints nil because cluster info has not been
covered when print:
```
2015-10-02 14:00:24.353631 I | etcdserver: loaded cluster information
from store: <nil>
```
snapshotStore is the store of snapshot, and it supports to get latest snapshot
and save incoming snapshot.
raftStorage supports to get latest snapshot when v3demo is open.
etcd is going to support incremental snapshot, and we design to let it
send at most one snapshot out at first stage. So when one snapshot is in
flight, snapshot request will return error.
When failing to get snapshot when sending MsgSnap, raft prints out
related log and abort sending this message.
Like the commit 8ebc933111, this commit lets simple etcdctl commands
use a context with timeout value passed via -total-timeout.
This commit doesn't change complex commands like watch,
cluster-health, and import because it is not obvious that using the
context in the commands is good or not.
Before this PR, the log is
```
2015/09/1 13:18:31 etcdmain: client: etcd cluster is unavailable or
misconfigured
```
It is quite hard for people to understand what happens.
Now we print out the exact reason for the failure, and explains the way
to handle it.
etcdctl should be capable to use a user specified timeout value for
total command execution, not only per request timeout. This commit
adds a new option --total-timeout to the command. The value passed via
this option is used as a timeout value of entire command execution.
Fixes coreos#3517
New PR from https://github.com/coreos/etcd/pull/3575.
This add `defer` to `s.mu`. Current code does not `Unlock`
in the correct scope, I think.
(Sorry, I accidentally deleted my fork so the changes
might not sound continuous from my previous pull requests.)
Change to the function:
1. specify the meaning of startRev and endRev parameters
2. specify the meaning of returned nextRev
Moreover, it adds unit tests for the function.
It should continue to skip following operations.
The test from rev14 to rev0 fails if it doesn't call continue and append
all revisions of the same main rev to the list.
It should skip last empty generation when the key is just tombstoned.
The rev15 and rev16 in the test fails if it doesn't skip last empty generation
and find previous generations.
Like the commit 6974fc63ed, this commit lets etcdserver forbid
removing started member if quorum cannot be preserved after
reconfiguration if the option -strict-reconfig-check is passed to
etcd. The removal can cause deadlock if unstarted members have wrong
peer URLs.
After enabling v3 demo, it may change the underlying data organization
for v3 store. So we forbid to unset --experimental-v3demo once it has
been used.
I found some grammatical errors in comments.
This pull request was submitted https://github.com/coreos/etcd/pull/3513.
I am resubmitting following the correct guidlines.
Current membership changing functionality of etcd seems to have a
problem which can cause deadlock.
How to produce:
1. construct N node cluster
2. add N new nodes with etcdctl member add, without starting the new members
What happens:
After finishing add N nodes, a total number of the cluster becomes 2 *
N and a quorum number of the cluster becomes N + 1. It means
membership change requires at least N + 1 nodes because Raft treats
membership information in its log like other ordinal log append
requests.
Assume the peer URLs of the added nodes are wrong because of miss
operation or bugs in wrapping program which launch etcd. In such a
case, both of adding and removing members are impossible because the
quorum isn't preserved. Of course ordinal requests cannot be
served. The cluster would seem to be deadlock.
Of course, the best practice of adding new nodes is adding one node
and let the node start one by one. However, the effect of this problem
is so serious. I think preventing the problem forcibly would be
valuable.
Solution:
This patch lets etcd forbid adding a new node if the operation changes
quorum and the number of changed quorum is larger than a number of
running nodes. If etcd is launched with a newly added option
-strict-reconfig-check, the checking logic is activated. If the option
isn't passed, default behavior of reconfig is kept.
Fixes https://github.com/coreos/etcd/issues/3477
- Use one sentence per line for easier diffing
- Walkthrough the thought process and cleanup the grammar
- Move below the other sections
Original author: @philips
The docker build command will use whatever directory contains the
Dockerfile as the build context (including all of its subdirectories).
And the <src> path of ADD must be inside the context of the build.
So change it to build in a specific directory for clean and fast.
We made a mistake on the health endpoint by returning a string "true".
We have to make the etcdctl works for the next version of etcd which
will correct the mistake on the server side.
It is too late to change the server side right now since we already
released a version of etcdctl that only understands "true".
Using Go-style import paths in protos is not idiomatic. Normally, this
detail would be internal to etcd, but the path from which gogoproto
is imported affects downstream consumers (e.g. cockroachdb).
In cockroach, we want to avoid including `$GOPATH/src` in our protoc
include path for various reasons. This patch puts etcd on the same
convention, which allows this for cockroach.
More information: https://github.com/cockroachdb/cockroach/pull/2339#discussion_r38663417
This commit also regenerates all the protos, which seem to have
drifted a tiny bit.
It sets 10s timeout for public GetClusterFromRemotePeers.
This helps the following cases to work well in high latency scenario:
1. proxy sync members from the cluster
2. newly-joined member sync members from the cluster
Besides 10s request timeout, the request is also controlled by dial
timeout and read connection timeout.
"X-Content-Type-Options" was being autoadded, but none of the
test maps took it into account. I saw that "Content-Type" was
also being deleted, so I figured that was the best solution
for this as well.
It specifies request timeout error possibly caused by connection lost,
and print out better log for user to understand.
It handles two cases:
1. the leader cannot connect to majority of cluster.
2. the connection between follower and leader is down for a while,
and it losts proposals.
log format:
```
20:04:19 etcd3 | 2015-08-25 20:04:19.368126 E | etcdhttp: etcdserver:
request timed out, possibly due to connection lost
20:04:19 etcd3 | 2015-08-25 20:04:19.368227 E | etcdhttp: etcdserver:
request timed out, possibly due to connection lost
```
It is good to print it in debug output:
```
21:56:12 etcd1 | 2015-08-25 21:56:12.162406 I | etcdmain: peerTLS: cert
= certs/etcd1.pem, key = certs/etcd1-key.pem, ca = , trusted-ca =
certs/ca.pem, client-cert-auth = true
```
We always want to use GOMAXPROCS() as the way go parses it. When in go1.4, we
want to expose GOMAXPROCS value, so we set GOMAXPROCS explicitly as the
way go 1.4 does and print it out.
But it becomes a problem when go 1.5 changes the way to set GOMAXPROCS.
Fix the problem by stop setting GOMAXPROCS and get its value directly.
Due to this change, it sets default GOMAXPROCS to the
number of CPUs available when compiling in go 1.5, which matches how go 1.5 works:
https://docs.google.com/document/d/1At2Ls5_fhJQ59kDK2DFVhFu3g5mATSXqqV5QrxinasI/edit
This is a behavior change in etcd 2.2.
The original workflow may fail to cancel if stop() cancels the finished
request just before dial() assigning a new cancel. This commit checks
streamReader status before setting cancel to avoid this problem.
It is tested at travis for 300 times. go 1.5 always works well, while
go 1.4 fails to stop once.
This problem is totally fixed at 1.5.
go1.5 adds a Request.Cancel channel, which allows for "race free"
cancellation
(8b4278ffb7).
Our implementation relies on it to always cancel in-flight request.
It includes the initial read benchmark for etcd v3.
This is the first step to give some rough thoughts. I haven't digged
deeper to answer some questions, including why its performance is not
better than HTTP + json, why one put will cause performance downgrade.
Before this commit, it will return wrong create index, mod index.
It lets findGeneration return error when rev is at the gap of two
generations. This leads to the change of compact() code.
Currently, total duration of snapshot saving is exported for
prometheus. For more detailed analysis, this commit let etcd export
durations of marshalling for prometheus.
I think it makes sense to make the proxy listen on the default port so
we can give the proxy more testing by default. Also, this should make it
easy to kill a single etcd member and test that etcdctl still works,
etc.
However, I have hit a bug: the proxy takes several seconds
Non-canonical permission path is useless because the path received
by auth is always canonical, which is due to our ServeMux always
redirects request to canonical path().
This helps users to detect path permission setting error early.
Ref: http://godoc.org/net/http#ServeMux
The request can still time out because we have set dial timeout and
read/write timeout. It increases timeout expectation from 1s to 5s,
but it makes it workable in globally-deployer cluster.
Before this PR, the timeout caused by leader election returns:
```
14:45:37 etcd2 | 2015-08-12 14:45:37.786349 E | etcdhttp: got unexpected
response error (etcdserver: request timed out)
```
After this PR:
```
15:52:54 etcd1 | 2015-08-12 15:52:54.389523 E | etcdhttp: etcdserver:
request timed out, possibly due to leader down
```
This document talks about the technical details of discovery service
protocol. It helps users to learn about how discovery service works and
what behavior to expect.
It uses heartbeat interval and election timeout to estimate the
expected request timeout.
This PR helps etcd survive under high roundtrip-time environment,
e.g., globally-deployed cluster.
It uses heartbeat interval and election timeout to estimate the
commit timeout for internal requests.
This PR helps etcd survive under high roundtrip-time environment,
e.g., globally-deployed cluster.
Add HeaderTimeout field in Config, so users could set timeout for each request.
Before this, one hanged request may block the call for long time. After
this, if the network is good, the user could set short timeout and expect
that API call can attempt next available endpoint quickly.
Follow the simple rule in the atomic package:
"On both ARM and x86-32, it is the caller's responsibility to arrange
for 64-bit alignment of 64-bit words accessed atomically. The first word
in a global variable or in an allocated struct or slice can be relied
upon to be 64-bit aligned."
Tested on a system with /proc/cpuinfo reporting:
processor : 0
model name : ARMv7 Processor rev 1 (v7l)
Features : swp half thumb fastmult vfp edsp thumbee neon vfpv3
tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer : 0x41
CPU architecture: 7
CPU variant : 0x0
CPU part : 0xc0d
CPU revision : 1
The behavior accelarates the happen of the first-time leader election,
so the cluster could elect its leader fast. Technically, it could
help to reduce `electionMs - heartbeatMs` wait time for the first leader election.
Main usage:
1. Quick start for the local cluster when setting a little longer
election timeout
2. Quick start for the global cluster, which sets election timeout to
its maximum 50s.
If the body is closed to stop watching, it will ignore the error from
reading body and return context error.
Before this PR, the cancel when watching always returns error `read tcp
127.0.0.1:57824: use of closed network connection`. After this PR, it
will return expected context canceled error.
etcd always returns 500/503 response when it may have no leader.
So we should log the other 50x response in a normal way.
This helps to log correctly when discovery meets 504 error. Before this
PR, it logs like this:
```
18:31:58 etcd2 | 2015/08/4 18:31:58 discovery: error #0: client: etcd
member https://discovery.etcd.io has no leader
18:31:58 etcd2 | 2015/08/4 18:31:58 discovery: waiting for other nodes:
error connecting to https://discovery.etcd.io, retrying in 4s
```
After this PR:
```
22:20:25 etcd2 | 2015/08/4 22:20:25 discovery: error #0: client: etcd
member https://discovery.etcd.io returns server error [Gateway Timeout]
22:20:25 etcd2 | 2015/08/4 22:20:25 discovery: waiting for other nodes:
error connecting to https://discovery.etcd.io, retrying in 4s
```
This solves the problem that etcd may fatal because its critical path
cannot get file descriptor resource when the number of clients is too
big. The PR lets the client listener close client connections
immediately after they are accepted when
the file descriptor usage in the process reaches some pre-set limit, so
it ensures that the internal critical path could always get file
descriptor when it needs.
When there are tons to clients connecting to the server, the original
behavior is like this:
```
2015/08/4 16:42:08 etcdserver: cannot monitor file descriptor usage
(open /proc/self/fd: too many open files)
2015/08/4 16:42:33 etcdserver: failed to purge snap file open
default2.etcd/member/snap: too many open files
[halted]
```
Current behavior is like this:
```
2015/08/6 19:05:25 transport: accept error: closing connection,
exceed file descriptor usage limitation (fd limit=874)
2015/08/6 19:05:25 transport: accept error: closing connection,
exceed file descriptor usage limitation (fd limit=874)
2015/08/6 19:05:26 transport: accept error: closing connection,
exceed file descriptor usage limitation (fd limit=874)
2015/08/6 19:05:27 transport: accept error: closing connection,
exceed file descriptor usage limitation (fd limit=874)
2015/08/6 19:05:28 transport: accept error: closing connection,
exceed file descriptor usage limitation (fd limit=874)
2015/08/6 19:05:28 etcdserver: 80% of the file descriptor limit is
used [used = 873, limit = 1024]
```
It is available at linux system today because pkg/runtime only has linux
support.
This patch adds error logging in wal.Close() if unlocking and
destroying fail. Though it is hard to handling the errors, logging
would be helpful for trouble shooting.
Even though current etcd does not time out
watches, the client could be running against
an old etcd version or the server may close
polling connection for other reasons.
This patch ignores successful (as in 200)
responses with emtpy bodies instead
of producing JSON errors.
PUT on the endpoint sets the GlobalDebugLevel to json level value.
The action overwrites the origianl log level setting from
users. We need to write doc to warn this.
This method uses raft status exposed at /debug/varz to determine the
health of the cluster. It uses whether commit index increases to
determine the cluster health, and uses whether match index increases to
determine the member health.
This could fix the bug #2711 that fails to detect follower is unhealthy
because it doesn't rely on whether message in long-polling connection is sent.
This health check is stricter than the old one, and reflects the
situation that whether followers are healthy in the view of the leader. One
example is that if the follower is receiving the snapshot, it will turns
out to be unhealthy because it doesn't move forward.
`etcdctl cluster-health` will reflect the healthy view in the raft level,
while connectivity checks reflects the healthy view in transport level.
1. When reset endpoints, client will choose a random endpoint to pin.
2. If the pinned endpoint is healthy, client will keep using it.
3. If the pinned endpoint becomes unhealthy, client will attempt other
endpoints and update its pin.
At present it prints the whole usage and flags, which cause the exact
error message is hidden two screens above.
Fixes#3141
Signed-off-by: Guohua Ouyang <gouyang@redhat.com>
The previous flag parsing has a small issue. It uses
`recursive == true` and `after-index == 0` to determine
if user specifies the sub flags. This is incorrect since
user can specify `after-index = 0`. Then the flag parsing
would be confused.
This commit explicitly find the `--` in the remaining args
and determine the key and cmdArgs accordingly.
If the user sets TLS info, this implies that he wants to listen on TLS.
If etcd finds that urls to listen is still HTTP schema, it prints out
warning to notify user about possible wrong setting.
Change its default value from `127.0.0.1:4001,127.0.0.1:2379` to
`http://127.0.0.1:4001,http://127.0.0.1:2379`
Adding HTTP schema makes its format consistent with etcd's xxx-urls
flags.
If TLS config is empty, etcd downgrades keepalive listener from HTTPS to
HTTP without warning. This results in HTTPS downgrade bug for client urls.
The commit returns error if it cannot listen on TLS.
Reasons for the notice:
1. No users have reported about their feedback about auth feature so
far.
2. We haven't used it internally.
3. This is the first release that includes auth feature, so it is good
to be more cautious.
The discovery service doesn't return HTTP header early when watch
starts. This may trigger ResponseHeaderTimeout and cause the watch
request failed.
The fix on discovery service may take some time. Remove the
ResponseHeaderTimeout first so it behaves as before.
@xiang90 pointed out my earlier commit duplicated a lot of things that
were mentioned earlier in the doc.
This time around I tried just making some gotchas more explicit in the
existing docs instead of just tacking new stuff onto the end.
etcd shows an odd message on configuration error like this (partially):
```
... discovery or bootstrap flags are setChoose one of ...
^^^^^^^^^
```
This commit fixes the message format problem.
Waiting 3ms is not long enough for schedule to work well. The test suite
may fail once per 200 times in travis due to this. Extend this to 10ms
to ensure schedule could work. Now it could run 1000 times successfully
in travis.
The original test code before fb4b0b5cf0
doesn't work because reader side may update the
stream, while writer side writes message to the old stream and fails.
This PR removes unnecessary call to set term, and avoids this problem to
happen on term > 1 in the future.
It is possible to trigger the time.After case if the timer went off
between time.After setting the timer for its channel and the time that
select looked at the channel. So it needs to be longer.
refer: https://groups.google.com/forum/#!topic/golang-nuts/1tjcV80ccq8
The bug happens when restarted member wants to listen on its original
port, but finds out that it has been occupied by some client.
Use well-known port instead of ephemeral port, so client cannot occupy
the listen port anymore.
Update term when AddPeer, or the term in peer will not be updated until
the term is changed. This fixes the log flood happended when the v2.1
follower applies the snapshot from v2.0 leader:
```
rafthttp: cannot attach out of data stream server [0 / 17]
```
or
```
rafthttp: server streaming to 6e3bd23ae5f1eae0 at term 0 has been
stopped
```
If backend is closed, the operations on backend in compact
goroutine will panic. So this PR waits for compact goroutine to exit
before close backend.
This fixes the TestWorkflow failure too.
sendApp accesses the storage several times. Perviously, we
assume that the storage will not be modified during the read
opeartions. The assumption is not true since the storage can
be compacted between the read operations. If a compaction
causes a read entries error, we should not painc. Instead, we
can simply retry the sendApp logic until succeed.
Backup process should be able to read all WALs until io.EOF to
generate a point-in-time backup.
Our WAL file is append-only. And the backup process will lock all
files before start reading, which can prevent the gc routine from
removing any files in the middle.
I've been working on this project for a few weeks and I believe it has some features that could assist maintaining etcd registries. Please check it out and I hope you would like to include it in your list of etcd tools.
Kind regards,
Nikos
multiple cpu running may be slower than single cpu running, so it may
take longer time to remove files.
Increase from 5ms to 20ms to give it enough time.
So it ensures that the message will not be discarded because the receive
side has not been ready, which happens easily in multiple core test.
Use log.fatal instead of log.error. The test exits when there is
something wrong because the error may affect following test cases.
before:
```
2015/06/12 20:06:19 rafthttp: dropped MsgApp from %!s(uint64=2) since
receiving buffer is full
```
after:
```
2015/06/12 13:51:38 rafthttp: dropped MsgProp from 2 since receiving
buffer is full
```
Refactor the fake cancel implementation.
The old one may cancel other in-flight message in random, which leaves
the original target message blocked forever.
Before this PR, it may error like this:
```
--- FAIL: TestUpdateMember-2 (0.00s)
server_test.go:950: action =
[{ApplyConfChange:ConfChangeUpdateNode []}
{ProposeConfChange:ConfChangeUpdateNode []}], want
[{ProposeConfChange:ConfChangeUpdateNode []}
{ApplyConfChange:ConfChangeUpdateNode []}]
```
This fixes the test by recording the proposal event in time.
There exist the possiblity to update attributes of removed member in
reasonable workflow:
1. start member A
2. leader receives the proposal to remove member A
2. member A sends the proposal of update its attribute to the leader
3. leader commits the two proposals
So etcdserver should allow to update attributes of removed member.
-bind-addr is etcd 0.4 flag, and we should deprecate it.
Moreover, this makes Procfile fit the workflow we mention in the doc,
which helps ourselves find the problem first.
ForceGosched() performs bad when GOMAXPROCS>1. When GOMAXPROCS=1, it
could promise that other goroutines run long enough
because it always yield the processor to other goroutines. But it cannot
yield processor to goroutine running on other processors. So when
GOMAXPROCS>1, the yield may finish when goroutine on the other
processor just runs for little time.
Here is a test to confirm the case:
```
package main
import (
"fmt"
"runtime"
"testing"
)
func ForceGosched() {
// possibility enough to sched up to 10 go routines.
for i := 0; i < 10000; i++ {
runtime.Gosched()
}
}
var d int
func loop(c chan struct{}) {
for {
select {
case <-c:
for i := 0; i < 1000; i++ {
fmt.Sprintf("come to time %d", i)
}
d++
}
}
}
func TestLoop(t *testing.T) {
c := make(chan struct{}, 1)
go loop(c)
c <- struct{}{}
ForceGosched()
if d != 1 {
t.Fatal("d is not incremented")
}
}
```
`go test -v -race` runs well, but `GOMAXPROCS=2 go test -v -race` fails.
Change the functionality to waiting for schedule to happen.
add godep for speakeasy and auth entry parsing
add security_user to client
add role to client
add role commands
add auth support to etcdclient and etcdctl(member/user)
add enable/disable to etcdctl
better error messages, read/write/readwrite
Bump go-etcd to include codec changes, add new dependency
verify the error for revoke/add if nothing changed, remove security-merging prefix
Snapshot takes an io.Writer and writes the entire backend data to
the given writer. Snapshot writes a consistent view and does not
block other storage operations.
Restore restores the in-memory states (index and book keeping) of
the storage from the backend data.
When it waits for apply to be done, it should stop the loop if it
receives stop signal.
This helps to print out panic information. Before this PR, if the panic
happens when server loop is applying entries, server loop will wait for
raft loop to stop forever.
This PR sets etcd version and min cluster version in request header,
and let server check version compatibility. rafthttp server
will reject any message from peer with incompatible version(too low
version or too high version), and print out warning logs.
The call to transport.RoundTrip closes the request body regardless of
the value of request.Closed. This causes subsequent calls to RoundTrip
using the same request body to fail.
Fixes#2895
This PR set maxIdleConnsPerHost to 128 to let proxy handle 128 concurrent
requests in long term smoothly.
If the number of concurrent requests is bigger than this value,
proxy needs to create one new connection when handling each request in
the delta, which is bad because the creation consumes resource and may
eat up your ephemeral port.
Godeps should allow me to do
godep restore
godep save -r ./...
But that doesn't work. Try it.
This requires update to the following packages:
github.com/prometheus/client_golang/
github.com/prometheus/procfs
github.com/matttproud/golang_protobuf_extensions/
There were 2 major problems.
1. godeps have code.google.com/p/goprotobuf but that repo doesn't exist
2. prometheus/client_golang/_vendor moved to other packages and godep
(with -r) can't handle it.
At the end of this we should be able to use godeps again without tons of
black magic. uggh. what a pain in the ass.
The black magic to actually get godeps back in shape was:
```bash
# remove code.google.com/p/goprotobuf (doesn't exist)
# remove all _vendor lines from prometheus (we still have other
# prometheus lines so restore still works)
vi Godeps/Godeps.json
# remove all the crazy vendoring crud because godep doesn't handle it
# correctly
find . -name \*.go | xargs sed -i
's|github.com/coreos/etcd/Godeps/_workspace/src/||'
# ok now, restore as best we can (everything except it wines about
# goprotobuf
godep restore
# now update the packages which were using the old (dead) goprotobuf
go get -u github.com/prometheus/client_golang/
go get -u github.com/matttproud/golang_protobuf_extensions/
# update prometheus procfs because prometheus/client_golang/ has a
# dependancy on this update
go get -u github.com/prometheus/procfs
# get rid of Godeps directory entirely
git rm -rf Godeps
# ok, now, rewrite the Godeps directory and redo the path rewrites
godep save -r ./...
# now put Godeps back into git
git add Godeps/
# commit the new code
git commit -aA
# And now, you can use godeps!
godep restore
godep save -r ./...
git diff
# nothing!!
```
The commit > unstable might not true for follower. The leader only need
to ensure the entry is stored on the majority of nodes to commit an
entry. So the minority of the cluster might receive commit > unstable
append request. This is normal.
raft node should set initial prev hard state to empty.
Or it will not send the first hard coded state to application
until the state changes again.
This commit fixs the issue. It introduce a small overhead, that
the same tate might send to application twice when restarting.
But this is fine.
Because etcd 2.1 will build stream to any existing peers and etcd 2.0
requires the remote to provide most updated term, it is
necessary for streamReader to know the latest term.
This helps etcd 2.1 connect to msgappV1 handler when the remote member
doesn't support msgappV2. And it doesn't print out unsupported handler
error to make log clean.
Fix#2841.
From Prometheus developer:
```
the recommended way for etcd as an open source project and under
consideration of its size would be etcd_<subsystem>_<name>.
```
We made the naming change accordingly.
This reverts commit f8ce5996b0.
etcd no longer resolves TCP addresses passed in through flags,
so there is no need to compare hostname and IP slices anymore.
(for more details: a3892221ee)
Conflicts:
etcdserver/cluster.go
etcdserver/config.go
pkg/netutil/netutil.go
pkg/netutil/netutil_test.go
In all stream types, streamMsgApp needs to be closed when
updating term because its stream connection can only be used under
a certain term. But there is no need to close other streams, which
may waste time and reduce performance.
When failing to get leader stats, it said 'cluster is unhealthy' before.
This is confusing when it cannot get stats because advertised client urls
are set wrong and the cluster is healthy.
etcdhttp will check the cluster version and update its
capability version periodically.
Any new handler's after 2.0 needs to wrap by capability handler
to ensure it is not accessable until rolling upgrade finished.
After this PR, only cluster's interface Cluster is exposed, which makes
code much cleaner. And it avoids external packages to rely on cluster
struct in the future.
The PR extracts types.Cluster from etcdserver.Cluster. types.Cluster
is used for flag parsing and etcdserver config.
There is no need to expose etcdserver.Cluster public, which contains
lots of etcdserver internal details and methods. This is the first step
for it.
1. Persist the cluster version change through raft. When the member is restarted, it can recover
the previous known decided cluster version.
2. When there is a new leader, it is forced to do a version checking immediately. This helps to
update the first cluster version fast.
We store cluster related key in StoreAdminPrefix for some
historical reason. The previous API is called admin. But now,
the admin name is gone and `cluster` is a more clear and correct
name.
Cluster version is the min major.minor of all members in
the etcd cluster. Cluster version is set to the min version
that a etcd member is compatible with when first bootstrapp.
During a rolling upgrades, the cluster version will be updated
automatically.
For example:
```
Cluster [a:1, b:1 ,c:1] -> clusterVersion 1
update a -> 2, b -> 2
after a detection
Cluster [a:2, b:2 ,c:1] -> clusterVersion 1, since c is still 1
update c -> 2
after a detection
Cluster [a:2, b:2 ,c:2] -> clusterVersion 2
```
The API/raft component can utilize clusterVersion to determine if
it can accept a client request or a raft RPC.
We choose polling rather than pushing since we want to use the same
logic for cluster version detection and (TODO) cluster version checking.
Before a member actually joins a etcd cluster, it should check the version
of the cluster. Push does not work since the other members cannot push
version info to it before it actually joins. Moreover, we do not want our
raft RPC system (which is doing the heartbeat pushing) to coordinate cluster version.
Before this PR, people can set listen-client-urls without setting
advertise-client-urls, and leaves advertise-client-urls as default
localhost value. The client libraries which sync the cluster info
fetch wrong advertise-client-urls and cannot connect to the cluster.
This PR avoids this case and provides better UX.
On the other hand, this change is safe because people always want to set
advertise-client-urls if listen-client-urls is set. The default localhost
advertise url cannot be accessed from the outside, and should always be
set except that etcd is bootstrapped with no flag.
The original process is stopping etcd only when pipeline message finds itself
has been removed. After this PR, stream dial has this functionality too.
It helps fast etcd stop, which doesn't need to wait for stream break to
fall back to pipeline, and wait for election timeout to send out message
to detect self removal.
A bit was missing from the documentation on disaster recovery, the reset
of the advertised peer urls for the node recovered from backup. Without
that, any subsequent server joining the cluster would not be able to
speak to the first node.
When removing a member, etcdserver might return 410 that indicates
the member has been removed. To client, 410 is a vaild response since
the client might do internal retry.
Add remotes to rafthttp, who help newly joined members catch up the
progress of the cluster. It supports basic message sending to remote, and
has no stream connection for simplicity. remotes will not be used
after the latest peers have been added into rafthttp.
Subcommits:
decouple root and security enable/disable
create root role
prefix matching
godep: bump go-etcd to include credentials
add godep for speakeasy and auth entry parsing
appropriate errors for security enable/disable
WIP adding to etcd/client all the security client methods
add guest access
minor ui return tweaks
revert client changes
respond to comments, log more security operations
fix major ensure() bug, add better UX
block recursive access
fix some boneheaded mistakes
fix integration test
last comments
fix up security_api.md
philips nits
fix docs
Decoder should return error for any broken block including the
one that only contains the length field. We should change io.EOF
to io.ErrUnexpectedEOF before return the error.
It is more reasonable to init the variable before passing it as an
argument.
It fixes a bug that etcdserver may panic on server stats when processing
a message from rafthttp streamReader before server stats is initialized
in server.Start().
Based on code from discovery/srv.go. The returns the target as DNS
returns it. In the case of SSL, certs are tied to the hostname and not
the IP address generally.
Solves #2547
We start to resolve host into tcp addrs since we generate
tcp based initial-cluster during srv discovery. However it
creates problems around tls and cluster verification. The
srv discovery only needs to use resolved the tcp addr to
find the local node. It does not have to resolve everything
and use the resolved addrs.
This fixes#2488 and #2226
If the process dies during wal.cut(), it might leave a corrupted wal
file. This commit solves the problem by creating a temp wal file first,
then atomically rename it to a wal file when we are sure it is vaild.
It exposes the metrics of file descriptor limit and file descriptor used.
Moreover, it prints out warning when more than 80% of fd limit has been used.
```
2015/04/08 01:26:19 etcdserver: 80% of the file descriptor limit is open
[open = 969, limit = 1024]
```
SaveSnap and Save are called in separate goroutines now. Allow at most
one WAL function being called at one time to protect internal fields and
guarantee execution order.
Or one possible bug is that the new cut file is started with snapshot
entry instead of crc entry.
The patch decreases the allocs when sending one AppEntry in msgappv2
stream from 30 to 9. This helps reduce CPU load when etcd is under
high write load.
msgappv2 stream is used to send all MsgApp, and replaces the
functionality of msgapp stream. Compared to v1, it has several
advantanges:
1. The output message is exactly the same with the input one, which
cannot be done in v1.
2. It uses one connection to stream persistently, which prevents message
reorder and saves the time to request stream.
3. It transmits 10 addiontional bytes in the procedure of committing one
proposal, which is trivia for idle time.
4. It transmits less bytes when committing mutliple proposals or keep
committing proposals.
etcd now compact raft storage asynchronously, and append entry to raft
storage may happen at the same time. Add the lock to fix the bug that
the entries saved in storage may be organized in a wrong way.
Stop raftNode goroutine before stopping server goroutine, so
server.Stop does stop all underlying stuffs elegantly now. This fixes
the problem that previous-round lock on WAL may not be released when
etcd is restarted.
Upgrade test listens on a fixed port, which may fail with 'bind address
already in use' if the port was just used to send tcp sockets.
The commit makes it listen on a random available port to avoid this.
It is safe to repair the unexpectedEOF error in the last wal. raft
will not send out message before the entry successfully comitted
into wal. Thus we can safely truncate the last entry in the wal
to repair.
When talking about missing events on a particular key, the 1000 event history
limit can be understood as being per key, instead of etcd-wide events. Make it
clear that it is across all etcd keys.
Each progress has a inflighs sliding window. When the progress
is in replicate state, inflights will control the sending speed
of the leader.
The leader can have at most maxInflight number of inflight
messages for each replicate progress. Receving a appResp moves
forward the sliding window. Heartbeat response free one
slot if the window is full.
Currently this doesn't work if a user wants to try out a single machine
cluster but change the name for whatever reason. This is because the
name is always "default" and the
```
./bin/etcd -name 'baz'
```
This solves our problem on CoreOS where the default is `ETCD_NAME=%m`.
limit the max size of entries sent per message.
Lower the cost at probing state as we limit the size per message;
lower the penalty when aggressively decrease to a too low next.
stub out security
further wip
Last stub before CRUD for roles
Complete role merging
start tests
add Godep for golang.org/x/crypto/bcrypt
first round of comments
add tests, remove root addition (will be added back as part of creation)
Add security checks for /v2/machines and /v2/keys
Allow non-root to determine if security is enabled, get machine list.
Responding to comments, remove multiple verbs (like /v2/security/user/foo/password)
add some prefixes to the logging
etcd does not provide enough flexibility to configure server SSL and
client authentication separately. When configuring server SSL the
`--ca-file` flag is required to trust self-signed SSL certificates
used to service client requests.
The `--ca-file` has the side effect of enabling client cert
authentication. This can be surprising for those looking to simply
secure communication between an etcd server and client.
Resolve this issue by introducing four new flags:
--client-cert-auth
--peer-client-cert-auth
--trusted-ca-file
--peer-trusted-ca-file
These new flags will allow etcd to support a more explicit SSL
configuration for both etcd clients and peers.
Example usage:
Start etcd with server SSL and no client cert authentication:
etcd -name etcd0 \
--advertise-client-urls https://etcd0.example.com:2379 \
--cert-file etcd0.example.com.crt \
--key-file etcd0.example.com.key \
--trusted-ca-file ca.crt
Start etcd with server SSL and enable client cert authentication:
etcd -name etcd0 \
--advertise-client-urls https://etcd0.example.com:2379 \
--cert-file etcd0.example.com.crt \
--key-file etcd0.example.com.key \
--trusted-ca-file ca.crt \
--client-cert-auth
Start etcd with server SSL and client cert authentication for both
peer and client endpoints:
etcd -name etcd0 \
--advertise-client-urls https://etcd0.example.com:2379 \
--cert-file etcd0.example.com.crt \
--key-file etcd0.example.com.key \
--trusted-ca-file ca.crt \
--client-cert-auth \
--peer-cert-file etcd0.example.com.crt \
--peer-key-file etcd0.example.com.key \
--peer-trusted-ca-file ca.crt \
--peer-client-cert-auth
This change is backwards compatible with etcd versions 2.0.0+. The
current behavior of the `--ca-file` flag is preserved.
Fixes#2499.
Support IPv6 address for ETCD_ADDR and ETCD_PEER_ADDR
pkg/flags: Support IPv6 address for ETCD_ADDR and ETCD_PEER_ADDR
pkg/flags: tests for IPv6 addr and bind-addr flags
pkg/flags: IPAddressPort.Host: do not enclose IPv6 address in square brackets
pkg/flags: set default bind address to [::] instead of 0.0.0.0
pkg/flags: we don't need fmt any more
also, one minor fix: net.JoinHostPort takes string as a port value
pkg/flags: fix ipv6 tests
pkg/flags: test both IPv4 and IPv6 addresses in TestIPAddressPortString
etcdmain: test: use [::] instead of 0.0.0.0
The messages in channel are outdated, and there is no need to send
them in the future. It also reports unreachable if there are messages
in the channel.
Follower should not reject the append message with a smaller index than its commit
index. Or it will trigger the leader's resending logic, which might have a high cost.
ReleaseLockTo should not release the lock on the WAL
segment that is right before the given index. When
restarting etcd, etcd needs to read from the WAL segment
that has a smaller index than the snapshot index.
The correct behavior is that ReleaseLockTo releases
the locks w is holding so that w only holds one lock
that has an index smaller than the given index.
If the TLS config is empty, etcd downgrades https to http without a warning.
This commit avoid the downgrade and stoping etcd from bootstrap if it cannot
listen on TLS.
MultiNode is an alternative to raft.Node that is more efficient
when a node may participate in many consensus groups. It is currently
used in the CockroachDB project; this commit merges the
github.com/cockroachdb/etcd fork back into the mainline.
Currently the `etcdctl ls` subcommand help output is a bit misleading.
It mentions that using the `--recursive` flag will output all keys and
values for a given path:
--recursive returns all values for key and child keys
This is inaccurate. The `--recursive` will only output the key names
under the given path. Fix the issue by updating the help string for
the `--recursive` flag.
--recursive returns all key names recursively for the given path
Fixes#2379.
for transport that are using timeout connections, we set the
maxIdleConnsPerHost to -1. The default transport does not clear
the timeout for the connections it sets to be idle. So the connections
with timeout cannot be reused.
The size of MsgSnap may be very big, e.g., 1G.
If its size is big and general streaming is used to send it, it may block
the following messages for several ten seconds, which interrupts the
heartbeat heavily.
Only use pipeline to send MsgSnap.
Do not aggressively compact raft log entries. After a snapshot,
etcd server can compact the raft log upto snapshot index. etcd server
compacts to an index smaller than snapshot to keep some entries in memory.
The leader can still read out the in memory entries to send to a slightly
slow follower. If all the entries are compacted, the leader will send the
whole snapshot or read entries from disk if possible.
Encoding store into json snapshot has quite high CPU cost. And it
will block for a while. This commit makes the encoding process non-
blocking by running it in another go-routine.
raft relies on the link layer to report the status of the sent snapshot.
If the snapshot is still sending, the replication to that remote peer will
be paused. If the snapshot finish sending, the replication will begin
optimistically after electionTimeout. If the snapshot fails, raft will
try to resend it.
New rafthttp uses /raft/stream/msgapp for MsgApp stream, but v2.0 rafthttp
cannot understand it. Use the old endpoint /raft/stream instead for backward
compatibility, and plan to move to new endpoint in the version after the
next one.
Dial timeout is set shorter because
1. etcd is supposed to work in good environment, and the new value is long
enough
2. shorter dial timeout makes dial fail faster, which is good for
performance
WAL should control the cut logic itself. We want to do falloc to
per allocate the space for a segmented wal file at the beginning
and cut it when it size reaches the limit.
etcdctl reconnects to the leader, but was not picking up ssl settings in this
case, which causes it to show unhealthy when this is not the case.
Fixes#2373
docs: Add v2 ACL RFC
Add workflow, fix terminology, make the API JSON, and general cleanup
fixes from xiang90s comments
add permissions struct
update regarding glob matches
rename file
The functionality in pkg/osutil ensures that all interrupt handlers finish
and the process kills itself with the proper signal.
Test for interrupt handling added.
The server shutsdown gracefully by stopping on interrupt (Issue #2277.)
It is not sent out because it is useless to let remote raft step the
message.
Moreover, MsgApp stream reader can always assume that the length
of entries sent is > 0.
doozer in particular is rather confusing to mention since the project
hasn't been worked on in years. While we are at it it might simplify
people's understanding if we remove zookeeper too.
When using very similar flags to our examples, the cluster doesn't bootstrap due to mismatched protocols (`http` vs `https`) in the `-initial-advertise-peer-urls` and `initial-cluster` list:
```
./etcd -name infra0 -initial-advertise-peer-urls https://127.0.0.1:2380 \
> -listen-peer-urls https://127.0.0.1:2380 \
> -initial-cluster-token etcd-cluster-1 \
> -initial-cluster infra0=http://127.0.0.1:2380,infra1=http://127.0.0.1:2381,infra2=http://127.0.0.1:2382 \
> -initial-cluster-state new
2015/01/29 10:32:16 no data-dir provided, using default data-dir ./infra0.etcd
2015/01/29 10:32:16 etcd: listening for peers on https://127.0.0.1:2380
2015/01/29 10:32:16 etcd: listening for client requests on http://localhost:2379
2015/01/29 10:32:16 etcd: listening for client requests on http://localhost:4001
2015/01/29 10:32:16 etcd: stopping listening for client requests on http://localhost:4001
2015/01/29 10:32:16 etcd: stopping listening for client requests on http://localhost:2379
2015/01/29 10:32:16 etcd: stopping listening for peers on https://127.0.0.1:2380
2015/01/29 10:32:16 etcd: infra0 has different advertised URLs in the cluster and advertised peer URLs list
```
This addresses a problem that comes up in the cockroach tests,
in which the order of messages may lead to deadlocks (due to
the fact that we don't have regular heartbeat timers in most
of our tests).
Example output of `$ etcdctl exec-watch` command was wrong and showing non-existing ETCD_VALUE, ETCD_KEY, etc keys. Submitting a PR, which updates those examples with correct keys (ETCD_WATCH_VALUE, ETCD_WATCH_KEY, etc).
The ./test script will no longer run the integration tests. To run the
integration test, set the INTEGRATION env var to a nonzero value. For
example, `INTEGRATION=y ./test`.
So that it can easily be used in a container.
Symptoms:
$ sudo bin/rkt run ../etcd/etcd-${VERSION}-linux-amd64.aci
Error: Unable to open "/lib64/ld-linux-x86-64.so.2": No such file or directory
Now that heartbeats are distinct from MsgApp{,Resp}, the retries
currently performed in stepLeader's MsgAppResp section are only
performed on an actual MsgAppResp (or a new MsgProp). This means
that it may take a long time to recover from a dropped MsgAppResp
in a quiet cluster.
This commit adds a dedicated heartbeat response message. This message
does not convey the follower's current log position because the
MsgHeartbeat does not include the leaders term and index. Upon receipt
of a heartbeat response, the leader may retry the latest MsgApp if it
believes the follower to be behind.
In code outside the raft package, we cannot call raft.bcastHeartbeat
directly. Instead, to control heartbeats we set heartbeatInterval to 1
and call Tick().
According to http://godoc.org/fmt#Scan, if scan number is less than the
number of arguments, err will report why. So we don't need to handle
this error case.
etcd resolves DNS hostnames to IP addresses for client and peer URLs
before creating any listening sockets.
The following messages are logged during startup:
etcd: Resolving infra0.coreos.com:2380 to 10.0.1.10:2380
Fixes#1991
It is reasonable for the leader to wait for the reply before sending out the next
msgApp or msgSnap for the follower in bad path. Or the leader will send out useless
messages if the previous message is rejected or the previous message is a snapshot.
Especially for the snapshot case, the leader will be 100% to send out duplicate message
including the snapshot, which is a huge waste.
This commit implement a timeout based wait mechanism. The timeout for normal msgApp is a
heartbeatTimeout and the timeout for snapshot is electionTimeout(snapshot is larger). We
can implement a piggyback mechanism(application notifies the msg lost) in the future
if necessary.
As we move to container-based infrastructure testing env
on travis, the tcp write buffer is more than 1MB. Change
the test according to the change on the testing env.
Perviously, etcd retries three times for timeout error. After this
commit, etcd retries forever. Also this commit make etcd client
aware of gateway timetout.
If amounts of MsgProp goes to the follower, it could batch them and
forward to the leader. This can avoid dropping MsgProp in good path
due to exceed maximal serving in sender.
Moreover, batching MsgProp can increase the throughput of proposals that
come from follower.
This panic can never be reached when using raft.Node, because we only
read from propc when there is a leader. However, it is possible to see
this error when using raft the raft object directly (as in MultiNode),
and in this case it is better to simply drop the proposal (as if we had
sent it to a leader that immediately vanished).
Add an error return to MemoryStorage.Append for consistency.
If we cannot find the `m.from` from current peers in the raft and it is a response
message, we should filter it out or raft panics. We are not targetting to avoid
malicious peers.
It has to be done in the raft node layer syncchronously. Although we can check
it at the application layer asynchronously, but after the checking and before
the message going into raft, the raft state machine might make progress and
unfortunately remove the `m.from` peer.
It is necessary to make this check because of the following case:
1. memory storage contains ents from index 0 to 50, and unstable has
ents from index 50 to 60.
2. raft receives an incoming snapshot with index 100.
3. raft restores its unstable to 100, but has not applied snapshot on memory storage.
4. raft receives an out-dated MsgApp from index 60.
5. raft finds the term of index 60 to check the match.
6. raft asks memory storage about the term of index 60 after it failed to get
it from unstable.
7. memory storage panics because it knows nothing about index 60.
Lock implementation for etcd. It has three go routines:
a) acquire - loop that watches for the lock to be free and tries to acquire it.
b) watch - to watch for lock changes
c) refresh - to refresh the ttl when the lock is acquired
All the changes in lock ownership are notified on the events channel. Any feedback welcome!
Memory storage should append all entries that have greater index
than the snap.Matedata.Index. We first truncate the old parts of
incoming entries. Then truncate the existing entries in the storage.
At last, we append the incoming entries to the existing entries.
stableTo should only mark the index stable if the term is matched. After raft sends out unstable
entries to application, raft makes progress without waiting for reply. When the appliaction
calls the stableTo to notify the entries up to "index" are stable, raft might have truncated
some entries before "index" due to leader lost. raft must verify the (index,term) of stableTo,
before marking the entries as stable.
etcdmain: Fix misuse "-addr" flag
In code, it uses "-advertise-client-urls" or "-addr" flags to get the list of this member's peer URLs,
It should be using "-peer-addr" flag instead of "-addr" flag.
Streaming buffer is used for:
1. hand over data to io goroutine in non-blocking way
2. hold pressure for temprorary network delay
3. be able to wait on I/O instead of data coming under high throughput
The old 1024 value is too small and is very likely to be full and
break the streaming when suffering temprorary network delay.
It should ignore the compact operation instead of panic because the case that
the log is restored from snapshot before executing compact is reasonable.
When loading from a backup with a snapshot and WAL, the length of WAL entries
can be lower than the current index integer value, causing a panic when
slicing off uncommitted entries. This looks for WAL entries higher than
the current index before slicing.
* coreos/master:
rafthttp: fix import
raft: should not decrease match and next when handling out of order msgAppResp
Fix migration to allow snapshots to have the right IDs
add snapshotted integration test
fix test import loop
fix import loop, add set to types, and fix comments
etcdserver: autodetect v0.4 WALs and upgrade them to v0.5 automatically
wal: add a bench for write entry
rafthttp: add streaming server and client
dep: use vendored imports in codegangsta/cli
dep: bump golang.org/x/net/context
Conflicts:
etcdserver/server.go
etcdserver/server_test.go
migrate/snapshot.go
* coreos/master:
scripts: build-docker tag and use ENTRYPOINT
scripts: build-release add etcd-migrate
create .godir
raft: optimistically increase the next if the follower is already matched
raft: add handleHeartbeat handleHeartbeat commits to the commit index in the message. It never decreases the commit index of the raft state machine.
rafthttp: send takes raft message instead of bytes
*: add rafthttp pkg into test list
raft: include commitIndex in heartbeat
rafthttp: move server stats in raftHandler to etcdserver
*: etcdhttp.raftHandler -> rafthttp.RaftHandler
etcdserver: rename sender.go -> sendhub.go
*: etcdserver.sender -> rafthttp.Sender
Conflicts:
raft/log.go
raft/raft_paper_test.go
Compaction is now treated as an implementation detail of Storage
implementations; Node.Compact() and related functionality have been
removed. Ready.Snapshot is now used only for incoming snapshots.
A return value has been added to ApplyConfChange to allow applications
to track the node information that must be stored in the snapshot.
raftpb.Snapshot has been split into Snapshot and SnapshotMetadata, to
allow the full snapshot data to be read from disk only when needed.
raft.Storage has new methods Snapshot, ApplySnapshot, HardState, and
SetHardState. The Snapshot and HardState parameters have been removed
from RestartNode() and will now be loaded from Storage instead.
The only remaining difference between StartNode and RestartNode is that
the former bootstraps an initial list of Peers.
This is useful since we want to pipeline the appendEntry requests. Without
enabling optimistic increasing, the second pipelining appendEntry request
will include the entries the first one has already sent out. We decrease
the next directly to match if the leader receives a rejection for a matched
follower. This happens if one pipelining request get lost and following ones
arrives at the follower.
* coreos/master: (21 commits)
etcdserver: refactor ValidateClusterAndAssignIDs
integration: add integration test for remove member
integration: add test for member restart
version: bump to alpha.3
etcdserver: add buffer to the sender queue
*: gracefully stop etcdserver
Fix up migration tool, add snapshot migration
etcd4: migration from v0.4 -> v0.5
etcdserver: export Member.StoreKey
etcdserver: recover cluster when receiving newer snapshot
etcdserver: check and select committed entries to apply
etcdserver: recover from snapshot before applying requests
raft: not set applied when restored from snapshot
sender: support elegant stop
etcdserver: add StopNotify
etcdserver: fix TestDoProposalStopped test
etcdserver: minor cleanup
etcdserver: validate new node is not registered before in best effort
etcdserver: fix server.Stop()
*: print out configuration when necessary
...
Conflicts:
etcdserver/server.go
etcdserver/server_test.go
raft/log.go
The first entry in the log is a dummy which is used for matchTerm
but may not have an actual payload. This change permits Storage
implementations to treat this term value specially instead of
storing it as a dummy Entry.
Storage.FirstIndex() no longer includes the term-only entry.
This reverses a recent decision to create entry zero as initially
unstable; Storage implementations are now required to make
Term(0) == 0 and the first unstable entry is now index 1.
stableTo(0) is no longer allowed.
Fixes all updates since bcwaldon sketched the original, with cleanup and
into an acutal working state. The commit log follows:
fix pb reference and remove unused file post rebase
unbreak the migrate folder
correctly detect node IDs
fix snapshotting
Fix previous broken snapshot
Add raft log entries to the translation; fix test for all timezones. (Still in progress, but passing)
Fix etcd:join and etcd:remove
print more data when dumping the log
Cleanup based on yichengq's comments
more comments
Fix the commited index based on the snapshot, if one exists
detect nodeIDs from snapshot
add initial tool documentation and match the semantics in the build script and main
formalize migration doc
rename function and clarify docs
fix nil pointer
fix the record conversion test
add migration to test suite and fix govet
We start etcd server in this test without the cluster. Sometimes it panics when
accessing the cluster. Most of the time it does not panic, since we can stop the
server fast enough before applying the first configuration change entry.
Stop should be idempotent. It should simply send a stop signal to the server.
It is the server's responsibility to stop the go-routines and related components.
* coreos/master:
etcdserver: add sender tests
raft: Only call stableTo when we have ready entries or a snapshot.
etcdserver: add ID() function to the Server interface.
sender: use RoundTripper instead of Client in sender
The first Ready after RestartNode (with no snapshot) will have no
unstable entries, so we don't have the correct prevLastUnstablei
when Advance is called. This would cause raftLog.unstable to move
backwards and previously-stable entries would be returned to
the application again.
This should have been caught by the "unexpected Ready" portion of
TestNodeRestart, but it went unnoticed because the Node's goroutine
takes some time to read from advancec and prepare the write to read to
readyc. Added a small (1ms) delay to all such tests to ensure that the
goroutine has time to enter its select wait.
* coreos/master: (27 commits)
pkg/wait: move wait to pkg/wait
etcdserver: do not add/remove/update local member to/from sender hub
etcdserver: not record attributes when add member
raft: add a test for proposeConfChange
raft: block Stop() on n.done, support idempotency
raft: add a test for node proposal
integration: add increase cluster size test
integration: remove unnecessary t.Testing argument
raft: stop the node synchronously
integration: fix test to propagate NewServer errors
etcdserver: move peer URLs check to config
etcdserver: ensure initial-advertise-peer-urls match initial-cluster
raft: add a test for node.Tick
raft: add comment string for TestNodeStart
etcdserver: use member instead of node at etcd level
raft: nodes return sorted ids
raft: update unstable when calling stableTo with 0
*: support updating advertise-peer-url Users might want to update the peerurl of the etcd member in several cases. For example, if the IP address of the physical machine etcd running on is changed, user need to update the adversite-pee-rurl accordingly. This commit makes etcd support updating the advertise-peer-url of its members.
transport: create a tls listener only if the tlsInfo is not empty and the scheme is HTTPS
etcdserver: use member pointer for all tests
...
Conflicts:
etcdserver/server.go
raft/log.go
raft/log_test.go
raft/node.go
This entry is now persisted through the normal flow instead of appearing
in the stored log at creation time. This is how things worked before
the Storage interface was introduced. (see coreos/etcd#1689)
Callers must in general have a reference to their Storage objects to
transfer entries from Ready to Storage, so it doesn't make sense to
create a hidden Storage for them.
By explicitly creating Storage objects in tests we can remove a
few casts of raftLog's storage field.
This adds a check to setupCluster to ensure that the list of URLs
specified in `initial-advertise-peer-urls` matches those configured in
`initial-cluster` for this node. Also updates the documentation to
clarify this and address some changes in wording.
Users might want to update the peerurl of the etcd member in several cases.
For example, if the IP address of the physical machine etcd running on is
changed, user need to update the adversite-pee-rurl accordingly.
This commit makes etcd support updating the advertise-peer-url of its members.
This change splits the raftLog.entries array into an in-memory
"unstable" list and a pluggable interface for retrieving entries that
have been persisted to disk. An in-memory implementation of this
interface is provided which behaves the same as the old version;
in a future commit etcdserver could replace the MemoryStorage with
one backed by the WAL.
The wait ensures that cluster goes into the stable stage, which means that
leader has been elected and starts to heartbeat to followers.
This makes future client requests always handled in time, and there is no
need to retry sending requests.
This function was fundamentally buggy, as a panic could be trivially
triggered by setting the wrong environment variable (e.g.
ETCD_BIND_ADDR=foo). Instead, let's propagate the error and present it
to the user in a cleaner way.
It is the correct thing to do to ensure that the communication is full
of out-of-date messages.
It results in that integration testing is very easy to throw MsgProp away,
and makes client wait until 5 min timeout. Sync interval and heartbeat are
increased to alleviate the traffic.
There's no real need to expose a Discoverer interface/struct when the
only use of the interface (and indeed the module) is to invoke a single
function. This isn't Java, after all. So instead, simplify to Discovery
exposing just two functions: JoinCluster (i.e. what was formerly called
"discovery"), and GetCluster (hitherto "ProxyDiscovery")
Node set the applied to committed right after it sends out Ready to application. This is not
correct since the application has not actually applied the entries at that point. We add a
Advance interface to Node. Application needs to call Advance to tell raft Node its progress.
Also this change can avoid unnecessary copying when application is still applying entires but
there are more entries to be applied.
1. etcd fatals if there is critical error in the system and operator should
do something for it
2. etcd panics if there happens something unexpected, and it should be
reported to us to debug.
When waiting for a watch result, we expect the good path to complete
quickly here so we don't need to time out so aggressively. (Failure
noted in #1600)
In certain cases (for example, if a cluster peer is accessible but it
has no members listed), the httpClusterClient could have an empty set of
endpoints as a result of the Sync. This means that its Do function could
potentially return a nil response and nil error, with catastrophic
consequences for callers.
To be safe (particularly about this latter behaviour), this change
errors in both Sync and Do if no endpoints are available.
The docs for the other APIs use curl for example usage, which matches
the docs for the etcd APIs.
Other cleanup include fixing usage of peer ports and using 10.0.0.x IPs
throughout.
- Provide more concrete examples and explanation
- Cleanup the formatting to one sentence per line, this makes reviewing
easier
- Point to existing docs on wal and snap instead of trying to duplicate
it here again.
This creates a simple ID type (wrapped around uint64) to provide for
standard serialization/deserialization to a string (i.e. base 16
encoded). This replaces strutil so now that package is removed.
There's no real need for do and doWithTimeout to return Responses when
the only field of interest is the status code.
This also removes the superfluous httpMembersAPIResponse struct.
If any of this initialization fails, something very bad has happened,
and we should not continue as-is (this has previously manifested in
strange bugs)
etcd checks that the data dir is writable by writing and removing an
empty file to the data dir during startup and exits non-zero if that
fails.
fixes#876
etcd is Apache 2.0 licensed and accepts contributions via Github pull requests. This document outlines some of the conventions on commit message formatting, contact points for developers and other resources to make getting your contribution into etcd easier.
etcd is Apache 2.0 licensed and accepts contributions via GitHub pull requests. This document outlines some of the conventions on commit message formatting, contact points for developers and other resources to make getting your contribution into etcd easier.
# Email and chat
@ -12,6 +12,14 @@ etcd is Apache 2.0 licensed and accepts contributions via Github pull requests.
- Fork the repository on GitHub
- Read the README.md for build instructions
## Reporting bugs and creating issues
Reporting bugs is one of the best ways to contribute. However, a good bug report
has some very specific qualities, so please read over our short document on
before you submit your bug report. This document might contain links known
issues, another good reason to take a look there, before reporting your bug.
## Contribution flow
This is a rough outline of what a contributor's workflow looks like:
@ -27,11 +35,11 @@ Thanks for your contributions!
### Code style
The coding style suggested by the Golang community is used in etcd. See the [style doc](https://code.google.com/p/go-wiki/wiki/CodeReviewComments) for details.
The coding style suggested by the Golang community is used in etcd. See the [style doc](https://github.com/golang/go/wiki/CodeReviewComments) for details.
Please follow this style to make etcd easy to review, maintain and develop.
### Format of the Commit Message
### Format of the commit message
We follow a rough convention for commit messages that is designed to answer two
questions: what changed and why. The subject line should feature the what and
Return an HTTP 200 OK response code and a representation of all members in the etcd cluster:
```
Example Request: GET
http://localhost:2379/v2/admin/members/
Response formats: JSON
Example Response:
```
```json
{
"members":[
{
"id":"272e204152",
"name":"node1",
"peerURLs":[
"http://10.0.0.10:2379"
],
"clientURLs":[
"http://10.0.0.10:2380"
]
},
{
"id":"2225373f43",
"name":"node2",
"peerURLs":[
"http://127.0.0.11:2379"
],
"clientURLs":[
"http://127.0.0.11:2380"
]
},
]
}
```
### POST /v2/admin/members/
Add a member to the cluster.
Returns an HTTP 201 response code and the representation of added member with a newly generated a memberID when successful. Returns a string describing the failure condition when unsuccessful.
If the POST body is malformed an HTTP 400 will be returned. If the member exists in the cluster or existed in the cluster at some point in the past an HTTP 500(TODO: fix this) will be returned. If the cluster fails to process the request within timeout an HTTP 500 will be returned, though the request may be processed later.
```
Example Request: POST
http://localhost:2379/v2/admin/members/
Body:
[{"PeerURLs":["http://10.0.0.10:2379"]}]
Respose formats: JSON
Example Response:
```
```json
[
{
"id":"3777296169",
"peerURLs":[
"http://10.0.0.10:2379"
],
},
]
```
### DELETE /v2/admin/members/:id
Remove a member from the cluster.
Returns empty when successful. Returns a string describing the failure condition when unsuccessful.
If the member does not exist in the cluster an HTTP 500(TODO: fix this) will be returned. If the cluster fails to process the request within timeout an HTTP 500 will be returned, though the request may be processed later.
|-peer-addr |-initial-advertise-peer-urls |If specified, peer-addr will be used as the only peer URL. Error if both flags specified.|
|-addr |-advertise-client-urls |If specified, addr will be used as the only client URL. Error if both flags specified.|
|-peer-bind-addr |-listen-peer-urls |If specified, peer-bind-addr will be used as the only peer bind URL. Error if both flags specified.|
|-bind-addr |-listen-client-urls |If specified, bind-addr will be used as the only client bind URL. Error if both flags specified.|
|-peers |none |Deprecated. The -initial-cluster flag provides a similar concept with different semantics. Please read this guide on cluster startup.|
|-peers-file |none |Deprecated. The -initial-cluster flag provides a similar concept with different semantics. Please read this guide on cluster startup.|
- 1x dedicated local SSD mounted under /var/lib/etcd
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
- etcd version 2.1.0 alpha
## etcd Cluster
3 etcd members, each runs on a single machine
## Testing
Bootstrap another machine and use the [boom HTTP benchmark tool][boom] to send requests to each etcd member. Check the [benchmark hacking guide][hack-benchmark] for detailed instructions.
## Performance
### reading one single key
| key size in bytes | number of clients | target etcd server | read QPS | 90th Percentile Latency (ms) |
- 1x dedicated local SSD mounted as etcd data directory
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
## etcd Cluster
3 etcd 2.2.0 members, each runs on a single machine.
Detailed versions:
```
etcd Version: 2.2.0
Git SHA: e4561dd
Go Version: go1.5
Go OS/Arch: linux/amd64
```
## Testing
Bootstrap another machine, outside of the etcd cluster, and run the [`boom` HTTP benchmark tool](https://github.com/rakyll/boom) with a connection reuse patch to send requests to each etcd cluster member. See the [benchmark instructions](../../hack/benchmark/) for the patch and the steps to reproduce our procedures.
The performance is calulated through results of 100 benchmark rounds.
## Performance
### Single Key Read Performance
| key size in bytes | number of clients | target etcd server | average read QPS | read QPS stddev | average 90th Percentile Latency (ms) | latency stddev |
- Because etcd now records metrics for each API call, read QPS performance seems to see a minor decrease in most scenarios. This minimal performance impact was judged a reasonable investment for the breadth of monitoring and debugging information returned.
- Write QPS to cluster leaders seems to be increased by a small margin. This is because the main loop and entry apply loops were decoupled in the etcd raft logic, eliminating several blocks between them.
- Write QPS to all members seems to be increased by a significant margin, because followers now receive the latest commit index sooner, and commit proposals more quickly.
- 1x dedicated local SSD mounted under /var/lib/etcd
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
## etcd Cluster
3 etcd 2.2.0-rc members, each runs on a single machine.
Detailed versions:
```
etcd Version: 2.2.0-alpha.1+git
Git SHA: 59a5a7e
Go Version: go1.4.2
Go OS/Arch: linux/amd64
```
Also, we use 3 etcd 2.1.0 alpha-stage members to form cluster to get base performance. etcd's commit head is at [c7146bd5][c7146bd5], which is the same as the one that we use in [etcd 2.1 benchmark][etcd-2.1-benchmark].
## Testing
Bootstrap another machine and use the [boom HTTP benchmark tool][boom] to send requests to each etcd member. Check the [benchmark hacking guide][hack-benchmark] for detailed instructions.
## Performance
### reading one single key
| key size in bytes | number of clients | target etcd server | read QPS | 90th Percentile Latency (ms) |
- read QPS in most scenarios is decreased by 5~8%. The reason is that etcd records store metrics for each store operation. The metrics is important for monitoring and debugging, so this is acceptable.
- write QPS to leader is increased by 20~30%. This is because we decouple raft main loop and entry apply loop, which avoids them blocking each other.
- write QPS to all servers is increased by 30~80% because follower could receive latest commit index earlier and commit proposals faster.
- When etcd reaches data size threshold, it may trigger leader election easily and drop part of proposals.
- At most cases, etcd cluster should work smoothly if it doesn't hit the threshold. If it doesn't work well due to insufficient resources, you need to decrease its data size.
| value bytes | key number limitation | suggested data size threshold(MB) | consumed RSS(MB) |
*NOTE*: The watch features are under active development, and their memory usage may change as that development progresses. We do not expect it to significantly increase beyond the figures stated below.
A primary goal of etcd is supporting a very large number of watchers doing a massively large amount of watching. etcd aims to support O(10k) clients, O(100K) watch streams (O(10) streams per client) and O(10M) total watchings (O(100) watching per stream). The memory consumed by each individual watching accounts for the largest portion of etcd's overall usage, and is therefore the focus of current and future optimizations.
Three related components of etcd watch consume physical memory: each `grpc.Conn`, each watch stream, and each instance of the watching activity. `grpc.Conn` maintains the actual TCP connection and other gRPC connection state. Each `grpc.Conn` consumes O(10kb) of memory, and might have multiple watch streams attached.
Each watch stream is an independent HTTP2 connection which consumes another O(10kb) of memory.
Multiple watchings might share one watch stream.
Watching is the actual struct that tracks the changes on the key-value store. Each watching should only consume <O(1kb).
Two components of etcd storage consume physical memory. The etcd process allocates an *in-memory index* to speed key lookup. The process's *page cache*, managed by the operating system, stores recently-accessed data from disk for quick re-use.
The in-memory index holds all the keys in a [B-tree][btree] data structure, along with pointers to the on-disk data (the values). Each key in the B-tree may contain multiple pointers, pointing to different versions of its values. The theoretical memory consumption of the in-memory index can hence be approximated with the formula:
where `c1` is the key metadata overhead and `c2` is the version metadata overhead.
The graph shows the detailed structure of the in-memory index B-tree.
```
In mem index
+------------+
| key || ... |
+--------------+ | || |
| | +------------+
| | | v1 || ... |
| disk <----------------| || | Tree Node
| | +------------+
| | | v2 || ... |
| <----------------+ || |
| | +------------+
+--------------+ +-----+ | | |
| | | | |
| +------------+
|
|
^
------+
| ... |
| |
+-----+
| ... | Tree Node
| |
+-----+
| ... |
| |
------+
```
[Page cache memory][pagecache] is managed by the operating system and is not covered in detail in this document.
## Testing Environment
etcd version
- git head https://github.com/coreos/etcd/commit/776e9fb7be7eee5e6b58ab977c8887b4fe4d48db
GCE n1-standard-2 machine type
- 7.5 GB memory
- 2x CPUs
## In-memory index memory usage
In this test, we only benchmark the memory usage of the in-memory index. The goal is to find `c1` and `c2` mentioned above and to understand the hard limit of memory consumption of the storage.
We calculate the memory usage consumption via the Go runtime.ReadMemStats. We calculate the total allocated bytes difference before creating the index and after creating the index. It cannot perfectly reflect the memory usage of the in-memory index itself but can show the rough consumption pattern.
| N | versions | key size | memory usage |
|------|----------|----------|--------------|
| 100K | 1 | 64bytes | 22MB |
| 100K | 5 | 64bytes | 39MB |
| 1M | 1 | 64bytes | 218MB |
| 1M | 5 | 64bytes | 432MB |
| 100K | 1 | 256bytes | 41MB |
| 100K | 5 | 256bytes | 65MB |
| 1M | 1 | 256bytes | 409MB |
| 1M | 5 | 256bytes | 506MB |
Based on the result, we can calculate `c1=120bytes`, `c2=30bytes`. We only need two sets of data to calculate `c1` and `c2`, since they are the only unknown variable in the formula. The `c1=120bytes` and `c2=30bytes` are the average value of the 4 sets of `c1` and `c2` we calculated. The key metadata overhead is still relatively nontrivial (50%) for small key-value pairs. However, this is a significant improvement over the old store, which had at least 1000% overhead.
## Overall memory usage
The overall memory usage captures how much RSS etcd consumes with the storage. The value size should have very little impact on the overall memory usage of etcd, since we keep values on disk and only retain hot values in memory, managed by the OS page cache.
| N | versions | key size | value size | memory usage |
Based on the result, we know the value size does not significantly impact the memory consumption. There is some minor increase due to more data held in the OS page cache.
* New development occurs on the [master branch][master].
* Master branch should always have a green build!
* Backwards-compatible bug fixes should target the master branch and subsequently be ported to stable branches.
* Once the master branch is ready for release, it will be tagged and become the new stable branch.
The etcd team has adopted a *rolling release model* and supports one stable version of etcd.
### Master branch
The `master` branch is our development branch. All new features land here first.
To try new and experimental features, pull `master` and play with it. Note that `master` may not be stable because new features may introduce bugs.
Before the release of the next stable version, feature PRs will be frozen. We will focus on the testing, bug-fix and documentation for one to two weeks.
### Stable branches
All branches with prefix `release-` are considered _stable_ branches.
After every minor release (http://semver.org/), we will have a new stable branch for that release. We will keep fixing the backwards-compatible bugs for the latest stable release, but not previous releases. The _patch_ release, incorporating any bug fixes, will be once every two weeks, given any patches.
Starting an etcd cluster requires that each node knows another in the cluster. If you are trying to bring up a cluster all at once, say using a cloud formation, you also need to coordinate who will be the initial cluster leader. The discovery protocol helps you by providing an automated way to discover other existing peers in a cluster.
For more information on how etcd can locate the cluster, see the [finding the cluster][cluster-finding] documentation.
Please note - at least 3 nodes are required for [cluster availability][optimal-cluster-size].
To use the discovery API, you must first create a unique discovery URL for your etcd cluster. Visit [https://discovery.etcd.io/new](https://discovery.etcd.io/new) to create a new discovery URL.
You can inspect the list of peers by viewing `https://discovery.etcd.io/<cluster id>`.
### Start etcd With the Discovery Flag
Specify the `-discovery` flag when you start each etcd instance. The list of existing peers in the cluster will be downloaded and configured. If the instance is the first peer, it will start as the leader of the cluster.
The discovery API communicates with a separate etcd cluster to store and retrieve the list of peers. CoreOS provides [https://discovery.etcd.io](https://discovery.etcd.io) as a free service, but you can easily run your own etcd cluster for this purpose. Here's an example using an etcd cluster located at `10.10.10.10:4001`:
If you're interested in how to discovery API works behind the scenes, read about the [Discovery Protocol](https://github.com/coreos/etcd/blob/master/Documentation/discovery-protocol.md).
## Setting Peer Addresses Correctly
The Discovery API submits the `-peer-addr` of each etcd instance to the configured Discovery endpoint. It's important to select an address that *all* peers in the cluster can communicate with. For example, if you're located in two regions of a cloud provider, configuring a private `10.x` address will not work between the two regions, and communication will not be possible between all peers.
## Stale Peers
The discovery API will automatically clean up the address of a stale peer that is no longer part of the cluster. The TTL for this process is a week, which should be long enough to handle any extremely long outage you may encounter. There is no harm in having stale peers in the list until they are cleaned up, since an etcd instance only needs to connect to one valid peer in the cluster to join.
## Lifetime of a Discovery URL
A discovery URL identifies a single etcd cluster. Do not re-use discovery URLs for new clusters.
When a machine starts with a new discovery URL the discovery URL will be activated and record the machine's metadata. If you destroy the whole cluster and attempt to bring the cluster back up with the same discovery URL it will fail. This is intentional because all of the registered machines are gone including their logs so there is nothing to recover the killed cluster.
We use Raft as the underlying distributed protocol which provides consistency and persistence of the data across all of the etcd instances.
Let start by creating 3 new etcd instances.
We use `-peer-addr` to specify server port and `-addr` to specify client port and `-data-dir` to specify the directory to store the log and info of the machine in the cluster:
**Note:** If you want to run etcd on an external IP address and still have access locally, you'll need to add `-bind-addr 0.0.0.0` so that it will listen on both external and localhost addresses.
A similar argument `-peer-bind-addr` is used to setup the listening address for the server port.
Let's join two more machines to this cluster using the `-peers` argument. A single connection to any peer will allow a new machine to join, but multiple can be specified for greater resiliency.
If one machine disconnects from the cluster, it could rejoin the cluster automatically when the communication is recovered.
If one machine is killed, it could rejoin the cluster when started with old name. If the peer address is changed, etcd will treat the new peer address as the refreshed one, which benefits instance migration, or virtual machine boot with different IP. The peer-address-changing functionality is only supported when the majority of the cluster is alive, because this behavior needs the consensus of the etcd cluster.
**Note:** For now, it is user responsibility to ensure that the machine doesn't join the cluster that has the member with the same name. Or unexpected error will happen. It would be improved sooner or later.
### Killing Nodes in the Cluster
Now if we kill the leader of the cluster, we can get the value from one of the other two machines:
```sh
curl -L http://127.0.0.1:4002/v2/keys/foo
```
We can also see that a new leader has been elected:
```
curl -L http://127.0.0.1:4002/v2/leader
```
```
http://127.0.0.1:7002
```
or
```
http://127.0.0.1:7003
```
### Testing Persistence
Next we'll kill all the machines to test persistence.
Type `CTRL-C` on each terminal and then rerun the same command you used to start each machine.
Your request for the `foo` key will return the correct value:
```sh
curl -L http://127.0.0.1:4002/v2/keys/foo
```
```json
{
"action":"get",
"node":{
"createdIndex":4,
"key":"/foo",
"modifiedIndex":4,
"value":"bar"
}
}
```
### Using HTTPS between servers
In the previous example we showed how to use SSL client certs for client-to-server communication.
Etcd can also do internal server-to-server communication using SSL client certs.
To do this just change the `-*-file` flags to `-peer-*-file`.
If you are using SSL for server-to-server communication, you must use it on all instances of etcd.
### Bootstrapping a new cluster by name
An etcd server is uniquely defined by the peer addresses it listens to. Suppose, however, that you wish to start over, while maintaining the data from the previous cluster -- that is, to pretend that this machine has never joined a cluster before.
You can use `--initial-cluster-name` to generate a new unique ID for each node, as a shared token that every node understands. Nodes also take this into account for bootstrapping the new cluster ID, so it also provides a way for a machine to listen on the same interfaces, disconnect from one cluster, and join a different cluster.
Individual node configuration options can be set in three places:
1. Command line flags
2. Environment variables
3. Configuration file
Options set on the command line take precedence over all other sources.
Options set in environment variables take precedence over options set in
configuration files.
## Cluster Configuration
Cluster-wide settings are configured via the `/config` admin endpoint and additionally in the configuration file. Values contained in the configuration file will seed the cluster setting with the provided value. After the cluster is running, only the admin endpoint is used.
The full documentation is contained in the [API docs](https://github.com/coreos/etcd/blob/master/Documentation/api.md#cluster-config).
*`activeSize` - the maximum number of peers that can participate in the consensus protocol. Other peers will join as standbys.
*`removeDelay` - the minimum time in seconds that a machine has been observed to be unresponsive before it is removed from the cluster.
*`syncInterval` - the amount of time in seconds between cluster sync when it runs in standby mode.
## Command Line Flags
### Required
*`-name` - The node name. Defaults to a UUID.
### Optional
*`-addr` - The advertised public hostname:port for client communication. Defaults to `127.0.0.1:4001`.
*`-discovery` - A URL to use for discovering the peer list. (i.e `"https://discovery.etcd.io/your-unique-key"`).
*`-http-read-timeout` - The number of seconds before an HTTP read operation is timed out.
*`-http-write-timeout` - The number of seconds before an HTTP write operation is timed out.
*`-bind-addr` - The listening hostname for client communication. Defaults to advertised IP.
*`-peers` - A comma separated list of peers in the cluster (i.e `"203.0.113.101:7001,203.0.113.102:7001"`).
*`-peers-file` - The file path containing a comma separated list of peers in the cluster.
*`-ca-file` - The path of the client CAFile. Enables client cert authentication when present.
*`-cert-file` - The cert file of the client.
*`-key-file` - The key file of the client.
*`-config` - The path of the etcd configuration file. Defaults to `/etc/etcd/etcd.conf`.
*`-cors` - A comma separated white list of origins for cross-origin resource sharing.
*`-cpuprofile` - The path to a file to output CPU profile data. Enables CPU profiling when present.
*`-data-dir` - The directory to store log and snapshot. Defaults to the current working directory.
*`-max-result-buffer` - The max size of result buffer. Defaults to `1024`.
*`-max-retry-attempts` - The max retry attempts when trying to join a cluster. Defaults to `3`.
*`-peer-addr` - The advertised public hostname:port for server communication. Defaults to `127.0.0.1:7001`.
*`-peer-bind-addr` - The listening hostname for server communication. Defaults to advertised IP.
*`-peer-ca-file` - The path of the CAFile. Enables client/peer cert authentication when present.
*`-peer-cert-file` - The cert file of the server.
*`-peer-key-file` - The key file of the server.
*`-peer-election-timeout` - The number of milliseconds to wait before the leader is declared unhealthy.
*`-peer-heartbeat-interval` - The number of milliseconds in between heartbeat requests
*`-snapshot=false` - Disable log snapshots. Defaults to `true`.
*`-cluster-active-size` - The expected number of instances participating in the consensus protocol. Only applied if the etcd instance is the first peer in the cluster.
*`-cluster-remove-delay` - The number of seconds before one node is removed from the cluster since it cannot be connected at all. Only applied if the etcd instance is the first peer in the cluster.
*`-cluster-sync-interval` - The number of seconds between synchronization for standby-mode instance with the cluster. Only applied if the etcd instance is the first peer in the cluster.
*`-v` - Enable verbose logging. Defaults to `false`.
*`-vv` - Enable very verbose logging. Defaults to `false`.
*`-version` - Print the version and exit.
## Configuration File
The etcd configuration file is written in [TOML](https://github.com/mojombo/toml)
Diagnosing issues in a distributed application is hard.
etcd will help as much as it can - just enable these debug features using the CLI flag `-trace=*` or the config option `trace=*`.
## Logging
Log verbosity can be increased to the max using either the `-vvv` CLI flag or the `very_very_verbose=true` config option.
The only supported logging mode is to stdout.
## Metrics
etcd itself can generate a set of metrics.
These metrics represent many different internal data points that can be helpful when debugging etcd servers.
#### Metrics reference
Each individual metric name is prefixed with `etcd.<NAME>`, where \<NAME\> is the configured name of the etcd server.
*`timer.appendentries.handle`: amount of time a peer takes to process an AppendEntriesRequest from the POV of the peer itself
*`timer.peer.<PEER>.heartbeat`: amount of time a peer heartbeat operation takes from the POV of the leader that initiated that operation for peer \<PEER\>
*`timer.command.<COMMAND>`: amount of time a given command took to be processed through the local server's raft state machine. This does not include time waiting on locks.
#### Fetching metrics over HTTP
Once tracing has been enabled on a given etcd server, all metric data is available at the server's `/debug/metrics` HTTP endpoint (i.e. `http://127.0.0.1:4001/debug/metrics`).
Executing a GET HTTP command against the metrics endpoint will yield the current state of all metrics in the etcd server.
#### Sending metrics to Graphite
etcd supports [Graphite's Carbon plaintext protocol](https://graphite.readthedocs.org/en/latest/feeding-carbon.html#the-plaintext-protocol) - a TCP wire protocol designed for shipping metric data to an aggregator.
To send metrics to a Graphite endpoint using this protocol, use of the `-graphite-host` CLI flag or the `graphite_host` config option (i.e. `graphite_host=172.17.0.19:2003`).
See an [example graphite deploy script](https://github.com/coreos/etcd/contrib/graphite).
#### Generating additional metrics with Collectd
[Collectd](http://collectd.org/documentation.shtml) gathers metrics from the host running etcd.
While these aren't metrics generated by etcd itself, it can be invaluable to compare etcd's view of the world to that of a separate process running next to etcd.
See an [example collectd deploy script](https://github.com/coreos/etcd/contrib/collectd).
## Profiling
etcd exposes profiling information from the Go pprof package over HTTP.
The basic browsable interface is served by etcd at the `/debug/pprof` HTTP endpoint (i.e. `http://127.0.0.1:4001/debug/pprof`).
For more information on using profiling tools, see http://blog.golang.org/profiling-go-programs.
**NOTE**: In the following examples you need to ensure that the `./bin/etcd` is identical to the `./bin/etcd` that you are targeting (same git hash, arch, platform, etc).
#### Heap memory profile
```
go tool pprof ./bin/etcd http://127.0.0.1:4001/debug/pprof/heap
```
#### CPU profile
```
go tool pprof ./bin/etcd http://127.0.0.1:4001/debug/pprof/profile
```
#### Blocked goroutine profile
```
go tool pprof ./bin/etcd http://127.0.0.1:4001/debug/pprof/block
Peer discovery uses the following sources in this order: log data in `-data-dir`, `-discovery` and `-peers`.
If log data is provided, etcd will concatenate possible peers from three sources: the log data, the `-discovery` option, and `-peers`. Then it tries to join cluster through them one by one. If all connection attempts fail (which indicates that the majority of the cluster is currently down), it will restart itself based on the log data, which helps the cluster to recover from a full outage.
Without log data, the instance is assumed to be a brand new one. If possible targets are provided by `-discovery` and `-peers`, etcd will make a best effort attempt to join them, and if none is reachable it will exit. Otherwise, if no `-discovery` or `-peers` option is provided, a new cluster will always be started.
This ensures that users can always restart the node safely with the same command (without --force), and etcd will either reconnect to the old cluster if it is still running or recover its cluster from a outage.
Adding peers in an etcd cluster adds network, CPU, and disk overhead to the leader since each one requires replication.
Peers primarily provide resiliency in the event of a leader failure but the benefit of more failover nodes decreases as the cluster size increases.
A lightweight alternative is the standby.
Standbys are a way for an etcd node to forward requests along to the cluster but the standbys are not part of the Raft cluster themselves.
This provides an easier API for local applications while reducing the overhead required by a regular peer node.
Standbys also act as standby nodes in the event that a peer node in the cluster has not recovered after a long duration.
## Configuration Parameters
There are three configuration parameters used by standbys: active size, remove delay and standby sync interval.
The active size specifies a target size for the number of peers in the cluster.
If there are not enough peers to meet the active size, standbys will send join requests until the peer count is equal to the active size.
If there are more peers than the target active size then peers are removed by the leader and will become standbys.
The remove delay specifies how long the cluster should wait before removing a dead peer.
By default this is 30 minutes.
If a peer is inactive for 30 minutes then the peer is removed.
The standby sync interval specifies the synchronization interval of standbys with the cluster.
By default this is 5 seconds.
After each interval, standbys synchronize information with cluster.
## Logical Workflow
### Start a etcd machine
#### Main logic
```
If find existing standby cluster info:
Goto standby loop
Find cluster as required
If determine to start peer server:
Goto peer loop
Else:
Goto standby loop
Peer loop:
Start peer mode
If running:
Wait for stop
Goto standby loop
Standby loop:
Start standby mode
If running:
Wait for stop
Goto peer loop
```
#### [Cluster finding logic][cluster-finding.md]
#### Join request logic:
```
Fetch machine info
If cannot match version:
return false
If active size <= peer count:
return false
If it has existed in the cluster:
return true
If join request fails:
return false
return true
```
**Note**
1. [TODO] The running mode cannot be determined by log, because the log may be outdated. But the log could be used to estimate its state.
2. Even if sync cluster fails, it will restart still for recovery from full outage.
#### Peer mode start logic
```
Start raft server
Start other helper routines
```
#### Peer mode auto stop logic
```
When removed from the cluster:
Stop raft server
Stop other helper routines
```
#### Standby mode run logic
```
Loop:
Sleep for some time
Sync cluster, and write cluster info into disk
Check active size and send join request if needed
If succeed:
Clear cluster info from disk
Return
```
#### Serve Requests as Standby
Return '404 Page Not Found' always on peer address. This is because peer address is used for raft communication and cluster management, which should not be used in standby mode.
Serve requests from client:
```
Redirect all requests to client URL of leader
```
**Note**
1. The leader here implies the one in raft cluster when doing the latest successful synchronization.
2. [IDEA] We could extend HTTP Redirect to multiple possible targets.
### Join Request Handling
```
If machine has existed in the cluster:
Return
If peer count < active size:
Add peer
Increase peer count
```
### Remove Request Handling
```
If machine exists in the cluster:
Remove peer
Decrease peer count
```
## Cluster Monitor Logic
### Active Size Monitor:
This is only run by current cluster leader.
```
Loop:
Sleep for some time
If peer count > active size:
Remove randomly selected peer
```
### Peer Activity Monitor
This is only run by current cluster leader.
```
Loop:
Sleep for some time
For each peer:
If peer last activity time > remove delay:
Remove the peer
Goto Loop
```
## Cluster Cases
### Create Cluster with Thousands of Instances
First few machines run in peer mode.
All the others check the status of the cluster and run in standby mode.
### Recover from full outage
Machines with log data restart with join failure.
Machines in peer mode recover heartbeat between each other.
Machines in standby mode always sync the cluster. If sync fails, it uses the first address from data log as redirect target.
### Kill one peer machine
Leader of the cluster lose the connection with the peer.
When the time exceeds remove delay, it removes the peer from the cluster.
Machine in standby mode finds one available place of the cluster. It sends join request and joins the cluster.
**Note**
1. [TODO] Machine which was divided from majority and was removed from the cluster will distribute running of the cluster if the new node uses the same name.
### Kill one standby machine
No change for the cluster.
## Cons
1. New instance cannot join immediately after one peer is kicked out of the cluster, because the leader doesn't know the info about the standby instances.
2. It may introduce join collision
3. Cluster needs a good interval setting to balance the join delay and join collision.
## Future Attack Plans
1. Based on heartbeat miss and remove delay, standby could adjust its next check time.
2. Preregister the promotion target when heartbeat miss happens.
3. Get the estimated cluster size from the check happened in the sync interval, and adjust sync interval dynamically.
4. Accept join requests based on active size and alive peers.
etcd v3 uses [gRPC][grpc] for its messaging protocol. The etcd project includes a gRPC-based [Go client][go-client] and a command line utility, [etcdctl][etcdctl], for communicating with an etcd cluster through gRPC. For languages with no gRPC support, etcd provides a JSON [grpc-gateway][grpc-gateway]. This gateway serves a RESTful proxy that translates HTTP/JSON requests into gRPC messages.
## Using grpc-gateway
The gateway accepts a [JSON mapping][json-mapping] for etcd's [protocol buffer][api-ref] message definitions. Note that `key` and `value` fields are defined as byte arrays and therefore must be base64 encoded in JSON.
```bash
<<COMMENT
https://www.base64encode.org/
foo is 'Zm9v' in Base64
bar is 'YmFy'
COMMENT
curl -L http://localhost:2379/v3alpha/kv/put \
-X POST -d '{"key": "Zm9v", "value": "YmFy"}'
curl -L http://localhost:2379/v3alpha/kv/range \
-X POST -d '{"key": "Zm9v"}'
```
## Swagger
Generated [Swagger][swagger] API definitions can be found at [rpc.swagger.json][swagger-doc].
| RoleGrantPermission | AuthRoleGrantPermissionRequest | AuthRoleGrantPermissionResponse | RoleGrantPermission grants a permission of a specified key or range to a specified role. |
| RoleRevokePermission | AuthRoleRevokePermissionRequest | AuthRoleRevokePermissionResponse | RoleRevokePermission revokes a key or range permission of a specified role. |
##### service `Cluster` (etcdserver/etcdserverpb/rpc.proto)
| Method | Request Type | Response Type | Description |
| Range | RangeRequest | RangeResponse | Range gets the keys in the range from the key-value store. |
| Put | PutRequest | PutResponse | Put puts the given key into the key-value store. A put request increments the revision of the key-value store and generates one event in the event history. |
| DeleteRange | DeleteRangeRequest | DeleteRangeResponse | DeleteRange deletes the given range from the key-value store. A delete request increments the revision of the key-value store and generates a delete event in the event history for every deleted key. |
| Txn | TxnRequest | TxnResponse | Txn processes multiple requests in a single transaction. A txn request increments the revision of the key-value store and generates events with the same revision for every completed request. It is not allowed to modify the same key several times within one txn. |
| Compact | CompactionRequest | CompactionResponse | Compact compacts the event history in the etcd key-value store. The key-value store should be periodically compacted or the event history will continue to grow indefinitely. |
##### service `Lease` (etcdserver/etcdserverpb/rpc.proto)
| Method | Request Type | Response Type | Description |
| LeaseGrant | LeaseGrantRequest | LeaseGrantResponse | LeaseGrant creates a lease which expires if the server does not receive a keepAlive within a given time to live period. All keys attached to the lease will be expired and deleted if the lease expires. Each expired key generates a delete event in the event history. |
| LeaseRevoke | LeaseRevokeRequest | LeaseRevokeResponse | LeaseRevoke revokes a lease. All keys attached to the lease will expire and be deleted. |
| LeaseKeepAlive | LeaseKeepAliveRequest | LeaseKeepAliveResponse | LeaseKeepAlive keeps the lease alive by streaming keep alive requests from the client to the server and streaming keep alive responses from the server to the client. |
##### service `Maintenance` (etcdserver/etcdserverpb/rpc.proto)
| Method | Request Type | Response Type | Description |
| Status | StatusRequest | StatusResponse | Status gets the status of the member. |
| Defragment | DefragmentRequest | DefragmentResponse | Defragment defragments a member's backend database to recover storage space. |
| Hash | HashRequest | HashResponse | Hash returns the hash of the local KV state for consistency checking purpose. This is designed for testing; do not use this in production when there are ongoing transactions. |
| Snapshot | SnapshotRequest | SnapshotResponse | Snapshot sends a snapshot of the entire backend from a member over a stream to a client. |
##### service `Watch` (etcdserver/etcdserverpb/rpc.proto)
| Method | Request Type | Response Type | Description |
| Watch | WatchRequest | WatchResponse | Watch watches for events happening or that have happened. Both input and output are streams; the input stream is for creating and canceling watchers and the output stream sends events. One watch RPC can watch on multiple key ranges, streaming events for several watches at once. The entire event history can be watched starting from the last compaction revision. |
default, used to query if any alarm is active space quota is exhausted
| Field | Description | Type |
| ----- | ----------- | ---- |
| action | action is the kind of alarm request to issue. The action may GET alarm statuses, ACTIVATE an alarm, or DEACTIVATE a raised alarm. | AlarmAction |
| memberID | memberID is the ID of the member associated with the alarm. If memberID is 0, the alarm request covers all members. | uint64 |
| alarm | alarm is the type of alarm to consider for this request. | AlarmType |
CompactionRequest compacts the key-value store up to a given revision. All superseded keys with a revision less than the compaction revision will be removed.
| Field | Description | Type |
| ----- | ----------- | ---- |
| revision | revision is the key-value store revision for the compaction operation. | int64 |
| physical | physical is set so the RPC will wait until the compaction is physically applied to the local database such that compacted entries are totally removed from the backend database. | bool |
| key | key is the first key to delete in the range. | bytes |
| range_end | range_end is the key following the last key to delete for the range [key, range_end). If range_end is not given, the range is defined to contain only the key argument. If range_end is '\0', the range is all keys greater than or equal to the key argument. | bytes |
| ID | ID is the member ID for this member. | uint64 |
| name | name is the human-readable name of the member. If the member is not started, the name will be an empty string. | string |
| peerURLs | peerURLs is the list of URLs the member exposes to the cluster for communication. | (slice of) string |
| clientURLs | clientURLs is the list of URLs the member exposes to clients for communication. If the member is not started, clientURLs will be empty. | (slice of) string |
| key | default, no sorting lowest target value first highest target value first key is the first key for the range. If range_end is not given, the request only looks up key. | bytes |
| range_end | range_end is the upper bound on the requested range [key, range_end). If range_end is '\0', the range is all keys >= key. If the range_end is one bit larger than the given key, then the range requests get the all keys with the prefix (the given key). If both key and range_end are '\0', then range requests returns all keys. | bytes |
| limit | limit is a limit on the number of keys returned for the request. | int64 |
| revision | revision is the point-in-time of the key-value store to use for the range. If revision is less or equal to zero, the range is over the newest key-value store. If the revision has been compacted, ErrCompacted is returned as a response. | int64 |
| sort_order | sort_order is the order for returned sorted results. | SortOrder |
| sort_target | sort_target is the key-value field to use for sorting. | SortTarget |
| serializable | serializable sets the range request to use serializable member-local reads. Range requests are linearizable by default; linearizable requests have higher latency and lower throughput than serializable requests but reflect the current consensus of the cluster. For better performance, in exchange for possible stale reads, a serializable range request is served locally without needing to reach consensus with other nodes in the cluster. | bool |
| header | header has the current key-value store information. The first header in the snapshot stream indicates the point in time of the snapshot. | ResponseHeader |
| remaining_bytes | remaining_bytes is the number of blob bytes to be sent after this message | uint64 |
| blob | blob contains the next chunk of the snapshot in the snapshot stream. | bytes |
From google paxosdb paper: Our implementation hinges around a powerful primitive which we call MultiOp. All other database operations except for iteration are implemented as a single call to MultiOp. A MultiOp is applied atomically and consists of three components: 1. A list of tests called guard. Each test in guard checks a single entry in the database. It may check for the absence or presence of a value, or compare with a given value. Two different tests in the guard may apply to the same or different entries in the database. All tests in the guard are applied and MultiOp returns the results. If all tests are true, MultiOp executes t op (see item 2 below), otherwise it executes f op (see item 3 below). 2. A list of database operations called t op. Each operation in the list is either an insert, delete, or lookup operation, and applies to a single database entry. Two different operations in the list may apply to the same or different entries in the database. These operations are executed if guard evaluates to true. 3. A list of database operations called f op. Like t op, but executed if guard evaluates to false.
| Field | Description | Type |
| ----- | ----------- | ---- |
| compare | compare is a list of predicates representing a conjunction of terms. If the comparisons succeed, then the success requests will be processed in order, and the response will contain their respective responses in order. If the comparisons fail, then the failure requests will be processed in order, and the response will contain their respective responses in order. | (slice of) Compare |
| success | success is a list of requests which will be applied when compare evaluates to true. | (slice of) RequestOp |
| failure | failure is a list of requests which will be applied when compare evaluates to false. | (slice of) RequestOp |
| succeeded | succeeded is set to true if the compare evaluated to true or false otherwise. | bool |
| responses | responses is a list of responses corresponding to the results from applying success if succeeded is true or failure if succeeded is false. | (slice of) ResponseOp |
| key | key is the key to register for watching. | bytes |
| range_end | range_end is the end of the range [key, range_end) to watch. If range_end is not given, only the key argument is watched. If range_end is equal to '\0', all keys greater than or equal to the key argument are watched. | bytes |
| start_revision | start_revision is an optional revision to watch from (inclusive). No start_revision is "now". | int64 |
| progress_notify | progress_notify is set so that the etcd server will periodically send a WatchResponse with no events to the new watcher if there are no recent events. It is useful when clients wish to recover a disconnected watcher starting from a recent known revision. The etcd server may decide how often it will send notifications based on current load. | bool |
| watch_id | watch_id is the ID of the watcher that corresponds to the response. | int64 |
| created | created is set to true if the response is for a create watch request. The client should record the watch_id and expect to receive events for the created watcher from the same stream. All events sent to the created watcher will attach with the same watch_id. | bool |
| canceled | canceled is set to true if the response is for a cancel watch request. No further events will be sent to the canceled watcher. | bool |
| compact_revision | compact_revision is set to the minimum index if a watcher tries to watch at a compacted index. This happens when creating a watcher at a compacted revision or the watcher cannot catch up with the progress of the key-value store. The client should treat the watcher as canceled and should not try to create any watcher with the same start_revision again. | int64 |
| events | | (slice of) mvccpb.Event |
##### message `Event` (mvcc/mvccpb/kv.proto)
| Field | Description | Type |
| ----- | ----------- | ---- |
| type | type is the kind of event. If type is a PUT, it indicates new data has been stored to the key. If type is a DELETE, it indicates the key was deleted. | EventType |
| kv | kv holds the KeyValue for the event. A PUT event contains current kv pair. A PUT event with kv.Version=1 indicates the creation of a key. A DELETE/EXPIRE event contains the deleted key with its modification revision set to the revision of deletion. | KeyValue |
##### message `KeyValue` (mvcc/mvccpb/kv.proto)
| Field | Description | Type |
| ----- | ----------- | ---- |
| key | key is the key in bytes. An empty key is not allowed. | bytes |
| create_revision | create_revision is the revision of last creation on this key. | int64 |
| mod_revision | mod_revision is the revision of last modification on this key. | int64 |
| version | version is the version of the key. A deletion resets the version to zero and any modification of the key increases its version. | int64 |
| value | value is the value held by the key, in bytes. | bytes |
| lease | lease is the ID of the lease that attached to key. When the attached lease expires, the key will be deleted. If lease is 0, then no lease is attached to the key. | int64 |
For the most part, the etcd project is stable, but we are still moving fast! We believe in the release fast philosophy. We want to get early feedback on features still in development and stabilizing. Thus, there are, and will be more, experimental features and APIs. We plan to improve these features based on the early feedback from the community, or abandon them if there is little interest, in the next few releases. If you are running a production system, please do not rely on any experimental features or APIs.
## The current experimental API/features are:
- v3 auth API: expect to be stable in 3.1 release
- etcd gateway: expect to be stable in 3.1 release
Users mostly interact with etcd by putting or getting the value of a key. This section describes how to do that by using etcdctl, a command line tool for interacting with etcd server. The concepts described here should apply to the gRPC APIs or client library APIs.
By default, etcdctl talks to the etcd server with the v2 API for backward compatibility. For etcdctl to speak to etcd using the v3 API, the API version must be set to version 3 via the `ETCDCTL_API` environment variable.
``` bash
export ETCDCTL_API=3
```
## Write a key
Applications store keys into the etcd cluster by writing to keys. Every stored key is replicated to all etcd cluster members through the Raft protocol to achieve consistency and reliability.
Here is the command to set the value of key `foo` to `bar`:
``` bash
$ etcdctl put foo bar
OK
```
## Read keys
Applications can read values of keys from an etcd cluster. Queries may read a single key, or a range of keys.
Suppose the etcd cluster has stored the following keys:
```
foo = bar
foo1 = bar1
foo3 = bar3
```
Here is the command to read the value of key `foo`:
```bash
$ etcdctl get foo
foo
bar
```
Here is the command to range over the keys from `foo` to `foo9`:
```bash
$ etcdctl get foo foo9
foo
bar
foo1
bar1
foo3
bar3
```
## Read past version of keys
Applications may want to read superseded versions of a key. For example, an application may wish to roll back to an old configuration by accessing an earlier version of a key. Alternatively, an application may want a consistent view over multiple keys through multiple requests by accessing key history.
Since every modification to the etcd cluster key-value store increments the global revision of an etcd cluster, an application can read superseded keys by providing an older etcd revision.
Suppose an etcd cluster already has the following keys:
``` bash
$ etcdctl put foo bar # revision = 2
$ etcdctl put foo1 bar1 # revision = 3
$ etcdctl put foo bar_new # revision = 4
$ etcdctl put foo1 bar1_new # revision = 5
```
Here are an example to access the past versions of keys:
```bash
$ etcdctl get foo foo9 # access the most recent versions of keys
foo
bar_new
foo1
bar1_new
$ etcdctl get --rev=4 foo foo9 # access the versions of keys at revision 4
foo
bar_new
foo1
bar1
$ etcdctl get --rev=3 foo foo9 # access the versions of keys at revision 3
foo
bar
foo1
bar1
$ etcdctl get --rev=2 foo foo9 # access the versions of keys at revision 2
foo
bar
$ etcdctl get --rev=1 foo foo9 # access the versions of keys at revision 1
```
## Delete keys
Applications can delete a key or a range of keys from an etcd cluster.
Here is the command to delete key `foo`:
```bash
$ etcdctl del foo
1 # one key is deleted
```
Here is the command to delete keys ranging from `foo` to `foo9`:
```bash
$ etcdctl del foo foo9
2 # two keys are deleted
```
## Watch key changes
Applications can watch on a key or a range of keys to monitor for any updates.
Here is the command to watch on key `foo`:
```bash
$ etcdctl watch foo
# in another terminal: etcdctl put foo bar
foo
bar
```
Here is the command to watch on a range key from `foo` to `foo9`:
```bash
$ etcdctl watch foo foo9
# in another terminal: etcdctl put foo bar
foo
bar
# in another terminal: etcdctl put foo1 bar1
foo1
bar1
```
## Watch historical changes of keys
Applications may want to watch for historical changes of keys in etcd. For example, an application may wish to receive all the modifications of a key; if the application stays connected to etcd, then `watch` is good enough. However, if the application or etcd fails, a change may happen during the failure, and the application will not receive the update in real time. To guarantee the update is delivered, the application must be able to watch for historical changes to keys. To do this, an application can specify a historical revision on a watch, just like reading past version of keys.
Suppose we finished the following sequence of operations:
``` bash
etcdctl put foo bar # revision = 2
etcdctl put foo1 bar1 # revision = 3
etcdctl put foo bar_new # revision = 4
etcdctl put foo1 bar1_new # revision = 5
```
Here is an example to watch the historical changes:
```bash
# watch for changes on key `foo` since revision 2
$ etcdctl watch --rev=2 foo
PUT
foo
bar
PUT
foo
bar_new
# watch for changes on key `foo` since revision 3
$ etcdctl watch --rev=3 foo
PUT
foo
bar_new
```
## Compacted revisions
As we mentioned, etcd keeps revisions so that applications can read past versions of keys. However, to avoid accumulating an unbounded amount of history, it is important to compact past revisions. After compacting, etcd removes historical revisions, releasing resources for future use. All superseded data with revisions before the compacted revision will be unavailable.
Here is the command to compact the revisions:
```bash
$ etcdctl compact 5
compacted revision 5
# any revisions before the compacted one are not accessible
$ etcdctl get --rev=4 foo
Error: rpc error: code = 11 desc = etcdserver: mvcc: required revision has been compacted
```
## Grant leases
Applications can grant leases for keys from an etcd cluster. When a key is attached to a lease, its lifetime is bound to the lease's lifetime which in turn is governed by a time-to-live (TTL). Each lease has a minimum time-to-live (TTL) value specified by the application at grant time. The lease's actual TTL value is at least the minimum TTL and is chosen by the etcd cluster. Once a lease's TTL elapses, the lease expires and all attached keys are deleted.
Here is the command to grant a lease:
```
# grant a lease with 10 second TTL
$ etcdctl lease grant 10
lease 32695410dcc0ca06 granted with TTL(10s)
# attach key foo to lease 32695410dcc0ca06
$ etcdctl put --lease=32695410dcc0ca06 foo bar
OK
```
## Revoke leases
Applications revoke leases by lease ID. Revoking a lease deletes all of its attached keys.
Suppose we finished the following sequence of operations:
```
$ etcdctl lease grant 10
lease 32695410dcc0ca06 granted with TTL(10s)
$ etcdctl put --lease=32695410dcc0ca06 foo bar
OK
```
Here is the command to revoke the same lease:
```
$ etcdctl lease revoke 32695410dcc0ca06
lease 32695410dcc0ca06 revoked
$ etcdctl get foo
# empty response since foo is deleted due to lease revocation
```
## Keep leases alive
Applications can keep a lease alive by refreshing its TTL so it does not expire.
Suppose we finished the following sequence of operations:
For testing and development deployments, the quickest and easiest way is to set up a local cluster. For a production deployment, refer to the [clustering][clustering] section.
## Local standalone cluster
Deploying an etcd cluster as a standalone cluster is straightforward. Start it with just one command:
```
$ ./etcd
...
```
The started etcd member listens on `localhost:2379` for client requests.
To interact with the started cluster by using etcdctl:
```
# use API version 3
$ export ETCDCTL_API=3
$ ./etcdctl put foo bar
OK
$ ./etcdctl get foo
bar
```
## Local multi-member cluster
A Procfile is provided to easily set up a local multi-member cluster. Start a multi-member cluster with a few commands:
```
# install goreman program to control Profile-based applications.
$ go get github.com/mattn/goreman
$ goreman -f Procfile start
...
```
The started members listen on `localhost:12379`, `localhost:22379`, and `localhost:32379` for client requests respectively.
To interact with the started cluster by using etcdctl:
```
# use API version 3
$ export ETCDCTL_API=3
$ etcdctl --write-out=table --endpoints=localhost:12379 member list
Discovery service protocol helps new etcd member to discover all other members in cluster bootstrap phase using a shared discovery URL.
Discovery service protocol is _only_ used in cluster bootstrap phase, and cannot be used for runtime reconfiguration or cluster monitoring.
The protocol uses a new discovery token to bootstrap one _unique_ etcd cluster. Remember that one discovery token can represent only one etcd cluster. As long as discovery protocol on this token starts, even if it fails halfway, it must not be used to bootstrap another etcd cluster.
The rest of this article will walk through the discovery process with examples that correspond to a self-hosted discovery cluster. The public discovery service, discovery.etcd.io, functions the same way, but with a layer of polish to abstract away ugly URLs, generate UUIDs automatically, and provide some protections against excessive requests. At its core, the public discovery service still uses an etcd cluster as the data store as described in this document.
## Protocol workflow
The idea of discovery protocol is to use an internal etcd cluster to coordinate bootstrap of a new cluster. First, all new members interact with discovery service and help to generate the expected member list. Then each new member bootstraps its server using this list, which performs the same functionality as -initial-cluster flag.
In the following example workflow, we will list each step of protocol in curl format for ease of understanding.
By convention the etcd discovery protocol uses the key prefix `_etcd/registry`. If `http://example.com` hosts an etcd cluster for discovery service, a full URL to discovery keyspace will be `http://example.com/v2/keys/_etcd/registry`. We will use this as the URL prefix in the example.
### Creating a new discovery token
Generate a unique token that will identify the new cluster. This will be used as a unique prefix in discovery keyspace in the following steps. An easy way to do this is to use `uuidgen`:
```
UUID=$(uuidgen)
```
### Specifying the expected cluster size
The discovery token expects a cluster size that must be specified. The size is used by the discovery service to know when it has found all members that will initially form the cluster.
```
curl -X PUT http://example.com/v2/keys/_etcd/registry/${UUID}/_config/size -d value=${cluster_size}
```
Usually the cluster size is 3, 5 or 7. Check [optimal cluster size][cluster-size] for more details.
### Bringing up etcd processes
Given the discovery URL, use it as `-discovery` flag and bring up etcd processes. Every etcd process will follow this next few steps internally if given a `-discovery` flag.
### Registering itself
The first thing for etcd process is to register itself into the discovery URL as a member. This is done by creating member ID as a key in the discovery URL.
```
curl -X PUT http://example.com/v2/keys/_etcd/registry/${UUID}/${member_id}?prevExist=false -d value="${member_name}=${member_peer_url_1}&${member_name}=${member_peer_url_2}"
```
### Checking the status
It checks the expected cluster size and registration status in discovery URL, and decides what the next action is.
```
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}/_config/size
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}
```
If registered members are still not enough, it will wait for left members to appear.
If the number of registered members is bigger than the expected size N, it treats the first N registered members as the member list for the cluster. If the member itself is in the member list, the discovery procedure succeeds and it fetches all peers through the member list. If it is not in the member list, the discovery procedure finishes with the failure that the cluster has been full.
In etcd implementation, the member may check the cluster status even before registering itself. So it could fail quickly if the cluster has been full.
### Waiting for all members
The wait process is described in detail in the [etcd API documentation][api].
```
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}?wait=true&waitIndex=${current_etcd_index}
```
It keeps waiting until finding all members.
## Public discovery service
CoreOS Inc. hosts a public discovery service at https://discovery.etcd.io/ , which provides some nice features for ease of use.
### Mask key prefix
Public discovery service will redirect `https://discovery.etcd.io/${UUID}` to etcd cluster behind for the key at `/v2/keys/_etcd/registry`. It masks register key prefix for short and readable discovery url.
### Get new token
```
GET /new
Sent query:
size=${cluster_size}
Possible status codes:
200 OK
400 Bad Request
200 Body:
generated discovery url
```
The generation process in the service follows the steps from [Creating a New Discovery Token][new-discovery-token] to [Specifying the Expected Cluster Size][expected-cluster-size].
### Check discovery status
```
GET /${UUID}
```
The status for this discovery token, including the machines that have been registered, can be checked by requesting the value of the UUID.
### Open-source repository
The repository is located at https://github.com/coreos/discovery.etcd.io. It could be used to build a custom discovery service.
etcd uses the [capnslog][capnslog] library for logging application output categorized into *levels*. A log message's level is determined according to these conventions:
* Error: Data has been lost, a request has failed for a bad reason, or a required resource has been lost
* Examples:
* A failure to allocate disk space for WAL
* Warning: (Hopefully) Temporary conditions that may cause errors, but may work fine. A replica disappearing (that may reconnect) is a warning.
* Examples:
* Failure to send raft message to a remote peer
* Failure to receive heartbeat message within the configured election timeout
* Notice: Normal, but important (uncommon) log information.
* Examples:
* Add a new node into the cluster
* Add a new user into auth subsystem
* Info: Normal, working log information, everything is fine, but helpful notices for auditing or common operations.
* Examples:
* Startup configuration
* Start to do snapshot
* Debug: Everything is still fine, but even common operations may be logged, and less helpful but more quantity of notices.
The guide talks about how to release a new version of etcd.
The procedure includes some manual steps for sanity checking but it can probably be further scripted. Please keep this document up-to-date if making changes to the release process.
## Prepare release
Set desired version as environment variable for following steps. Here is an example to release 2.3.0:
```
export VERSION=v2.3.0
export PREV_VERSION=v2.2.5
```
All releases version numbers follow the format of [semantic versioning 2.0.0](http://semver.org/).
### Major, minor version release, or its pre-release
- Ensure the relevant milestone on GitHub is complete. All referenced issues should be closed, or moved elsewhere.
- Remove this release from [roadmap](https://github.com/coreos/etcd/blob/master/ROADMAP.md), if necessary.
- Ensure the latest upgrade documentation is available.
- Bump [hardcoded MinClusterVerion in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L29), if necessary.
- Add feature capability maps for the new version, if necessary.
### Patch version release
- Discuss about commits that are backported to the patch release. The commits should not include merge commits.
- Cherry-pick these commits starting from the oldest one into stable branch.
## Write release note
- Write introduction for the new release. For example, what major bug we fix, what new features we introduce or what performance improvement we make.
- Write changelog for the last release. ChangeLog should be straightforward and easy to understand for the end-user.
- Put `[GH XXXX]` at the head of change line to reference Pull Request that introduces the change. Moreover, add a link on it to jump to the Pull Request.
## Tag version
- Bump [hardcoded Version in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L30) to the latest version `${VERSION}`.
- Ensure all tests on CI system are passed.
- Manually check etcd is buildable in Linux, Darwin and Windows.
- Manually check upgrade etcd cluster of previous minor version works well.
- Manually check new features work well.
- Add a signed tag through `git tag -s ${VERSION}`.
- Sanity check tag correctness through `git show tags/$VERSION`.
- Push the tag to GitHub through `git push origin tags/$VERSION`. This assumes `origin` corresponds to "https://github.com/coreos/etcd".
## Build release binaries and images
- Ensure `actool` is available, or installing it through `go get github.com/appc/spec/actool`.
- Ensure `docker` is available.
Run release script in root directory:
```
./scripts/release.sh ${VERSION}
```
It generates all release binaries and images under directory ./release.
## Sign binaries and images
etcd project key must be used to sign the generated binaries and images.`$SUBKEYID` is the key ID of etcd project Yubikey. Connect the key and run `gpg2 --card-status` to get the ID.
The following commands are used for public release sign:
```
cd release
for i in etcd-*{.zip,.tar.gz}; do gpg2 --default-key $SUBKEYID --armor --output ${i}.asc --detach-sign ${i}; done
for i in etcd-*{.zip,.tar.gz}; do gpg2 --verify ${i}.asc ${i}; done
```
The public key for GPG signing can be found at [CoreOS Application Signing Key](https://coreos.com/security/app-signing-key)
## Publish release page in GitHub
- Set release title as the version name.
- Follow the format of previous release pages.
- Attach the generated binaries, aci image and signatures.
- Select whether it is a pre-release.
- Publish the release!
## Publish docker image in Quay.io
- Push docker image:
```
docker login quay.io
docker push quay.io/coreos/etcd:${VERSION}
```
- Add `latest` tag to the new image on [quay.io](https://quay.io/repository/coreos/etcd?tag=latest&tab=tags) if this is a stable release.
## Announce to the etcd-dev Googlegroup
- Follow the format of [previous release emails](https://groups.google.com/forum/#!forum/etcd-dev).
- Make sure to include a list of authors that contributed since the previous release - something like the following might be handy:
- Create new stable branch through `git push origin ${VERSION_MAJOR}.${VERSION_MINOR}` if this is a major stable release. This assumes `origin` corresponds to "https://github.com/coreos/etcd".
- Bump [hardcoded Version in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L30) to the version `${VERSION}+git`.
Starting a new etcd cluster can be painful since each machine needs to know of at least one live machine in the cluster. If you are trying to bring up a new cluster all at once, say using an AWS cloud formation, you also need to coordinate who will be the initial cluster leader. The discovery protocol uses an existing running etcd cluster to start a second etcd cluster.
To use this feature you add the command line flag `-discovery` to your etcd args. In this example we will use `http://example.com/v2/keys/_etcd/registry` as the URL prefix.
## The Protocol
By convention the etcd discovery protocol uses the key prefix `_etcd/registry`. A full URL to the keyspace will be `http://example.com/v2/keys/_etcd/registry`.
### Creating a New Cluster
Generate a unique token that will identify the new cluster. This will be used as a key prefix in the following steps. An easy way to do this is to use uuidgen:
```
UUID=$(uuidgen)
```
### Bringing up Machines
Now that you have your cluster ID you can start bringing up machines. Every machine will follow this protocol internally in etcd if given a `-discovery`.
### Registering your Machine
The first thing etcd must do is register your machine. This is done by using the machine name (from the `-name` arg) and posting it with a long TTL to the given key.
```
curl -X PUT "http://example.com/v2/keys/_etcd/registry/${UUID}/${etcd_machine_name}?ttl=604800" -d value=${peer_addr}
```
### Discovering Peers
Now that this etcd machine is registered it must discover its peers.
But, the tricky bit of starting a new cluster is that one machine needs to assume the initial role of leader and will have no peers. To figure out if another machine has already started the cluster etcd needs to create the `_state` key and set its value to "started":
```
curl -X PUT "http://example.com/v2/keys/_etcd/registry/${UUID}/_state?prevExist=false" -d value=started
```
If this returns a `200 OK` response then this machine is the initial leader and should start with no peers configured. If, however, this returns a `412 Precondition Failed` then you need to find all of the registered peers:
```
curl -X GET "http://example.com/v2/keys/_etcd/registry/${UUID}?recursive=true"
The etcd performance benchmarks run etcd on 8 vCPU, 16GB RAM, 50GB SSD GCE instances, but any relatively modern machine with low latency storage and a few gigabytes of memory should suffice for most use cases. Applications with large v2 data stores will require more memory than a large v3 data store since data is kept in anonymous memory instead of memory mapped from a file. than For running etcd on a cloud provider, we suggest at least a medium instance on AWS or a standard-1 instance on GCE.
## Download the pre-built binary
The easiest way to get etcd is to use one of the pre-built release binaries which are available for OSX, Linux, Windows, appc, and Docker. Instructions for using these binaries are on the [GitHub releases page][github-release].
## Build the latest version
For those wanting to try the very latest version, build etcd from the `master` branch.
[Go](https://golang.org/) version 1.6+ (with HTTP2 support) is required to build the latest version of etcd.
Here are the commands to build an etcd binary from the `master` branch:
```
# go is required
$ go version
go version go1.6 darwin/amd64
# GOPATH should be set correctly
$ echo $GOPATH
/Users/example/go
$ mkdir -p $GOPATH/src/github.com/coreos
$ cd $GOPATH/src/github.com/coreos
$ git clone github.com:coreos/etcd.git
$ cd etcd
$ ./build
$ ./bin/etcd
...
```
## Test the installation
Check the etcd binary is built correctly by starting etcd and setting a key.
etcd is a distributed key-value store designed to reliably and quickly preserve and provide access to critical data. It enables reliable distributed coordination through distributed locking, leader elections, and write barriers. An etcd cluster is intended for high availability and permanent data storage and retrieval.
## Getting started
New etcd users and developers should get started by [downloading and building][download_build] etcd. After getting etcd, follow this [quick demo][demo] to see the basics of creating and working with an etcd cluster.
## Developing with etcd
The easiest way to get started using etcd as a distributed key-value store is to [set up a local cluster][local_cluster].
- [Setting up local clusters][local_cluster]
- [Interacting with etcd][interacting]
- [API references][api_ref]
- [gRPC gateway][api_grpc_gateway]
- [Experimental features and APIs][experimental]
## Operating etcd clusters
Administrators who need to create reliable and scalable key-value stores for the developers they support should begin with a [cluster on multiple machines][clustering].
Atomic *test and set* value to a file. If test succeeds, this operation will change the previous value of the file to the given value.
- If the prevValue is given, it will test against previous value of
the node.
- If the prevValue is empty, it will test if the node is not existing.
- If the prevValue is not empty, it will test if the prevValue is equal to the current value of the file.
- If the prevIndex is given, it will test if the create/last modified index of the node is equal to prevIndex.
- **Renew** (path, ttl)
Set the node's expiration time to (current time + ttl)
## ACL
### Theory
Etcd exports a Unix-like file system interface consisting of files and directories, collectively called nodes.
Each node has various meta-data, including three names of the access control lists used to control reading, writing and changing (change ACL names for the node).
We are storing the ACL names for nodes under a special *ACL* directory.
Each node has ACL name corresponding to one file within *ACL* dir.
Unless overridden, a node naturally inherits the ACL names of its parent directory on creation.
For each ACL name, it has three children: *R (Reading)*, *W (Writing)*, *C (Changing)*
Each permission is also a node. Under the node it contains the users who have this permission for the file referring to this ACL name.
All Responses from etcd API have a [response header][response_header] attached. The response header includes the metadata of the response.
```proto
messageResponseHeader{
uint64cluster_id=1;
uint64member_id=2;
int64revision=3;
uint64raft_term=4;
}
```
* Cluster_ID - the ID of the cluster that generates the response
* Member_ID - the ID of the member that generates the response
* Revision - the revision of the key-value store when the response is generated
* Raft_Term - the Raft term of the member when the response is generated
An application may read the Cluster_ID (Member_ID) field to ensure it is communicating with the intended cluster (member).
Applications can use the `Revision` to know the latest revision of the key-value store. This is especially useful when applications specify a historical revision to make time `travel query` and wishes to know the latest revision at the time of the request.
Applications can use `Raft_Term` to detect when the cluster completes a new leader election.
## Key-Value API
Key-Value API is used to manipulate key-value pairs stored inside etcd. The key-value API is defined as a [gRPC service][kv-service]. The Key-Value pair is defined as structured data in [protobuf format][kv-proto].
### Key-Value pair
A key-value pair is the smallest unit that the key-value API can manipulate. Each key-value pair has a number of fields:
```protobuf
messageKeyValue{
byteskey=1;
int64create_revision=2;
int64mod_revision=3;
int64version=4;
bytesvalue=5;
int64lease=6;
}
```
* Key - key in bytes. An empty key is not allowed.
* Value - value in bytes.
* Version - version is the version of the key. A deletion resets the version to zero and any modification of the key increases its version.
* Create_Revision - revision of the last creation on the key.
* Mod_Revision - revision of the last modification on the key.
* Lease - the ID of the lease attached to the key. If lease is 0, then no lease is attached to the key.
etcd is a consistent and durable key value store with mini-transaction(TODO: link to txn doc when we have it) support. The key value store is exposed through the KV APIs. etcd tries to ensure the strongest consistency and durability guarantees for a distributed system. This specification enumerates the KV API guarantees made by etcd.
### APIs to consider
* Read APIs
* range
* watch
* Write APIs
* put
* delete
* Combination (read-modify-write) APIs
* txn
### etcd specific definitions
#### Operation completed
An etcd operation is considered complete when it is committed through consensus, and therefore “executed” -- permanently stored -- by the etcd storage engine. The client knows an operation is completed when it receives a response from the etcd server. Note that the client may be uncertain about the status of an operation if it times out, or there is a network disruption between the client and the etcd member. etcd may also abort operations when there is a leader election. etcd does not send `abort` responses to clients’ outstanding requests in this event.
#### Revision
An etcd operation that modifies the key value store is assigned with a single increasing revision. A transaction operation might modifies the key value store multiple times, but only one revision is assigned. The revision attribute of a key value pair that modified by the operation has the same value as the revision of the operation. The revision can be used as a logical clock for key value store. A key value pair that has a larger revision is modified after a key value pair with a smaller revision. Two key value pairs that have the same revision are modified by an operation "concurrently".
### Guarantees provided
#### Atomicity
All API requests are atomic; an operation either completes entirely or not at all. For watch requests, all events generated by one operation will be in one watch response. Watch never observes partial events for a single operation.
#### Consistency
All API calls ensure [sequential consistency][seq_consistency], the strongest consistency guarantee available from distributed systems. No matter which etcd member server a client makes requests to, a client reads the same events in the same order. If two members complete the same number of operations, the state of the two members is consistent.
For watch operations, etcd guarantees to return the same value for the same key across all members for the same revision. For range operations, etcd has a similar guarantee for [linearized][Linearizability] access; serialized access may be behind the quorum state, so that the later revision is not yet available.
As with all distributed systems, it is impossible for etcd to ensure [strict consistency][strict_consistency]. etcd does not guarantee that it will return to a read the “most recent” value (as measured by a wall clock when a request is completed) available on any cluster member.
#### Isolation
etcd ensures [serializable isolation][serializable_isolation], which is the highest isolation level available in distributed systems. Read operations will never observe any intermediate data.
#### Durability
Any completed operations are durable. All accessible data is also durable data. A read will never return data that has not been made durable.
#### Linearizability
Linearizability (also known as Atomic Consistency or External Consistency) is a consistency level between strict consistency and sequential consistency.
For linearizability, suppose each operation receives a timestamp from a loosely synchronized global clock. Operations are linearized if and only if they always complete as though they were executed in a sequential order and each operation appears to complete in the order specified by the program. Likewise, if an operation’s timestamp precedes another, that operation must also precede the other operation in the sequence.
For example, consider a client completing a write at time point 1 (*t1*). A client issuing a read at *t2* (for *t2* > *t1*) should receive a value at least as recent as the previous write, completed at *t1*. However, the read might actually complete only by *t3*, and the returned value, current at *t2* when the read began, might be "stale" by *t3*.
etcd does not ensure linearizability for watch operations. Users are expected to verify the revision of watch responses to ensure correct ordering.
etcd ensures linearizability for all other operations by default. Linearizability comes with a cost, however, because linearized requests must go through the Raft consensus process. To obtain lower latencies and higher throughput for read requests, clients can configure a request’s consistency mode to `serializable`, which may access stale data with respect to quorum, but removes the performance penalty of linearized accesses' reliance on live consensus.
etcd is designed to reliably store infrequently updated data and provide reliable watch queries. etcd exposes previous versions of key-value pairs to support inexpensive snapshots and watch history events (“time travel queries”). A persistent, multi-version, concurrency-control data model is a good fit for these use cases.
etcd stores data in a multiversion [persistent][persistent-ds] key-value store. The persistent key-value store preserves the previous version of a key-value pair when its value is superseded with new data. The key-value store is effectively immutable; its operations do not update the structure in-place, but instead always generates a new updated structure. All past versions of keys are still accessible and watchable after modification. To prevent the data store from growing indefinitely over time from maintaining old versions, the store may be compacted to shed the oldest versions of superseded data.
### Logical view
The store’s logical view is a flat binary key space. The key space has a lexically sorted index on byte string keys so range queries are inexpensive.
The key space maintains multiple revisions. Each atomic mutative operation (e.g., a transaction operation may contain multiple operations) creates a new revision on the key space. All data held by previous revisions remains unchanged. Old versions of key can still be accessed through previous revisions. Likewise, revisions are indexed as well; ranging over revisions with watchers is efficient. If the store is compacted to recover space, revisions before the compact revision will be removed.
A key’s lifetime spans a generation. Each key may have one or multiple generations. Creating a key increments the generation of that key, starting at 1 if the key never existed. Deleting a key generates a key tombstone, concluding the key’s current generation. Each modification of a key creates a new version of the key. Once a compaction happens, any generation ended before the given revision will be removed and values set before the compaction revision except the latest one will be removed.
### Physical view
etcd stores the physical data as key-value pairs in a persistent [b+tree][b+tree]. Each revision of the store’s state only contains the delta from its previous revision to be efficient. A single revision may correspond to multiple keys in the tree.
The key of key-value pair is a 3-tuple (major, sub, type). Major is the store revision holding the key. Sub differentiates among keys within the same revision. Type is an optional suffix for special value (e.g., `t` if the value contains a tombstone). The value of the key-value pair contains the modification from previous revision, thus one delta from previous revision. The b+tree is ordered by key in lexical byte-order. Ranged lookups over revision deltas are fast; this enables quickly finding modifications from one specific revision to another. Compaction removes out-of-date keys-value pairs.
etcd also keeps a secondary in-memory [btree][btree] index to speed up range queries over keys. The keys in the btree index are the keys of the store exposed to user. The value is a pointer to the modification of the persistent b+tree. Compaction removes dead pointers.
This document defines the various terms used in etcd documentation, command line and source code.
### Node
## Node
Node is an instance of raft state machine.
It has a unique identification, and records other nodes' progress internally when it is the leader.
### Member
## Member
Member is an instance of etcd. It hosts a node, and provides service to clients.
### Cluster
## Cluster
Cluster consists of several members.
The node in each member follows raft consensus protocol to replicate logs. Cluster receives proposals from members, commits them and apply to local store.
### Peer
## Peer
Peer is another member of the same cluster.
### Client
## Proposal
A proposal is a request (for example a write request, a configuration change request) that needs to go through raft protocol.
etcd uses [Prometheus][prometheus] for metrics reporting. The metrics can be used for real-time monitoring and debugging. etcd does not persist its metrics; if a member restarts, the metrics will be reset.
The simplest way to see the available metrics is to cURL the metrics endpoint `/metrics`. The format is described [here](http://prometheus.io/docs/instrumenting/exposition_formats/).
Follow the [Prometheus getting started doc][prometheus-getting-started] to spin up a Prometheus server to collect etcd metrics.
The naming of metrics follows the suggested [Prometheus best practices][prometheus-naming]. A metric name has an `etcd` or `etcd_debugging` prefix as its namespace and a subsystem prefix (for example `wal` and `etcdserver`).
## etcd namespace metrics
The metrics under the `etcd` prefix are for monitoring and alerting. They are stable high level metrics. If there is any change of these metrics, it will be included in release notes.
Metrics that are etcd2 related are documented [v2 metrics guide][v2-http-metrics].
### Server
These metrics describe the status of the etcd server. In order to detect outages or problems for troubleshooting, the server metrics of every production etcd cluster should be closely monitored.
All these metrics are prefixed with `etcd_server_`
| has_leader | Whether or not a leader exists. 1 is existence, 0 is not.| Gauge |
| leader_changes_seen_total | The number of leader changes seen. | Counter |
| proposals_committed_total | The total number of consensus proposals committed. | Gauge |
| proposals_applied_total | The total number of consensus proposals applied. | Gauge |
| proposals_pending | The current number of pending proposals. | Gauge |
| proposals_failed_total | The total number of failed proposals seen. | Counter |
`has_leader` indicates whether the member has a leader. If a member does not have a leader, it is
totally unavailable. If all the members in the cluster do not have any leader, the entire cluster
is totally unavailable.
`leader_changes_seen_total` counts the number of leader changes the member has seen since its start. Rapid leadership changes impact the performance of etcd significantly. It also signals that the leader is unstable, perhaps due to network connectivity issues or excessive load hitting the etcd cluster.
`proposals_committed_total` records the total number of consensus proposals committed. This gauge should increase over time if the cluster is healthy. Several healthy members of an etcd cluster may have different total committed proposals at once. This discrepancy may be due to recovering from peers after starting, lagging behind the leader, or being the leader and therefore having the most commits. It is important to monitor this metric across all the members in the cluster; a consistently large lag between a single member and its leader indicates that member is slow or unhealthy.
`proposals_applied_total` records the total number of consensus proposals applied. The etcd server applies every committed proposal asynchronously. The difference between `proposals_committed_total` and `proposals_applied_total` should usually be small (within a few thousands even under high load). If the difference between them continues to rise, it indicates that the etcd server is overloaded. This might happen when applying expensive queries like heavy range queries or large txn operations.
`proposals_pending` indicates how many proposals are queued to commit. Rising pending proposals suggests there is a high client load or the member cannot commit proposals.
`proposals_failed_total` are normally related to two issues: temporary failures related to a leader election or longer downtime caused by a loss of quorum in the cluster.
### Disk
These metrics describe the status of the disk operations.
| wal_fsync_duration_seconds | The latency distributions of fsync called by wal | Histogram |
| backend_commit_duration_seconds | The latency distributions of commit called by backend.| Histogram |
A `wal_fsync` is called when etcd persists its log entries to disk before applying them.
A `backend_commit` is called when etcd commits an incremental snapshot of its most recent changes to disk.
High disk operation latencies (`wal_fsync_duration_seconds` or `backend_commit_duration_seconds`) often indicate disk issues. It may cause high request latency or make the cluster unstable.
### Network
These metrics describe the status of the network.
All these metrics are prefixed with `etcd_network_`
| peer_sent_bytes_total | The total number of bytes sent to the peer with ID `To`. | Counter(To) |
| peer_received_bytes_total | The total number of bytes received from the peer with ID `From`. | Counter(From) |
| peer_round_trip_time_seconds | Round-Trip-Time histogram between peers. | Histogram(To) |
| client_grpc_sent_bytes_total | The total number of bytes sent to grpc clients. | Counter |
| client_grpc_received_bytes_total| The total number of bytes received to grpc clients. | Counter |
`peer_sent_bytes_total` counts the total number of bytes sent to a specific peer. Usually the leader member sends more data than other members since it is responsible for transmitting replicated data.
`peer_received_bytes_total` counts the total number of bytes received from a specific peer. Usually follower members receive data only from the leader member.
### gRPC requests
These metrics describe the requests served by a specific etcd member: total received requests, total failed requests, and processing latency. They are useful for tracking user-generated traffic hitting the etcd cluster.
| requests_total | Total number of received requests | Counter(method) |
| requests_failed_total | Total number of failed requests. | Counter(method,error) |
| unary_requests_duration_seconds | Bucketed handling duration of the requests. | Histogram(method) |
Example Prometheus queries that may be useful from these metrics (across all etcd members):
*`sum(rate(etcd_grpc_requests_failed_total{job="etcd"}[1m]) by (grpc_method) / sum(rate(etcd_grpc_total{job="etcd"})[1m]) by (grpc_method)`
Shows the fraction of events that failed by gRPC method across all members, across a time window of `1m`.
*`sum(rate(etcd_grpc_requests_total{job="etcd",grpc_method="PUT"})[1m]) by (grpc_method)`
Shows the rate of PUT requests across all members, across a time window of `1m`.
*`histogram_quantile(0.9, sum(rate(etcd_grpc_unary_requests_duration_seconds{job="etcd",grpc_method="PUT"}[5m]) ) by (le))`
Show the 0.90-tile latency (in seconds) of PUT request handling across all members, with a window of `5m`.
## etcd_debugging namespace metrics
The metrics under the `etcd_debugging` prefix are for debugging. They are very implementation dependent and volatile. They might be changed or removed without any warning in new etcd releases. Some of the metrics might be moved to the `etcd` prefix when they become more stable.
| process_open_fds | Number of open file descriptors. | Gauge |
| process_max_fds | Maximum number of open file descriptors. | Gauge |
Heavy file descriptor (`process_open_fds`) usage (i.e., near the process's file descriptor limit, `process_max_fds`) indicates a potential file descriptor exhaustion issue. If the file descriptors are exhausted, etcd may panic because it cannot create new WAL files.
etcd has a number of modules that are built on top of the core etcd API.
These modules provide things like dashboards, locks and leader election (removed).
**Warning**: Modules are deprecated from v0.4 until we have a solid base we can apply them back onto.
For now, we are choosing to focus on raft algorithm and core etcd to make sure that it works correctly and fast.
And it is time consuming to maintain these modules in this period, given that etcd's API changes from time to time.
Moreover, the lock module has some unfixed bugs, which may mislead users.
But we also notice that these modules are popular and useful, and plan to add them back with full functionality as soon as possible.
### Dashboard
An HTML dashboard can be found at `http://127.0.0.1:4001/mod/dashboard/`.
This dashboard is compiled into the etcd binary and uses the same API as regular etcd clients.
Use the `-cors='*'` flag to allow your browser to request information from the current master as it changes.
### Lock
The Lock module implements a fair lock that can be used when lots of clients want access to a single resource.
A lock can be associated with a value.
The value is unique so if a lock tries to request a value that is already queued for a lock then it will find it and watch until that value obtains the lock.
You may supply a `timeout` which will cancel the lock request if it is not obtained within `timeout` seconds. If `timeout` is not supplied, it is presumed to be infinite. If `timeout` is `0`, the lock request will fail if it is not immediately acquired.
If you lock the same value on a key from two separate curl sessions they'll both return at the same time.
Here's the API:
**Acquire a lock (with no value) for "customer1"**
```sh
curl -X POST http://127.0.0.1:4001/mod/v2/lock/customer1?ttl=60
```
**Acquire a lock for "customer1" that is associated with the value "bar"**
```sh
curl -X POST http://127.0.0.1:4001/mod/v2/lock/customer1?ttl=60 -d value=bar
```
**Acquire a lock for "customer1" that is associated with the value "bar" only if it is done within 2 seconds**
```sh
curl -X POST http://127.0.0.1:4001/mod/v2/lock/customer1?ttl=60 -d value=bar -d timeout=2
```
**Renew the TTL on the "customer1" lock for index 2**
```sh
curl -X PUT http://127.0.0.1:4001/mod/v2/lock/customer1?ttl=60 -d index=2
```
**Renew the TTL on the "customer1" lock for value "bar"**
```sh
curl -X PUT http://127.0.0.1:4001/mod/v2/lock/customer1?ttl=60 -d value=bar
```
**Retrieve the current value for the "customer1" lock.**
```sh
curl http://127.0.0.1:4001/mod/v2/lock/customer1
```
**Retrieve the current index for the "customer1" lock**
Starting an etcd cluster statically requires that each member knows another in the cluster. In a number of cases, the IPs of the cluster members may be unknown ahead of time. In these cases, the etcd cluster can be bootstrapped with the help of a discovery service.
Once an etcd cluster is up and running, adding or removing members is done via [runtime reconfiguration][runtime-conf]. To better understand the design behind runtime reconfiguration, we suggest reading [the runtime configuration design document][runtime-reconf-design].
This guide will cover the following mechanisms for bootstrapping an etcd cluster:
* [Static](#static)
* [etcd Discovery](#etcd-discovery)
* [DNS Discovery](#dns-discovery)
Each of the bootstrapping mechanisms will be used to create a three machine etcd cluster with the following details:
|Name|Address|Hostname|
|------|---------|------------------|
|infra0|10.0.1.10|infra0.example.com|
|infra1|10.0.1.11|infra1.example.com|
|infra2|10.0.1.12|infra2.example.com|
## Static
As we know the cluster members, their addresses and the size of the cluster before starting, we can use an offline bootstrap configuration by setting the `initial-cluster` flag. Each machine will get either the following environment variables or command line:
Note that the URLs specified in `initial-cluster` are the _advertised peer URLs_, i.e. they should match the value of `initial-advertise-peer-urls` on the respective nodes.
If spinning up multiple clusters (or creating and destroying a single cluster) with same configuration for testing purpose, it is highly recommended that each cluster is given a unique `initial-cluster-token`. By doing this, etcd can generate unique cluster IDs and member IDs for the clusters even if they otherwise have the exact same configuration. This can protect etcd from cross-cluster-interaction, which might corrupt the clusters.
etcd listens on [`listen-client-urls`][conf-listen-client] to accept client traffic. etcd member advertises the URLs specified in [`advertise-client-urls`][conf-adv-client] to other members, proxies, clients. Please make sure the `advertise-client-urls` are reachable from intended clients. A common mistake is setting `advertise-client-urls` to localhost or leave it as default if the remote clients should reach etcd.
The command line parameters starting with `--initial-cluster` will be ignored on subsequent runs of etcd. Feel free to remove the environment variables or command line flags after the initial bootstrap process. If the configuration needs changes later (for example, adding or removing members to/from the cluster), see the [runtime configuration][runtime-conf] guide.
### TLS
etcd supports encrypted communication through the TLS protocol. TLS channels can be used for encrypted internal cluster communication between peers as well as encrypted client traffic. This section provides examples for setting up a cluster with peer and client TLS. Additional information detailing etcd's TLS support can be found in the [security guide][security-guide].
#### Self-signed certificates
A cluster using self-signed certificates both encrypts traffic and authenticates its connections. To start a cluster with self-signed certificates, each cluster member should have a unique key pair (`member.crt`, `member.key`) signed by a shared cluster CA certificate (`ca.crt`) for both peer connections and client connections. Certificates may be generated by following the etcd [TLS setup][tls-setup] example.
On each machine, etcd would be started with these flags:
If the cluster needs encrypted communication but does not require authenticated connections, etcd can be configured to automatically generate its keys. On initialization, each member creates its own set of keys based on its advertised IP addresses and hosts.
On each machine, etcd would be started with these flag:
In the following example, we have not included our new host in the list of enumerated nodes. If this is a new cluster, the node _must_ be added to the list of initial cluster members.
etcd: infra1 not listed in the initial cluster config
exit 1
```
In this example, we are attempting to map a node (infra0) on a different address (127.0.0.1:2380) than its enumerated address in the cluster list (10.0.1.10:2380). If this node is to listen on multiple addresses, all addresses _must_ be reflected in the "initial-cluster" configuration directive.
etcd: error setting up initial cluster: infra0 has different advertised URLs in the cluster and advertised peer URLs list
exit 1
```
If a peer is configured with a different set of configuration arguments and attempts to join this cluster, etcd will report a cluster ID mismatch will exit.
etcd: conflicting cluster ID to the target cluster (c6ab534d07e8fcc4 != bc25ea2a74fb18b0). Exiting.
exit 1
```
## Discovery
In a number of cases, the IPs of the cluster peers may not be known ahead of time. This is common when utilizing cloud providers or when the network uses DHCP. In these cases, rather than specifying a static configuration, use an existing etcd cluster to bootstrap a new one. We call this process "discovery".
There two methods that can be used for discovery:
* etcd discovery service
* DNS SRV records
### etcd discovery
To better understand the design about discovery service protocol, we suggest reading the discovery service protocol [documentation][discovery-proto].
#### Lifetime of a discovery URL
A discovery URL identifies a unique etcd cluster. Instead of reusing a discovery URL, always create discovery URLs for new clusters.
Moreover, discovery URLs should ONLY be used for the initial bootstrapping of a cluster. To change cluster membership after the cluster is already running, see the [runtime reconfiguration][runtime-conf] guide.
#### Custom etcd discovery service
Discovery uses an existing cluster to bootstrap itself. If using a private etcd cluster, can create a URL like so:
```
$ curl -X PUT https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83/_config/size -d value=3
```
By setting the size key to the URL, a discovery URL is created with an expected cluster size of 3.
The URL to use in this case will be `https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83` and the etcd members will use the `https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83` directory for registration as they start.
**Each member must have a different name flag specified. `Hostname` or `machine-id` can be a good choice. Or discovery will fail due to duplicated name.**
Now we start etcd with those relevant flags for each member:
This will cause each member to register itself with the custom etcd discovery service and begin the cluster once all machines have been registered.
#### Public etcd discovery service
If no exiting cluster is available, use the public discovery service hosted at `discovery.etcd.io`. To create a private discovery URL using the "new" endpoint, use the command:
**Each member must have a different name flag specified. `Hostname` or `machine-id` can be a good choice. Or discovery will fail due to duplicated name.**
Now we start etcd with those relevant flags for each member:
etcd cluster members can listen on domain names or IP address, the bootstrap process will resolve DNS A records.
The resolved address in `--initial-advertise-peer-urls`*must match* one of the resolved addresses in the SRV targets. The etcd member reads the resolved address to find out if it belongs to the cluster defined in the SRV records.
When the `--proxy` flag is set, etcd runs in [proxy mode][proxy]. This proxy mode only supports the etcd v2 API; there are no plans to support the v3 API. Instead, for v3 API support, there will be a new proxy with enhanced features following the etcd 3.0 release.
To setup an etcd cluster with proxies of v2 API, please read the the [clustering doc in etcd 2.3 release][clustering_etcd2].
etcd is configurable through command-line flags and environment variables. Options set on the command line take precedence over those from the environment.
The format of environment variable for flag `--my-flag` is `ETCD_MY_FLAG`. It applies to all flags.
The [official etcd ports][iana-ports] are 2379 for client requests and 2380 for peer communication. The etcd ports can be set to accept TLS traffic, non-TLS traffic, or both TLS and non-TLS traffic.
To start etcd automatically using custom settings at startup in Linux, using a [systemd][systemd-intro] unit is highly recommended.
## Member flags
### --name
+ Human-readable name for this member.
+ default: "default"
+ env variable: ETCD_NAME
+ This value is referenced as this node's own entries listed in the `--initial-cluster` flag (e.g., `default=http://localhost:2380`). This needs to match the key used in the flag if using [static bootstrapping][build-cluster]. When using discovery, each member must have a unique name. `Hostname` or `machine-id` can be a good choice.
### --data-dir
+ Path to the data directory.
+ default: "${name}.etcd"
+ env variable: ETCD_DATA_DIR
### --wal-dir
+ Path to the dedicated wal directory. If this flag is set, etcd will write the WAL files to the walDir rather than the dataDir. This allows a dedicated disk to be used, and helps avoid io competition between logging and other IO operations.
+ default: ""
+ env variable: ETCD_WAL_DIR
### --snapshot-count
+ Number of committed transactions to trigger a snapshot to disk.
+ default: "10000"
+ env variable: ETCD_SNAPSHOT_COUNT
### --heartbeat-interval
+ Time (in milliseconds) of a heartbeat interval.
+ default: "100"
+ env variable: ETCD_HEARTBEAT_INTERVAL
### --election-timeout
+ Time (in milliseconds) for an election to timeout. See [Documentation/tuning.md][tuning] for details.
+ default: "1000"
+ env variable: ETCD_ELECTION_TIMEOUT
### --listen-peer-urls
+ List of URLs to listen on for peer traffic. This flag tells the etcd to accept incoming requests from its peers on the specified scheme://IP:port combinations. Scheme can be either http or https.If 0.0.0.0 is specified as the IP, etcd listens to the given port on all interfaces. If an IP address is given as well as a port, etcd will listen on the given port and interface. Multiple URLs may be used to specify a number of addresses and ports to listen on. The etcd will respond to requests from any of the listed addresses and ports.
+ default: "http://localhost:2380"
+ env variable: ETCD_LISTEN_PEER_URLS
+ example: "http://10.0.0.1:2380"
+ invalid example: "http://example.com:2380" (domain name is invalid for binding)
### --listen-client-urls
+ List of URLs to listen on for client traffic. This flag tells the etcd to accept incoming requests from the clients on the specified scheme://IP:port combinations. Scheme can be either http or https. If 0.0.0.0 is specified as the IP, etcd listens to the given port on all interfaces. If an IP address is given as well as a port, etcd will listen on the given port and interface. Multiple URLs may be used to specify a number of addresses and ports to listen on. The etcd will respond to requests from any of the listed addresses and ports.
+ default: "http://localhost:2379"
+ env variable: ETCD_LISTEN_CLIENT_URLS
+ example: "http://10.0.0.1:2379"
+ invalid example: "http://example.com:2379" (domain name is invalid for binding)
### --max-snapshots
+ Maximum number of snapshot files to retain (0 is unlimited)
+ default: 5
+ env variable: ETCD_MAX_SNAPSHOTS
+ The default for users on Windows is unlimited, and manual purging down to 5 (or some preference for safety) is recommended.
### --max-wals
+ Maximum number of wal files to retain (0 is unlimited)
+ default: 5
+ env variable: ETCD_MAX_WALS
+ The default for users on Windows is unlimited, and manual purging down to 5 (or some preference for safety) is recommended.
### --cors
+ Comma-separated white list of origins for CORS (cross-origin resource sharing).
+ default: none
+ env variable: ETCD_CORS
## Clustering flags
`--initial` prefix flags are used in bootstrapping ([static bootstrap][build-cluster], [discovery-service bootstrap][discovery] or [runtime reconfiguration][reconfig]) a new member, and ignored when restarting an existing member.
`--discovery` prefix flags need to be set when using [discovery service][discovery].
### --initial-advertise-peer-urls
+ List of this member's peer URLs to advertise to the rest of the cluster. These addresses are used for communicating etcd data around the cluster. At least one must be routable to all cluster members. These URLs can contain domain names.
+ Initial cluster configuration for bootstrapping.
+ default: "default=http://localhost:2380"
+ env variable: ETCD_INITIAL_CLUSTER
+ The key is the value of the `--name` flag for each node provided. The default uses `default` for the key because this is the default for the `--name` flag.
### --initial-cluster-state
+ Initial cluster state ("new" or "existing"). Set to `new` for all members present during initial static or DNS bootstrapping. If this option is set to `existing`, etcd will attempt to join the existing cluster. If the wrong value is set, etcd will attempt to start but fail safely.
+ default: "new"
+ env variable: ETCD_INITIAL_CLUSTER_STATE
[static bootstrap]: clustering.md#static
### --initial-cluster-token
+ Initial cluster token for the etcd cluster during bootstrap.
+ default: "etcd-cluster"
+ env variable: ETCD_INITIAL_CLUSTER_TOKEN
### --advertise-client-urls
+ List of this member's client URLs to advertise to the rest of the cluster. These URLs can contain domain names.
+ Be careful if advertising URLs such as http://localhost:2379 from a cluster member and are using the proxy feature of etcd. This will cause loops, because the proxy will be forwarding requests to itself until its resources (memory, file descriptors) are eventually depleted.
### --discovery
+ Discovery URL used to bootstrap the cluster.
+ default: none
+ env variable: ETCD_DISCOVERY
### --discovery-srv
+ DNS srv domain used to bootstrap the cluster.
+ default: none
+ env variable: ETCD_DISCOVERY_SRV
### --discovery-fallback
+ Expected behavior ("exit" or "proxy") when discovery services fails. "proxy" supports v2 API only.
+ default: "proxy"
+ env variable: ETCD_DISCOVERY_FALLBACK
### --discovery-proxy
+ HTTP proxy to use for traffic to discovery service.
+ default: none
+ env variable: ETCD_DISCOVERY_PROXY
### --strict-reconfig-check
+ Reject reconfiguration requests that would cause quorum loss.
+ default: false
+ env variable: ETCD_STRICT_RECONFIG_CHECK
### --auto-compaction-retention
+ Auto compaction retention for mvcc key value store in hour. 0 means disable auto compaction.
+ default: 0
+ env variable: ETCD_AUTO_COMPACTION_RETENTION
## Proxy flags
`--proxy` prefix flags configures etcd to run in [proxy mode][proxy]. "proxy" supports v2 API only.
### --proxy
+ Proxy mode setting ("off", "readonly" or "on").
+ default: "off"
+ env variable: ETCD_PROXY
### --proxy-failure-wait
+ Time (in milliseconds) an endpoint will be held in a failed state before being reconsidered for proxied requests.
+ default: 5000
+ env variable: ETCD_PROXY_FAILURE_WAIT
### --proxy-refresh-interval
+ Time (in milliseconds) of the endpoints refresh interval.
+ default: 30000
+ env variable: ETCD_PROXY_REFRESH_INTERVAL
### --proxy-dial-timeout
+ Time (in milliseconds) for a dial to timeout or 0 to disable the timeout
+ default: 1000
+ env variable: ETCD_PROXY_DIAL_TIMEOUT
### --proxy-write-timeout
+ Time (in milliseconds) for a write to timeout or 0 to disable the timeout.
+ default: 5000
+ env variable: ETCD_PROXY_WRITE_TIMEOUT
### --proxy-read-timeout
+ Time (in milliseconds) for a read to timeout or 0 to disable the timeout.
+ Don't change this value if using watches because use long polling requests.
+ default: 0
+ env variable: ETCD_PROXY_READ_TIMEOUT
## Security flags
The security flags help to [build a secure etcd cluster][security].
### --ca-file [DEPRECATED]
+ Path to the client server TLS CA file. `--ca-file ca.crt` could be replaced by `--trusted-ca-file ca.crt --client-cert-auth` and etcd will perform the same.
+ default: none
+ env variable: ETCD_CA_FILE
### --cert-file
+ Path to the client server TLS cert file.
+ default: none
+ env variable: ETCD_CERT_FILE
### --key-file
+ Path to the client server TLS key file.
+ default: none
+ env variable: ETCD_KEY_FILE
### --client-cert-auth
+ Enable client cert authentication.
+ default: false
+ env variable: ETCD_CLIENT_CERT_AUTH
### --trusted-ca-file
+ Path to the client server TLS trusted CA key file.
+ default: none
+ env variable: ETCD_TRUSTED_CA_FILE
### --auto-tls
+ Client TLS using generated certificates
+ default: false
+ env variable: ETCD_AUTO_TLS
### --peer-ca-file [DEPRECATED]
+ Path to the peer server TLS CA file. `--peer-ca-file ca.crt` could be replaced by `--peer-trusted-ca-file ca.crt --peer-client-cert-auth` and etcd will perform the same.
+ default: none
+ env variable: ETCD_PEER_CA_FILE
### --peer-cert-file
+ Path to the peer server TLS cert file.
+ default: none
+ env variable: ETCD_PEER_CERT_FILE
### --peer-key-file
+ Path to the peer server TLS key file.
+ default: none
+ env variable: ETCD_PEER_KEY_FILE
### --peer-client-cert-auth
+ Enable peer client cert authentication.
+ default: false
+ env variable: ETCD_PEER_CLIENT_CERT_AUTH
### --peer-trusted-ca-file
+ Path to the peer server TLS trusted CA file.
+ default: none
+ env variable: ETCD_PEER_TRUSTED_CA_FILE
### --peer-auto-tls
+ Peer TLS using generated certificates
+ default: false
+ env variable: ETCD_PEER_AUTO_TLS
## Logging flags
### --debug
+ Drop the default log level to DEBUG for all subpackages.
+ default: false (INFO for all packages)
+ env variable: ETCD_DEBUG
### --log-package-levels
+ Set individual etcd subpackages to specific log levels. An example being `etcdserver=WARNING,security=DEBUG`
+ default: none (INFO for all packages)
+ env variable: ETCD_LOG_PACKAGE_LEVELS
## Unsafe flags
Please be CAUTIOUS when using unsafe flags because it will break the guarantees given by the consensus protocol.
For example, it may panic if other members in the cluster are still alive.
Follow the instructions when using these flags.
### --force-new-cluster
+ Force to create a new one-member cluster. It commits configuration changes forcing to remove all existing members in the cluster and add itself. It needs to be set to [restore a backup][restore].
+ default: false
+ env variable: ETCD_FORCE_NEW_CLUSTER
## Miscellaneous flags
### --version
+ Print the version and exit.
+ default: false
### --config-file
+ Load server configuration from a file.
+ default: none
## Profiling flags
### --enable-pprof
+ Enable runtime profiling data via HTTP server. Address is at client URL + "/debug/pprof"
The following guide shows how to run etcd with rkt and Docker using the [static bootstrap process](clustering.md#static).
## Docker
In order to expose the etcd API to clients outside of Docker host, use the host IP address of the container. Please see [`docker inspect`](https://docs.docker.com/engine/reference/commandline/inspect) for more detail on how to get the IP address. Alternatively, specify `--net=host` flag to `docker run` command to skip placing the container inside of a separate network stack.
Failures are common in a large deployment of machines. A machine fails when its hardware or software malfunctions. Multiple machines fail together when there are power failures or network issues. Multiple kinds of failures can also happen at once; it is almost impossible to enumerate all possible failure cases.
In this section, we catalog kinds of failures and discuss how etcd is designed to tolerate these failures. Most users, if not all, can map a particular failure into one kind of failure. To prepare for rare or [unrecoverable failures][unrecoverable], always [back up][backup] the etcd cluster.
## Minor followers failure
When fewer than half of the followers fail, the etcd cluster can still accept requests and make progress without any major disruption. For example, two follower failures will not affect a five member etcd cluster’s operation. However, clients will lose connectivity to the failed members. Client libraries should hide these interruptions from users for read requests by automatically reconnecting to other members. Operators should expect the system load on the other members to increase due to the reconnections.
## Leader failure
When a leader fails, the etcd cluster automatically elects a new leader. The election does not happen instantly once the leader fails. It takes about an election timeout to elect a new leader since the failure detection model is timeout based.
During the leader election the cluster cannot process any writes. Write requests sent during the election are queued for processing until a new leader is elected.
Writes already sent to the old leader but not yet committed may be lost. The new leader has the power to rewrite any uncommitted entries from the previous leader. From the user perspective, some write requests might time out after a new leader election. However, no committed writes are ever lost.
The new leader extends timeouts automatically for all leases. This mechanism ensures a lease will not expire before the granted TTL even if it was granted by the old leader.
## Majority failure
When the majority members of the cluster fail, the etcd cluster fails and cannot accept more writes.
The etcd cluster can only recover from a majority failure once the majority of members become available. If a majority of members cannot come back online, then the operator must start [disaster recovery][unrecoverable] to recover the cluster.
Once a majority of members works, the etcd cluster elects a new leader automatically and returns to a healthy state. The new leader extends timeouts automatically for all leases. This mechanism ensures no lease expires due to server side unavailability.
## Network partition
A network partition is similar to a minor followers failure or a leader failure. A network partition divides the etcd cluster into two parts; one with a member majority and the other with a member minority. The majority side becomes the available cluster and the minority side is unavailable; there is no “split-brain” in etcd.
If the leader is on the majority side, then from the majority point of view the failure is a minority follower failure. If the leader is on the minority side, then it is a leader failure. The leader on the minority side steps down and the majority side elects a new leader.
Once the network partition clears, the minority side automatically recognizes the leader from the majority side and recovers its state.
## Failure during bootstrapping
A cluster bootstrap is only successful if all required members successfully start. If any failure happens during bootstrapping, remove the data directories on all members and re-bootstrap the cluster with a new cluster-token or new discovery token.
Of course, it is possible to recover a failed bootstrapped cluster like recovering a running cluster. However, it almost always takes more time and resources to recover that cluster than bootstrapping a new one, since there is no data to recover.
An etcd cluster needs periodic maintenance to remain reliable. Depending on an etcd application's needs, this maintenance can usually be automated and performed without downtime or significantly degraded performance.
All etcd maintenance manages storage resources consumed by the etcd keyspace. Failure to adequately control the keyspace size is guarded by storage space quotas; if an etcd member runs low on space, a quota will trigger cluster-wide alarms which will put the system into a limited-operation maintenance mode. To avoid running out of space for writes to the keyspace, the etcd keyspace history must be compacted. Storage space itself may be reclaimed by defragmenting etcd members. Finally, periodic snapshot backups of etcd member state makes it possible to recover any unintended logical data loss or corruption caused by operational error.
## History compaction
Since etcd keeps an exact history of its keyspace, this history should be periodically compacted to avoid performance degradation and eventual storage space exhaustion. Compacting the keyspace history drops all information about keys superseded prior to a given keyspace revision. The space used by these keys then becomes available for additional writes to the keyspace.
The keyspace can be compacted automatically with `etcd`'s time windowed history retention policy, or manually with `etcdctl`. The `etcdctl` method provides fine-grained control over the compacting process whereas automatic compacting fits applications that only need key history for some length of time.
`etcd` can be set to automatically compact the keyspace with the `--auto-compaction` option with a period of hours:
```sh
# keep one hour of history
$ etcd --auto-compaction-retention=1
```
An `etcdctl` initiated compaction works as follows:
```sh
# compact up to revision 3
$ etcdctl compact 3
```
Revisions prior to the compaction revision become inaccessible:
```sh
$ etcdctl get --rev=2 somekey
Error: rpc error: code=11desc= etcdserver: mvcc: required revision has been compacted
```
## Defragmentation
After compacting the keyspace, the backend database may exhibit internal fragmentation. Any internal fragmentation is space that is free to use by the backend but still consumes storage space. The process of defragmentation releases this storage space back to the file system. Defragmentation is issued on a per-member so that cluster-wide latency spikes may be avoided.
Compacting old revisions internally fragments `etcd` by leaving gaps in backend database. Fragmented space is available for use by `etcd` but unavailable to the host filesystem.
To defragment an etcd member, use the `etcdctl defrag` command:
The space quota in `etcd` ensures the cluster operates in a reliable fashion. Without a space quota, `etcd` may suffer from poor performance if the keyspace grows excessively large, or it may simply run out of storage space, leading to unpredictable cluster behavior. If the keyspace's backend database for any member exceeds the space quota, `etcd` raises a cluster-wide alarm that puts the cluster into a maintenance mode which only accepts key reads and deletes. After freeing enough space in the keyspace, the alarm can be disarmed and the cluster will resume normal operation.
By default, `etcd` sets a conservative space quota suitable for most applications, but it may be configured on the command line, in bytes:
```sh
# set a very small 16MB quota
$ etcd --quota-backend-bytes=16777216
```
The space quota can be triggered with a loop:
```sh
# fill keyspace
$ while[1];do dd if=/dev/urandom bs=1024count=1024| etcdctl put key || break;done
...
Error: rpc error: code=8desc= etcdserver: mvcc: database space exceeded
Snapshotting the `etcd` cluster on a regular basis serves as a durable backup for an etcd keyspace. By taking periodic snapshots of an etcd member's backend database, an `etcd` cluster can be recovered to a point in time with a known good state.
A snapshot is taken with `etcdctl`:
```sh
$ etcdctl snapshot save backup.db
$ etcdctl --write-out=table snapshot status backup.db
etcd provides stable, sustained high performance. Two factors define performance: latency and throughput. Latency is the time taken to complete an operation. Throughput is the total operations completed within some time period. Usually average latency increases as the overall throughput increases when etcd accepts concurrent client requests. In common cloud environments, like a standard `n-4` on Google Compute Engine (GCE) or a comparable machine type on AWS, a three member etcd cluster finishes a request in less than one millisecond under light load, and can complete more than 30,000 requests per second under heavy load.
etcd uses the Raft consensus algorithm to replicate requests among members and reach agreement. Consensus performance, especially commit latency, is limited by two physical constraints: network IO latency and disk IO latency. The minimum time to finish an etcd request is the network Round Trip Time (RTT) between members, plus the time `fdatasync` requires to commit the data to permanant storage. The RTT within a datacenter may be as long as several hundred microseconds. A typical RTT within the United States is around 50ms, and can be as slow as 400ms between continents. The typical fdatasync latency for a spinning disk is about 10ms. For SSDs, the latency is often lower than 1ms. To increase throughput, etcd batches multiple requests together and submits them to Raft. This batching policy lets etcd attain high throughput despite heavy load.
There are other sub-systems which impact the overall performance of etcd. Each serialized etcd request must run through etcd’s boltdb-backed MVCC storage engine, which usually takes tens of microseconds to finish. Periodically etcd incrementally snapshots its recently applied requests, merging them back with the previous on-disk snapshot. This process may lead to a latency spike. Although this is usually not a problem on SSDs, it may double the observed latency on HDD. Likewise, inflight compactions can impact etcd’s performance. Fortunately, the impact is often insignificant since the compaction is staggered so it does not compete for resources with regular requests. The RPC system, gRPC, gives etcd a well-defined, extensible API, but it also introduces additional latency, especially for local reads.
## Benchmarks
Benchmarking etcd performance can be done with the [benchmark](https://github.com/coreos/etcd/tree/master/tools/benchmark) CLI tool included with etcd.
For some baseline performance numbers, we consider a three member etcd cluster with the following hardware configuration:
- etcd v3 master branch (commit SHA d8f325d), Go 1.6.2
With this configuration, etcd can approximately write:
| Number of keys | Key size in bytes | Value size in bytes | Number of connections | Number of clients | Target etcd server | Average write QPS | Average latency per request | Memory |
put --key-size=8 --sequential-keys --total=100000 --val-size=256
```
Linearizable read requests go through a quorum of cluster members for consensus to fetch the most recent data. Serializable read requests are cheaper than linearizable reads since they are served by any single etcd member, instead of a quorum of members, in exchange for possibly serving stale data. etcd can read:
| Number of requests | Key size in bytes | Value size in bytes | Number of connections | Number of clients | Consistency | Average latency per request | Average read QPS |
We encourage running the benchmark test when setting up an etcd cluster for the first time in a new environment to ensure the cluster achieves adequate performance; cluster latency and throughput can be sensitive to minor environment differences.
etcd is designed to withstand machine failures. An etcd cluster automatically recovers from temporary failures (e.g., machine reboots) and tolerates up to *(N-1)/2* permanent failures for a cluster of N members. When a member permanently fails, whether due to hardware failure or disk corruption, it loses access to the cluster. If the cluster permanently loses more than *(N-1)/2* members then it disastrously fails, irrevocably losing quorum. Once quorum is lost, the cluster cannot reach consensus and therefore cannot continue accepting updates.
To recover from disastrous failure, etcd v3 provides snapshot and restore facilities to recreate the cluster without v3 key data loss. To recover v2 keys, refer to the [v2 admin guide][v2_recover].
Recovering a cluster first needs a snapshot of the keyspace from an etcd member. A snapshot may either be taken from a live member with the `etcdctl snapshot save` command or by copying the `member/snap/db` file from an etcd data directory. For example, the following command snapshots the keyspace served by `$ENDPOINT` to the file `snapshot.db`:
```sh
$ etcdctl --endpoints $ENDPOINT snapshot save snapshot.db
```
### Restoring a cluster
To restore a cluster, all that is needed is a single snapshot "db" file. A cluster restore with `etcdctl snapshot restore` creates new etcd data directories; all members should restore using the same snapshot. Restoring overwrites some snapshot metadata (specifically, the member ID and cluster ID); the member loses its former identity. This metadata overwrite prevents the new member from inadvertently joining an existing cluster. Therefore in order to start a cluster from a snapshot, the restore must start a new logical cluster.
Snapshot integrity may be optionally verified at restore time. If the snapshot is taken with `etcdctl snapshot save`, it will have an integrity hash that is checked by `etcdctl snapshot restore`. If the snapshot is copied from the data directory, there is no integrity hash and it will only restore by using `--skip-hash-check`.
A restore initializes a new member of a new cluster, with a fresh cluster configuration using `etcd`'s cluster configuration flags, but preserves the contents of the etcd keyspace. Continuing from the previous example, the following creates new etcd data directories (`m1.etcd`, `m2.etcd`, `m3.etcd`) for a three member cluster:
etcd comes with support for incremental runtime reconfiguration, which allows users to update the membership of the cluster at run time.
Reconfiguration requests can only be processed when the majority of the cluster members are functioning. It is **highly recommended** to always have a cluster size greater than two in production. It is unsafe to remove a member from a two member cluster. The majority of a two member cluster is also two. If there is a failure during the removal process, the cluster might not able to make progress and need to [restart from majority failure][majority failure].
To better understand the design behind runtime reconfiguration, we suggest reading [the runtime reconfiguration document][runtime-reconf].
## Reconfiguration use cases
Let's walk through some common reasons for reconfiguring a cluster. Most of these just involve combinations of adding or removing a member, which are explained below under [Cluster Reconfiguration Operations][cluster-reconf].
### Cycle or upgrade multiple machines
If multiple cluster members need to move due to planned maintenance (hardware upgrades, network downtime, etc.), it is recommended to modify members one at a time.
It is safe to remove the leader, however there is a brief period of downtime while the election process takes place. If the cluster holds more than 50MB, it is recommended to [migrate the member's data directory][member migration].
### Change the cluster size
Increasing the cluster size can enhance [failure tolerance][fault tolerance table] and provide better read performance. Since clients can read from any member, increasing the number of members increases the overall read throughput.
Decreasing the cluster size can improve the write performance of a cluster, with a trade-off of decreased resilience. Writes into the cluster are replicated to a majority of members of the cluster before considered committed. Decreasing the cluster size lowers the majority, and each write is committed more quickly.
### Replace a failed machine
If a machine fails due to hardware failure, data directory corruption, or some other fatal situation, it should be replaced as soon as possible. Machines that have failed but haven't been removed adversely affect the quorum and reduce the tolerance for an additional failure.
To replace the machine, follow the instructions for [removing the member][remove member] from the cluster, and then [add a new member][add member] in its place. If the cluster holds more than 50MB, it is recommended to [migrate the failed member's data directory][member migration] if it is still accessible.
### Restart cluster from majority failure
If the majority of the cluster is lost or all of the nodes have changed IP addresses, then manual action is necessary to recover safely.
The basic steps in the recovery process include [creating a new cluster using the old data][disaster recovery], forcing a single member to act as the leader, and finally using runtime configuration to [add new members][add member] to this new cluster one at a time.
## Cluster reconfiguration operations
Now that we have the use cases in mind, let us lay out the operations involved in each.
Before making any change, the simple majority (quorum) of etcd members must be available.
This is essentially the same requirement as for any other write to etcd.
All changes to the cluster are done one at a time:
* To update a single member peerURLs, make an update operation
* To replace a single member, make an add then a remove operation
* To increase from 3 to 5 members, make two add operations
* To decrease from 5 to 3, make two remove operations
All of these examples will use the `etcdctl` command line tool that ships with etcd.
To change membership without `etcdctl`, use the [v2 HTTP members API][member-api] or the [v3 gRPC members API][member-api-grpc].
### Update a member
#### Update advertise client URLs
To update the advertise client URLs of a member, simply restart
that member with updated client urls flag (`--advertise-client-urls`) or environment variable
(`ETCD_ADVERTISE_CLIENT_URLS`). The restarted member will self publish the updated URLs.
A wrongly updated client URL will not affect the health of the etcd cluster.
#### Update advertise peer URLs
To update the advertise peer URLs of a member, first update
it explicitly via member command and then restart the member. The additional action is required
since updating peer URLs changes the cluster wide configuration and can affect the health of the etcd cluster.
To update the peer URLs, first, we need to find the target member's ID. To list all members with `etcdctl`:
In this example let's `update` a8266ecf031671f3 member ID and change its peerURLs value to http://10.0.1.10:2380
```sh
$ etcdctl member update a8266ecf031671f3 http://10.0.1.10:2380
Updated member with ID a8266ecf031671f3 in cluster
```
### Remove a member
Let us say the member ID we want to remove is a8266ecf031671f3.
We then use the `remove` command to perform the removal:
```sh
$ etcdctl member remove a8266ecf031671f3
Removed member a8266ecf031671f3 from cluster
```
The target member will stop itself at this point and print out the removal in the log:
```
etcd: this member has been permanently removed from the cluster. Exiting.
```
It is safe to remove the leader, however the cluster will be inactive while a new leader is elected. This duration is normally the period of election timeout plus the voting process.
### Add a new member
Adding a member is a two step process:
* Add the new member to the cluster via the [HTTP members API][member-api], the [gRPC members API][member-api-grpc], or the `etcdctl member add` command.
* Start the new member with the new cluster configuration, including a list of the updated members (existing members + the new member).
Using `etcdctl` let's add the new member to the cluster by specifying its [name][conf-name] and [advertised peer URLs][conf-adv-peer]:
The new member will run as a part of the cluster and immediately begin catching up with the rest of the cluster.
If adding multiple members the best practice is to configure a single member at a time and verify it starts correctly before adding more new members.
If adding a new member to a 1-node cluster, the cluster cannot make progress before the new member starts because it needs two members as majority to agree on the consensus. This behavior only happens between the time `etcdctl member add` informs the cluster about the new member and the new member successfully establishing a connection to the existing one.
#### Error cases when adding members
In the following case we have not included our new host in the list of enumerated nodes.
If this is a new cluster, the node must be added to the list of initial cluster members.
As described in the above, the best practice of adding new members is to configure a single member at a time and verify it starts correctly before adding more new members. This step by step approach is very important because if newly added members is not configured correctly (for example the peer URLs are incorrect), the cluster can lose quorum. The quorum loss happens since the newly added member are counted in the quorum even if that member is not reachable from other existing members. Also quorum loss might happen if there is a connectivity issue or there are operational issues.
For avoiding this problem, etcd provides an option `-strict-reconfig-check`. If this option is passed to etcd, etcd rejects reconfiguration requests if the number of started members will be less than a quorum of the reconfigured cluster.
It is recommended to enable this option. However, it is disabled by default because of keeping compatibility.
Runtime reconfiguration is one of the hardest and most error prone features in a distributed system, especially in a consensus based system like etcd.
Read on to learn about the design of etcd's runtime reconfiguration commands and how we tackled these problems.
## Two phase config changes keep the cluster safe
In etcd, every runtime reconfiguration has to go through [two phases][add-member] for safety reasons. For example, to add a member, first inform cluster of new configuration and then start the new member.
Phase 1 - Inform cluster of new configuration
To add a member into etcd cluster, make an API call to request a new member to be added to the cluster. This is only way to add a new member into an existing cluster. The API call returns when the cluster agrees on the configuration change.
Phase 2 - Start new member
To join the etcd member into the existing cluster, specify the correct `initial-cluster` and set `initial-cluster-state` to `existing`. When the member starts, it will contact the existing cluster first and verify the current cluster configuration matches the expected one specified in `initial-cluster`. When the new member successfully starts, the cluster has reached the expected configuration.
By splitting the process into two discrete phases users are forced to be explicit regarding cluster membership changes. This actually gives users more flexibility and makes things easier to reason about. For example, if there is an attempt to add a new member with the same ID as an existing member in an etcd cluster, the action will fail immediately during phase one without impacting the running cluster. Similar protection is provided to prevent adding new members by mistake. If a new etcd member attempts to join the cluster before the cluster has accepted the configuration change,, it will not be accepted by the cluster.
Without the explicit workflow around cluster membership etcd would be vulnerable to unexpected cluster membership changes. For example, if etcd is running under an init system such as systemd, etcd would be restarted after being removed via the membership API, and attempt to rejoin the cluster on startup. This cycle would continue every time a member is removed via the API and systemd is set to restart etcd after failing, which is unexpected.
We expect runtime reconfiguration to be an infrequent operation. We decided to keep it explicit and user-driven to ensure configuration safety and keep the cluster always running smoothly under explicit control.
## Permanent loss of quorum requires new cluster
If a cluster permanently loses a majority of its members, a new cluster will need to be started from an old data directory to recover the previous state.
It is entirely possible to force removing the failed members from the existing cluster to recover. However, we decided not to support this method since it bypasses the normal consensus committing phase, which is unsafe. If the member to remove is not actually dead or force removed through different members in the same cluster, etcd will end up with a diverged cluster with same clusterID. This is very dangerous and hard to debug/fix afterwards.
With a correct deployment, the possibility of permanent majority lose is very low. But it is a severe enough problem that worth special care. We strongly suggest reading the [disaster recovery documentation][disaster-recovery] and prepare for permanent majority lose before putting etcd into production.
## Do not use public discovery service for runtime reconfiguration
The public discovery service should only be used for bootstrapping a cluster. To join member into an existing cluster, use runtime reconfiguration API.
Discovery service is designed for bootstrapping an etcd cluster in the cloud environment, when the IP addresses of all the members are not known beforehand. After successfully bootstrapping a cluster, the IP addresses of all the members are known. Technically, the discovery service should no longer be needed.
It seems that using public discovery service is a convenient way to do runtime reconfiguration, after all discovery service already has all the cluster configuration information. However relying on public discovery service brings troubles:
1. it introduces external dependencies for the entire life-cycle of the cluster, not just bootstrap time. If there is a network issue between the cluster and public discovery service, the cluster will suffer from it.
2. public discovery service must reflect correct runtime configuration of the cluster during it life-cycle. It has to provide security mechanism to avoid bad actions, and it is hard.
3. public discovery service has to keep tens of thousands of cluster configurations. Our public discovery service backend is not ready for that workload.
To have a discovery service that supports runtime reconfiguration, the best choice is to build a private one.
etcd supports automatic TLS as well as authentication through client certificates for both clients to server as well as peer (server to server / cluster) communication.
To get up and running, first have a CA certificate and a signed key pair for one member. It is recommended to create and sign a new key pair for every member in a cluster.
For convenience, the [cfssl] tool provides an easy interface to certificate generation, and we provide an example using the tool [here][tls-setup]. Alternatively, try this [guide to generating self-signed key pairs][tls-guide].
## Basic setup
etcd takes several certificate related configuration options, either through command-line flags or environment variables:
**Client-to-server communication:**
`--cert-file=<path>`: Certificate used for SSL/TLS connections **to** etcd. When this option is set, advertise-client-urls can use the HTTPS schema.
`--key-file=<path>`: Key for the certificate. Must be unencrypted.
`--client-cert-auth`: When this is set etcd will check all incoming HTTPS requests for a client certificate signed by the trusted CA, requests that don't supply a valid client certificate will fail.
The peer options work the same way as the client-to-server options:
`--peer-cert-file=<path>`: Certificate used for SSL/TLS connections between peers. This will be used both for listening on the peer address as well as sending requests to other peers.
`--peer-key-file=<path>`: Key for the certificate. Must be unencrypted.
`--peer-client-cert-auth`: When set, etcd will check all incoming peer requests from the cluster for valid client certificates signed by the supplied CA.
`--peer-auto-tls`: Use automatically generated self-signed certificates for TLS connections between peers.
If either a client-to-server or peer certificate is supplied the key must also be set. All of these configuration options are also available through the environment variables, `ETCD_CA_FILE`, `ETCD_PEER_CA_FILE` and so on.
## Example 1: Client-to-server transport security with HTTPS
For this, have a CA certificate (`ca.crt`) and signed key pair (`server.crt`, `server.key`) ready.
Let us configure etcd to provide simple HTTPS transport security step by step:
The command should show that the handshake succeed. Since we use self-signed certificates with our own certificate authority, the CA must be passed to curl using the `--cacert` option. Another possibility would be to add the CA certificate to the system's trusted certificates directory (usually in `/etc/pki/tls/certs` or `/etc/ssl/certs`).
**OSX 10.9+ Users**: curl 7.30.0 on OSX 10.9+ doesn't understand certificates passed in on the command line.
Instead, import the dummy ca.crt directly into the keychain or add the `-k` flag to curl to ignore errors.
To test without the `-k` flag, run `open ./fixtures/ca/ca.crt` and follow the prompts.
Please remove this certificate after testing!
If there is a workaround, let us know.
## Example 2: Client-to-server authentication with HTTPS client certificates
For now we've given the etcd client the ability to verify the server identity and provide transport security. We can however also use client certificates to prevent unauthorized access to etcd.
The clients will provide their certificates to the server and the server will check whether the cert is signed by the supplied CA and decide whether to serve the request.
The same files mentioned in the first example are needed for this, as well as a key pair for the client (`client.crt`, `client.key`) signed by the same certificate authority.
## Example 3: Transport security & client certificates in a cluster
etcd supports the same model as above for **peer communication**, that means the communication between etcd members in a cluster.
Assuming we have our `ca.crt` and two members with their own keypairs (`member1.crt`&`member1.key`, `member2.crt`&`member2.key`) signed by this CA, we launch etcd as follows:
```sh
DISCOVERY_URL=... # from https://discovery.etcd.io/new
The etcd members will form a cluster and all communication between members in the cluster will be encrypted and authenticated using the client certificates. The output of etcd will show that the addresses it connects to use HTTPS.
## Example 4: Automatic self-signed transport security
For cases where communication encryption, but not authentication, is needed, etcd supports encrypting its messages with automatically generated self-signed certificates. This simplifies deployment because there is no need for managing certificates and keys outside of etcd.
Configure etcd to use self-signed certificates for client and peer connections with the flags `--auto-tls` and `--peer-auto-tls`:
```sh
DISCOVERY_URL=... # from https://discovery.etcd.io/new
etcd proxy terminates the TLS from its client if the connection is secure, and uses proxy's own key/cert specified in `--peer-key-file` and `--peer-cert-file` to communicate with etcd members.
The proxy communicates with etcd members through both the `--advertise-client-urls` and `--advertise-peer-urls` of a given member. It forwards client requests to etcd members’ advertised client urls, and it syncs the initial cluster configuration through etcd members’ advertised peer urls.
When client authentication is enabled for an etcd member, the administrator must ensure that the peer certificate specified in the proxy's `--peer-cert-file` option is valid for that authentication. The proxy's peer certificate must also be valid for peer authentication if peer authentication is enabled.
## Frequently asked questions
### I'm seeing a SSLv3 alert handshake failure when using TLS client authentication?
The `crypto/tls` package of `golang` checks the key usage of the certificate public key before using it.
To use the certificate public key to do client auth, we need to add `clientAuth` to `Extended Key Usage` when creating the certificate public key.
Here is how to do it:
Add the following section to openssl.cnf:
```
[ ssl_client ]
...
extendedKeyUsage = clientAuth
...
```
When creating the cert be sure to reference it in the `-extensions` flag:
### With peer certificate authentication I receive "certificate is valid for 127.0.0.1, not $MY_IP"
Make sure to sign the certificates with a Subject Name the member's public IP address. The `etcd-ca` tool for example provides an `--ip=` option for its `new-cert` command.
The certificate needs to be signed for the member's FQDN in its Subject Name, use Subject Alternative Names (short IP SANs) to add the IP address. The `etcd-ca` tool provides `--domain=` option for its `new-cert` command, and openssl can make [it][alt-name] too.
# Migrate applications from using API v2 to API v3
The data store v2 is still accessible from the API v2 after upgrading to etcd3. Thus, it will work as before and require no application changes. With etcd 3, applications use the new grpc API v3 to access the mvcc store, which provides more features and improved performance. The mvcc store and the old store v2 are separate and isolated; writes to the store v2 will not affect the mvcc store and, similarly, writes to the mvcc store will not affect the store v2.
Migrating an application from the API v2 to the API v3 involves two steps: 1) migrate the client library and, 2) migrate the data. If the application can rebuild the data, then migrating the data is unnecessary.
## Migrate client library
API v3 is different from API v2, thus application developers need to use a new client library to send requests to etcd API v3. The documentation of the client v3 is available at https://godoc.org/github.com/coreos/etcd/clientv3.
There are some notable differences between API v2 and API v3:
- Transaction: In v3, etcd provides multi-key conditional transactions. Applications should use transactions in place of `Compare-And-Swap` operations.
- Flat key space: There are no directories in API v3, only keys. For example, "/a/b/c/" is a key. Range queries support getting all keys matching a given prefix.
- Compacted responses: Operations like `Delete` no longer return previous values. To get the deleted value, a transaction can be used to atomically get the key and then delete its value.
- Leases: A replacement for v2 TTLs; the TTL is bound to a lease and keys attach to the lease. When the TTL expires, the lease is revoked and all attached keys are removed.
## Migrate data
Application data can be migrated either offline or online. Offline migration is much simpler than online migration and is recommended.
### Offline migration
Offline migration is very simple but requires etcd downtime. If an etcd downtime window spanning from seconds to minutes is acceptable, offline migration is a good choice and is easy to automate.
First, all members in the etcd cluster must converge to the same state. This can be achieved by stopping all applications that write keys to etcd. Alternatively, if the applications must remain running, configure etcd to listen on a different client URL and restart all etcd members. To check if the states converged, within a few seconds, use the `ETCDCTL_API=3 etcdctl endpoint status` command to confirm that the `raft index` of all members match (or differ by at most 1 due to an internal sync raft command).
Second, migrate the v2 keys into v3 with the [migrate][migrate_command] (`ETCDCTL_API=3 etcdctl migrate`) command. The migrate command writes keys in the v2 store to a user-provided transformer program and reads back transformed keys. It then writes transformed keys into the mvcc store. This usually takes at most tens of seconds.
Restart the etcd members and everything should just work.
### Online migration
If the application cannot tolerate any downtime, then it must migrate online. The implementation of online migration will vary from application to application but the overall idea is the same.
First, write application code using the v3 API. The application must support two modes: a migration mode and a normal mode. The application starts in migration mode. When running in migration mode, the application reads keys using the v3 API first, and, if it cannot find the key, it retries with the API v2. In normal mode, the application only reads keys using the v3 API. The application writes keys over the API v3 in both modes. To acknowledge a switch from migration mode to normal mode, the application watches on a switch mode key. When switch key’s value turns to `true`, the application switches over from migration mode to normal mode.
Second, start a background job to migrate data from the store v2 to the mvcc store by reading keys from the API v2 and writing keys to the API v3.
After finishing data migration, the background job writes `true` into the switch mode key to notify the application that it may switch modes.
Online migration can be difficult when the application logic depends on store v2 indexes. Applications will need additional logic to convert mvcc store revisions to store v2 indexes.
etcd's Raft consensus algorithm is most efficient in small clusters between 3 and 9 peers. For clusters larger than 9, etcd will select a subset of instances to participate in the algorithm in order to keep it efficient. The end of this document briefly explores how etcd works internally and why these choices have been made.
## Cluster Management
You can manage the active cluster size through the [cluster config API](https://github.com/coreos/etcd/blob/master/Documentation/api.md#cluster-config). `activeSize` represents the etcd peers allowed to actively participate in the consensus algorithm.
If the total number of etcd instances exceeds this number, additional peers are started as [standbys](https://github.com/coreos/etcd/blob/master/Documentation/design/standbys.md), which can be promoted to active participation if one of the existing active instances has failed or been removed.
## Internals of etcd
### Writing to etcd
Writes to an etcd peer are always redirected to the leader of the cluster and distributed to all of the peers immediately. A write is only considered successful when a majority of the peers acknowledge the write.
For example, in a cluster with 5 peers, a write operation is only as fast as the 3rd fastest machine. This is the main reason for keeping the number of active peers below 9. In practice, you only need to worry about write performance in high latency environments such as a cluster spanning multiple data centers.
### Leader Election
The leader election process is similar to writing a key — a majority of the active peers must acknowledge the new leader before cluster operations can continue. The longer each peer takes to elect a new leader means you have to wait longer before you can write to the cluster again. In low latency environments this process takes milliseconds.
### Odd Active Cluster Size
The other important cluster optimization is to always have an odd active cluster size (i.e. `activeSize`). Adding an odd node to the number of peers doesn't change the size of the majority and therefore doesn't increase the total latency of the majority as described above. But, you gain a higher tolerance for peer failure by adding the extra machine. You can see this in practice when comparing two even and odd sized clusters:
| Active Peers | Majority | Failure Tolerance |
|--------------|------------|-------------------|
| 1 peers | 1 peers | None |
| 3 peers | 2 peers | 1 peer |
| 4 peers | 3 peers | 1 peer |
| 5 peers | 3 peers | **2 peers** |
| 6 peers | 4 peers | 2 peers |
| 7 peers | 4 peers | **3 peers** |
| 8 peers | 5 peers | 3 peers |
| 9 peers | 5 peers | **4 peers** |
As you can see, adding another peer to bring the number of active peers up to an odd size is always worth it. During a network partition, an odd number of active peers also guarantees that there will almost always be a majority of the cluster that can continue to operate and be the source of truth when the partition ends.
This document tracks people and use cases for etcd in production. By creating a list of production use cases we hope to build a community of advisors that we can reach out to with experience using various etcd applications, operation environments, and cluster sizes. The etcd development team may reach out periodically to check-in on how etcd is working in the field and update this list.
discovery.etcd.io is the longest continuously running etcd backed service that we know about. It is the basis of automatic cluster bootstrap and was launched in Feb. 2014: https://coreos.com/blog/etcd-0.3.0-released/.
## OpenTable
- *Application*: OpenTable internal service discovery and cluster configuration management
- *Launched*: May 2014
- *Cluster Size*: 3 members each in 6 independent clusters; approximately 50 nodes reading / writing
- *Backups*: Periodic sync to Ceph RadosGW and DigitalOcean VM
CyCore Systems provides architecture and engineering for computing systems. This cluster provides microservices, virtual machines, databases, storage clusters to a number of clients. It is built on CoreOS machines, with each machine in the cluster running etcd as a peer or proxy.
## Radius Intelligence
- *Application*: multiple internal tools, Kubernetes clusters, bootstrappable system configs
- *Launched*: June 2015
- *Cluster Size*: 2 clusters of 5 and 3 members; approximately a dozen nodes read/write
- *Backups*: None, all data can be recreated if necessary.
Radius Intelligence uses Kubernetes running CoreOS to containerize and scale internal toolsets. Examples include running [JetBrains TeamCity][teamcity] and internal AWS security and cost reporting tools. etcd clusters back these clusters as well as provide some basic environment bootstrapping configuration keys.
## Vonage
- *Application*: system configuration for microservices, scheduling, locks (future - service discovery)
- *Launched*: August 2015
- *Cluster Size*: 2 clusters of 5 members in 2 DCs, n local proxies 1-to-1 with microservice, (ssl and SRV look up)
- *Order of Data Size*: kilobytes
- *Operator*: Vonage [devAdmin][raoofm]
- *Environment*: VMWare, AWS
- *Backups*: Daily snapshots on VMs. Backups done for upgrades.
If any part of the etcd project has bugs or documentation mistakes, please let us know by [opening an issue][issue]. We treat bugs and mistakes very seriously and believe no issue is too small. Before creating a bug report, please check that an issue reporting the same problem does not already exist.
To make the bug report accurate and easy to understand, please try to create bug reports that are:
- Specific. Include as much details as possible: which version, what environment, what configuration, etc. If the bug is related to running the etcd server, please attach the etcd log (the starting log with etcd configuration is especially important).
- Reproducible. Include the steps to reproduce the problem. We understand some issues might be hard to reproduce, please includes the steps that might lead to the problem. If possible, please attach the affected etcd data dir and stack strace to the bug report.
- Isolated. Please try to isolate and reproduce the bug with minimum dependencies. It would significantly slow down the speed to fix a bug if too many dependencies are involved in a bug report. Debugging external systems that rely on etcd is out of scope, but we are happy to provide guidance in the right direction or help with using etcd itself.
- Unique. Do not duplicate existing bug report.
- Scoped. One bug per report. Do not follow up with another bug inside one report.
It may be worthwhile to read [Elika Etemad’s article on filing good bug reports][filing-good-bugs] before creating a bug report.
We might ask for further information to locate a bug. A duplicated bug report will be closed.
## Frequently asked questions
### How to get a stack trace
``` bash
$ kill -QUIT $PID
```
### How to get etcd version
``` bash
$ etcd --version
```
### How to get etcd configuration and log when it runs as systemd service ‘etcd2.service’
``` bash
$ sudo systemctl cat etcd2
$ sudo journalctl -u etcd2
```
Due to an upstream systemd bug, journald may miss the last few log lines when its processes exit. If journalctl says etcd stopped without fatal or panic message, try `sudo journalctl -f -t etcd2` to get full log.
The etcd v3 API is designed to give users a more efficient and cleaner abstraction compared to etcd v2. There are a number of semantic and protocol changes in this new API. For an overview [see Xiang Li's video](https://youtu.be/J5AioGtEPeQ?t=211).
To prove out the design of the v3 API the team has also built [a number of example recipes](https://github.com/coreos/etcd/tree/master/contrib/recipes), there is a [video discussing these recipes too](https://www.youtube.com/watch?v=fj-2RY-3yVU&feature=youtu.be&t=590).
# Design
1. Flatten binary key-value space
2. Keep the event history until compaction
- access to old version of keys
- user controlled history compaction
3. Support range query
- Pagination support with limit argument
- Support consistency guarantee across multiple range queries
4. Replace TTL key with Lease
- more efficient/ low cost keep alive
- a logical group of TTL keys
5. Replace CAS/CAD with multi-object Txn
- MUCH MORE powerful and flexible
6. Support efficient watching with multiple ranges
7. RPC API supports the completed set of APIs.
- more efficient than JSON/HTTP
- additional txn/lease support
8. HTTP API supports a subset of APIs.
- easy for people to try out etcd
- easy for people to write simple etcd application
## Notes
### Request Size Limitation
The max request size is around 1MB. Since etcd replicates requests in a streaming fashion, a very large
request might block other requests for a long time. The use case for etcd is to store small configuration
values, so we prevent user from submitting large requests. This also applies to Txn requests. We might loosen
the size in the future a little bit or make it configurable.
## Protobuf Defined API
[api protobuf][api-protobuf]
[kv protobuf][kv-protobuf]
## Examples
### Put a key (foo=bar)
```
// A put is always successful
Put( PutRequest { key = foo, value = bar } )
PutResponse {
cluster_id = 0x1000,
member_id = 0x1,
revision = 1,
raft_term = 0x1,
}
```
### Get a key (assume we have foo=bar)
```
Get ( RangeRequest { key = foo } )
RangeResponse {
cluster_id = 0x1000,
member_id = 0x1,
revision = 1,
raft_term = 0x1,
kvs = {
{
key = foo,
value = bar,
create_revision = 1,
mod_revision = 1,
version = 1;
},
},
}
```
### Range over a key space (assume we have foo0=bar0… foo100=bar100)
Etcd supports SSL/TLS as well as authentication through client certificates, both for clients to server as well as peer (server to server / cluster) communication.
To get up and running you first need to have a CA certificate and a signed key pair for your node. It is recommended to create and sign a new key pair for every node in a cluster.
For convenience the [etcd-ca](https://github.com/coreos/etcd-ca) tool provides an easy interface to certificate generation, alternatively this site provides a good reference on how to generate self-signed key pairs:
http://www.g-loaded.eu/2005/11/10/be-your-own-ca/
## Basic setup
Etcd takes several certificate related configuration options, either through command-line flags or environment variables:
**Client-to-server communication:**
`--cert-file=<path>`: Certificate used for SSL/TLS connections **to** etcd. When this option is set, you can reach etcd through HTTPS - for example at `https://127.0.0.1:4001`
`--key-file=<path>`: Key for the certificate. Must be unencrypted.
`--ca-file=<path>`: When this is set etcd will check all incoming HTTPS requests for a client certificate signed by the supplied CA, requests that don't supply a valid client certificate will fail.
The peer options work the same way as the client-to-server options:
`--peer-cert-file=<path>`: Certificate used for SSL/TLS connections between peers. This will be used both for listening on the peer address as well as sending requests to other peers.
`--peer-key-file=<path>`: Key for the certificate. Must be unencrypted.
`--peer-ca-file=<path>`: When set, etcd will check all incoming peer requests from the cluster for valid client certificates signed by the supplied CA.
If either a client-to-server or peer certificate is supplied the key must also be set. All of these configuration options are also available through the environment variables, `ETCD_CA_FILE`, `ETCD_PEER_CA_FILE` and so on.
## Example 1: Client-to-server transport security with HTTPS
For this you need your CA certificate (`ca.crt`) and signed key pair (`server.crt`, `server.key`) ready. If you just want to test the functionality, there are example certificates provided in the [etcd git repository](https://github.com/coreos/etcd/tree/master/fixtures/ca) (namely `server.crt` and `server.key.insecure`).
Assuming you have these files ready, let's configure etcd to use them to provide simple HTTPS transport security.
You should be able to see the handshake succeed. Because we use self-signed certificates with our own certificate authorities you need to provide the CA to curl using the `--cacert` option. Another possibility would be to add your CA certificate to the trusted certificates on your system (usually in `/etc/ssl/certs`).
**OSX 10.9+ Users**: curl 7.30.0 on OSX 10.9+ doesn't understand certificates passed in on the command line.
Instead you must import the dummy ca.crt directly into the keychain or add the `-k` flag to curl to ignore errors.
If you want to test without the `-k` flag run `open ./fixtures/ca/ca.crt` and follow the prompts.
Please remove this certificate after you are done testing!
If you know of a workaround let us know.
## Example 2: Client-to-server authentication with HTTPS client certificates
For now we've given the etcd client the ability to verify the server identity and provide transport security. We can however also use client certificates to prevent unauthorized access to etcd.
The clients will provide their certificates to the server and the server will check whether the cert is signed by the supplied CA and decide whether to serve the request.
You need the same files mentioned in the first example for this, as well as a key pair for the client (`client.crt`, `client.key`) signed by the same certificate authority.
## Example 3: Transport security & client certificates in a cluster
Etcd supports the same model as above for **peer communication**, that means the communication between etcd nodes in a cluster.
Assuming we have our `ca.crt` and two nodes with their own keypairs (`node1.crt`&`node1.key`, `node2.crt`&`node2.key`) signed by this CA, we launch etcd as follows:
```sh
DISCOVERY_URL=... # from https://discovery.etcd.io/new
The etcd nodes will form a cluster and all communication between nodes in the cluster will be encrypted and authenticated using the client certificates. You will see in the output of etcd that the addresses it connects to use HTTPS.
## Frequently Asked Questions
### I'm seeing a SSLv3 alert handshake failure when using SSL client authentication?
The `crypto/tls` package of `golang` checks the key usage of the certificate public key before using it.
To use the certificate public key to do client auth, we need to add `clientAuth` to `Extended Key Usage` when creating the certificate public key.
Here is how to do it:
Add the following section to your openssl.cnf:
```
[ ssl_client ]
...
extendedKeyUsage = clientAuth
...
```
When creating the cert be sure to reference it in the `-extensions` flag:
### With peer certificate authentication I receive "certificate is valid for 127.0.0.1, not $MY_IP"
Make sure that you sign your certificates with a Subject Name your node's public IP address. The `etcd-ca` tool for example provides an `--ip=` option for its `new-cert` command.
If you need your certificate to be signed for your node's FQDN in its Subject Name then you could use Subject Alternative Names (short IP SNAs) to add your IP address. This is not [currently supported](https://github.com/coreos/etcd-ca/issues/29) by `etcd-ca` but can be done [with openssl](http://wiki.cacert.org/FAQ/subjectAltName).
The default settings in etcd should work well for installations on a local network where the average network latency is low.
However, when using etcd across multiple data centers or over networks with high latency you may need to tweak the heartbeat interval and election timeout settings.
The default settings in etcd should work well for installations on a local network where the average network latency is low. However, when using etcd across multiple data centers or over networks with high latency, the heartbeat interval and election timeout settings may need tuning.
The network isn't the only source of latency. Each request and response may be impacted by slow disks on both the leader and follower. Each of these timeouts represents the total time from request to successful response from the other machine.
### Time Parameters
## Time parameters
The underlying distributed consensus protocol relies on two separate time parameters to ensure that nodes can handoff leadership if one stalls or goes offline.
The first parameter is called the *Heartbeat Interval*.
This is the frequency with which the leader will notify followers that it is still the leader.
etcd batches commands together for higher throughput so this heartbeat interval is also a delay for how long it takes for commands to be committed.
By default, etcd uses a `50ms` heartbeat interval.
For best practices, the parameter should be set around round-trip time between members.
By default, etcd uses a `100ms` heartbeat interval.
The second parameter is the *Election Timeout*.
This timeout is how long a follower node will go without hearing a heartbeat before attempting to become leader itself.
By default, etcd uses a `200ms` election timeout.
By default, etcd uses a `1000ms` election timeout.
Adjusting these values is a trade off.
Lowering the heartbeat interval will cause individual commands to be committed faster but it will lower the overall throughput of etcd.
If your etcd instances have low utilization then lowering the heartbeat interval can improve your command response time.
The value of heartbeat interval is recommended to be around the maximum of average round-trip time (RTT) between members, normally around 0.5-1.5x the round-trip time.
If heartbeat interval is too low, etcd will send unnecessary messages that increase the usage of CPU and network resources.
On the other side, a too high heartbeat interval leads to high election timeout. Higher election timeout takes longer time to detect a leader failure.
The easiest way to measure round-trip time (RTT) is to use [PING utility][ping].
The election timeout should be set based on the heartbeat interval and your network ping time between nodes.
Election timeouts should be at least 10 times your ping time so it can account for variance in your network.
For example, if the ping time between your nodes is 10ms then you should have at least a 100ms election timeout.
The election timeout should be set based on the heartbeat interval and average round-trip time between members.
Election timeouts must be at least 10 times the round-trip time so it can account for variance in the network.
For example, if the round-trip time between members is 10ms then the election timeout should be at least 100ms.
You should also set your election timeout to at least 4 to 5 times your heartbeat interval to account for variance in leader replication.
For a heartbeat interval of 50ms you should set your election timeout to at least 200ms - 250ms.
The election timeout should be set to at least 5 to 10 times the heartbeat interval to account for variance in leader replication.
For a heartbeat interval of 50ms, set the election timeout to at least 250ms - 500ms.
You can override the default values on the command line:
The upper limit of election timeout is 50000ms (50s), which should only be used when deploying a globally-distributed etcd cluster.
A reasonable round-trip time for the continental United States is 130ms, and the time between US and Japan is around 350-400ms.
If the network has uneven performance or regular packet delays/loss then it is possible that a couple of retries may be necessary to successfully send a packet. So 5s is a safe upper limit of global round-trip time.
As the election timeout should be an order of magnitude bigger than broadcast time, in the case of ~5s for a globally distributed cluster, then 50 seconds becomes a reasonable maximum.
The heartbeat interval and election timeout value should be the same for all members in one cluster. Setting different values for etcd members may disrupt cluster stability.
The default values can be overridden on the command line:
etcd clusters can be upgraded by doing a rolling upgrade or all at once. We make every effort to test this process, but please be sure to backup your data [by etcd-dump](https://github.com/AaronO/etcd-dump), or make a copy of data directory beforehand.
## Upgrade Process
- Stop the old etcd processes
- Upgrade the etcd binary
- Restart the etcd instance using the original --name, --address, --peer-address and --data-dir.
## Rolling Upgrade
During an upgrade, etcd clusters are designed to continue working in a mix of old and new versions. It's recommended to converge on the new version quickly. Using new API features before the entire cluster has been upgraded is only supported as a best effort. Each instance's version can be found with `curl http://127.0.0.1:4001/version`.
## All at Once
If downtime is not an issue, the easiest way to upgrade your cluster is to shutdown all of the etcd instances and restart them with the new binary. The current state of the cluster is saved to disk and will be loaded into the cluster when it restarts.
In the general case, upgrading from etcd 2.3 to 3.0 can be a zero-downtime, rolling upgrade:
- one by one, stop the etcd v2.3 processes and replace them with etcd v3.0 processes
- after running all v3.0 processes, new features in v3.0 are available to the cluster
Before [starting an upgrade](#upgrade-procedure), read through the rest of this guide to prepare.
### Upgrade Checklists
#### Upgrade Requirements
To upgrade an existing etcd deployment to 3.0, the running cluster must be 2.3 or greater. If it's before 2.3, please upgrade to [2.3](https://github.com/coreos/etcd/releases/tag/v2.3.0) before upgrading to 3.0.
Also, to ensure a smooth rolling upgrade, the running cluster must be healthy. You can check the health of the cluster by using the `etcdctl cluster-health` command.
#### Preparation
Before upgrading etcd, always test the services relying on etcd in a staging environment before deploying the upgrade to the production environment.
Before beginning, [backup the etcd data directory](admin_guide.md#backing-up-the-datastore). Should something go wrong with the upgrade, it is possible to use this backup to [downgrade](#downgrade) back to existing etcd version.
#### Mixed Versions
While upgrading, an etcd cluster supports mixed versions of etcd members, and operates with the protocol of the lowest common version. The cluster is only considered upgraded once all of its members are upgraded to version 3.0. Internally, etcd members negotiate with each other to determine the overall cluster version, which controls the reported version and the supported features.
#### Limitations
It might take up to 2 minutes for the newly upgraded member to catch up with the existing cluster when the total data size is larger than 50MB. Check the size of a recent snapshot to estimate the total data size. In other words, it is safest to wait for 2 minutes between upgrading each member.
For a much larger total data size, 100MB or more , this one-time process might take even more time. Administrators of very large etcd clusters of this magnitude can feel free to contact the [etcd team][etcd-contact] before upgrading, and we’ll be happy to provide advice on the procedure.
#### Downgrade
If all members have been upgraded to v3.0, the cluster will be upgraded to v3.0, and downgrade from this completed state is **not possible**. If any single member is still v2.3, however, the cluster and its operations remains “v2.3”, and it is possible from this mixed cluster state to return to using a v2.3 etcd binary on all members.
Please [backup the data directory](admin_guide.md#backing-up-the-datastore) of all etcd members to make downgrading the cluster possible even after it has been completely upgraded.
### Upgrade Procedure
This example details the upgrade of a three-member v2.3 ectd cluster running on a local machine.
#### 1. Check upgrade requirements.
Is the the cluster healthy and running v.2.3.x?
```
$ etcdctl cluster-health
member 6e3bd23ae5f1eae0 is healthy: got healthy result from http://localhost:22379
member 924e2e83e93f2560 is healthy: got healthy result from http://localhost:32379
member 8211f1d0f64f3269 is healthy: got healthy result from http://localhost:12379
cluster is healthy
$ curl http://localhost:2379/version
{"etcdserver":"2.3.x","etcdcluster":"2.3.0"}
```
#### 2. Stop the existing etcd process
When each etcd process is stopped, expected errors will be logged by other cluster members. This is normal since a cluster member connection has been (temporarily) broken:
```
2016-06-27 15:21:48.624124 E | rafthttp: failed to dial 8211f1d0f64f3269 on stream Message (dial tcp 127.0.0.1:12380: getsockopt: connection refused)
2016-06-27 15:21:48.624175 I | rafthttp: the connection with 8211f1d0f64f3269 became inactive
```
It’s a good idea at this point to [backup the etcd data directory](https://github.com/coreos/etcd/blob/master/Documentation/v2/admin_guide.md#backing-up-the-datastore) to provide a downgrade path should any problems occur:
```
$ etcdctl backup \
--data-dir /var/lib/etcd \
--backup-dir /tmp/etcd_backup
```
#### 3. Drop-in etcd v3.0 binary and start the new etcd process
The new v3.0 etcd will publish its information to the cluster:
```
09:58:25.938673 I | etcdserver: published {Name:infra1 ClientURLs:[http://localhost:12379]} to cluster 524400597fb1d5f6
```
Verify that each member, and then the entire cluster, becomes healthy with the new v3.0 etcd binary:
```
$ etcdctl cluster-health
member 6e3bd23ae5f1eae0 is healthy: got healthy result from http://localhost:22379
member 924e2e83e93f2560 is healthy: got healthy result from http://localhost:32379
member 8211f1d0f64f3269 is healthy: got healthy result from http://localhost:12379
cluster is healthy
```
Upgraded members will log warnings like the following until the entire cluster is upgraded. This is expected and will cease after all etcd cluster members are upgraded to v3.0:
```
2016-06-27 15:22:05.679644 W | etcdserver: the local etcd version 2.3.7 is not up-to-date
2016-06-27 15:22:05.679660 W | etcdserver: member 8211f1d0f64f3269 has a higher version 3.0.0
```
#### 4. Repeat step 2 to step 3 for all other members
#### 5. Finish
When all members are upgraded, the cluster will report upgrading to 3.0 successfully:
```
2016-06-27 15:22:19.873751 N | membership: updated the cluster version from 2.3 to 3.0
2016-06-27 15:22:19.914574 I | api: enabled capabilities for version 3.0.0
```
```
$ ETCDCTL_API=3 etcdctl endpoint health
127.0.0.1:12379 is healthy: successfully committed proposal: took = 18.440155ms
127.0.0.1:32379 is healthy: successfully committed proposal: took = 13.651368ms
127.0.0.1:22379 is healthy: successfully committed proposal: took = 18.513301ms
You can migrate a snapshot of your data from a v0.4.9+ cluster into a new etcd 2.2 cluster using a snapshot migration. After snapshot migration, the etcd indexes of your data will change. Many etcd applications rely on these indexes to behave correctly. This operation should only be done while all etcd applications are stopped.
To get started get the newest data snapshot from the 0.4.9+ cluster:
If you have a large amount of data, you can specify more concurrent works to copy data in parallel by using `-c` flag.
If you have hidden keys to copy, you can use `--hidden` flag to specify. For example fleet uses `/_coreos.com/fleet` so to import those keys use `--hidden /_coreos.com`.
And the data will quickly copy into the new cluster:
[](https://quay.io/repository/coreos/etcd-git)
**Note**: The `master` branch may be in an *unstable or even broken state* during development. Please use [releases][github-release] instead of the `master` branch in order to get stable binaries.
* *Fast*: benchmarked 1000s of writes/s per instance
* *Reliable*: properly distributed using Raft
etcd is written in Go and uses the [Raft][raft] consensus algorithm to manage a highly-available replicated log.
etcd is used [in production by many companies](./production-users.md), and the development team stands behind it in critical deployment scenarios, where etcd is frequently teamed with applications such as [Kubernetes][k8s], [fleet][fleet], [locksmith][locksmith], [vulcand][vulcand], and many others.
See [etcdctl][etcdctl] for a simple command line client.
Or feel free to just use `curl`, as in the examples below.
The easiest way to get etcd is to use one of the pre-built release binaries which are available for OSX, Linux, Windows, AppC (ACI), and Docker. Instructions for using these binaries are on the [GitHub releases page][github-release].
For those wanting to try the very latest version, you can build the latest version of etcd from the `master` branch.
You will first need [*Go*](https://golang.org/) installed on your machine (version 1.5+ is required).
All development occurs on `master`, including new features and bug fixes.
Bug fixes are first targeted at `master` and subsequently ported to release branches, as described in the [branch management][branch-management] guide.
This will bring up etcd listening on port 2379 for client communication and on port 2380 for server-to-server communication.
Next, let's set a single key, and then retrieve it:
```
curl -L http://127.0.0.1:2379/v2/keys/mykey -XPUT -d value="this is awesome"
curl -L http://127.0.0.1:2379/v2/keys/mykey
```
You have successfully started an etcd and written a key to the store.
### etcd TCP ports
The [official etcd ports][iana-ports] are 2379 for client requests, and 2380 for peer communication. To maintain compatibility, some etcd configuration and documentation continues to refer to the legacy ports 4001 and 7001, but all new etcd use and discussion should adopt the IANA-assigned ports. The legacy ports 4001 and 7001 will be fully deprecated, and support for their use removed, in future etcd releases.
If the previous node is a key and client tries to overwrite it with `dir=true`, it does not give warnings such as `Not a directory`. Instead, the key is set to empty value.
When first started, etcd stores its configuration into a data directory specified by the data-dir configuration parameter.
Configuration is stored in the write ahead log and includes: the local member ID, cluster ID, and initial cluster configuration.
The write ahead log and snapshot files are used during member operation and to recover after a restart.
Having a dedicated disk to store wal files can improve the throughput and stabilize the cluster.
It is highly recommended to dedicate a wal disk and set `--wal-dir` to point to a directory on that device for a production cluster deployment.
If a member’s data directory is ever lost or corrupted then the user should [remove][remove-a-member] the etcd member from the cluster using `etcdctl` tool.
A user should avoid restarting an etcd member with a data directory from an out-of-date backup.
Using an out-of-date data directory can lead to inconsistency as the member had agreed to store information via raft then re-joins saying it needs that information again.
For maximum safety, if an etcd member suffers any sort of data corruption or loss, it must be removed from the cluster.
Once removed the member can be re-added with an empty data directory.
### Contents
The data directory has two sub-directories in it:
1. wal: write ahead log files are stored here. For details see the [wal package documentation][wal-pkg]
2. snap: log snapshots are stored here. For details see the [snap package documentation][snap-pkg]
If `--wal-dir` flag is set, etcd will write the write ahead log files to the specified directory instead of data directory.
## Cluster Management
### Lifecycle
If you are spinning up multiple clusters for testing it is recommended that you specify a unique initial-cluster-token for the different clusters.
This can protect you from cluster corruption in case of mis-configuration because two members started with different cluster tokens will refuse members from each other.
### Monitoring
It is important to monitor your production etcd cluster for healthy information and runtime metrics.
#### Health Monitoring
At lowest level, etcd exposes health information via HTTP at `/health` in JSON format. If it returns `{"health": "true"}`, then the cluster is healthy. Please note the `/health` endpoint is still an experimental one as in etcd 2.2.
```
$ curl -L http://127.0.0.1:2379/health
{"health": "true"}
```
You can also use etcdctl to check the cluster-wide health information. It will contact all the members of the cluster and collect the health information for you.
```
$./etcdctl cluster-health
member 8211f1d0f64f3269 is healthy: got healthy result from http://127.0.0.1:12379
member 91bc3c398fb3c146 is healthy: got healthy result from http://127.0.0.1:22379
member fd422379fda50e48 is healthy: got healthy result from http://127.0.0.1:32379
cluster is healthy
```
#### Runtime Metrics
etcd uses [Prometheus][prometheus] for metrics reporting in the server. You can read more through the runtime metrics [doc][metrics].
### Debugging
Debugging a distributed system can be difficult. etcd provides several ways to make debug
easier.
#### Enabling Debug Logging
When you want to debug etcd without stopping it, you can enable debug logging at runtime.
etcd exposes logging configuration at `/config/local/log`.
Debug variables are exposed for real-time debugging purposes. Developers who are familiar with etcd can utilize these variables to debug unexpected behavior. etcd exposes debug variables via HTTP at `/debug/vars` in JSON format. The debug variables contains
`cmdline`, `file_descriptor_limit`, `memstats` and `raft.status`.
`cmdline` is the command line arguments passed into etcd.
`file_descriptor_limit` is the max number of file descriptors etcd can utilize.
`memstats` is explained in detail in the [Go runtime documentation][golang-memstats].
`raft.status` is useful when you want to debug low level raft issues if you are familiar with raft internals. In most cases, you do not need to check `raft.status`.
The recommended etcd cluster size is 3, 5 or 7, which is decided by the fault tolerance requirement. A 7-member cluster can provide enough fault tolerance in most cases. While larger cluster provides better fault tolerance the write performance reduces since data needs to be replicated to more machines.
#### Fault Tolerance Table
It is recommended to have an odd number of members in a cluster. Having an odd cluster size doesn't change the number needed for majority, but you gain a higher tolerance for failure by adding the extra member. You can see this in practice when comparing even and odd sized clusters:
| Cluster Size | Majority | Failure Tolerance |
|--------------|------------|-------------------|
| 1 | 1 | 0 |
| 2 | 2 | 0 |
| 3 | 2 | **1** |
| 4 | 3 | 1 |
| 5 | 3 | **2** |
| 6 | 4 | 2 |
| 7 | 4 | **3** |
| 8 | 5 | 3 |
| 9 | 5 | **4** |
As you can see, adding another member to bring the size of cluster up to an odd size is always worth it. During a network partition, an odd number of members also guarantees that there will almost always be a majority of the cluster that can continue to operate and be the source of truth when the partition ends.
#### Changing Cluster Size
After your cluster is up and running, adding or removing members is done via [runtime reconfiguration][runtime-reconfig], which allows the cluster to be modified without downtime. The `etcdctl` tool has `member list`, `member add` and `member remove` commands to complete this process.
### Member Migration
When there is a scheduled machine maintenance or retirement, you might want to migrate an etcd member to another machine without losing the data and changing the member ID.
The data directory contains all the data to recover a member to its point-in-time state. To migrate a member:
* Stop the member process.
* Copy the data directory of the now-idle member to the new machine.
* Update the peer URLs for the replaced member to reflect the new machine according to the [runtime reconfiguration instructions][update-a-member].
* Start etcd on the new machine, using the same configuration and the copy of the data directory.
This example will walk you through the process of migrating the infra1 member to a new machine:
etcd is designed to be resilient to machine failures. An etcd cluster can automatically recover from any number of temporary failures (for example, machine reboots), and a cluster of N members can tolerate up to _(N-1)/2_ permanent failures (where a member can no longer access the cluster, due to hardware failure or disk corruption). However, in extreme circumstances, a cluster might permanently lose enough members such that quorum is irrevocably lost. For example, if a three-node cluster suffered two simultaneous and unrecoverable machine failures, it would be normally impossible for the cluster to restore quorum and continue functioning.
To recover from such scenarios, etcd provides functionality to backup and restore the datastore and recreate the cluster without data loss.
#### Backing up the datastore
**NB:** Windows users must stop etcd before running the backup command.
The first step of the recovery is to backup the data directory and wal directory, if stored separately, on a functioning etcd node. To do this, use the `etcdctl backup` command, passing in the original data (and wal) directory used by etcd. For example:
```sh
etcdctl backup \
--data-dir %data_dir% \
[--wal-dir %wal_dir%]\
--backup-dir %backup_data_dir%
[--backup-wal-dir %backup_wal_dir%]
```
This command will rewrite some of the metadata contained in the backup (specifically, the node ID and cluster ID), which means that the node will lose its former identity. In order to recreate a cluster from the backup, you will need to start a new, single-node cluster. The metadata is rewritten to prevent the new node from inadvertently being joined onto an existing cluster.
#### Restoring a backup
To restore a backup using the procedure created above, start etcd with the `-force-new-cluster` option and pointing to the backup directory. This will initialize a new, single-member cluster with the default advertised peer URLs, but preserve the entire contents of the etcd data store. Continuing from the previous example:
```sh
etcd \
-data-dir=%backup_data_dir% \
[-wal-dir=%backup_wal_dir%]\
-force-new-cluster \
...
```
Now etcd should be available on this node and serving the original datastore.
Once you have verified that etcd has started successfully, shut it down and move the data and wal, if stored separately, back to the previous location (you may wish to make another copy as well to be safe):
```sh
pkill etcd
rm -fr %data_dir%
rm -fr %wal_dir%
mv %backup_data_dir% %data_dir%
mv %backup_wal_dir% %wal_dir%
etcd \
-data-dir=%data_dir% \
[-wal-dir=%wal_dir%]\
...
```
#### Restoring the cluster
Now that the node is running successfully, [change its advertised peer URLs][update-a-member], as the `--force-new-cluster` option has set the peer URL to the default listening on localhost.
You can then add more nodes to the cluster and restore resiliency. See the [add a new member][add-a-member] guide for more details. **NB:** If you are trying to restore your cluster using old failed etcd nodes, please make sure you have stopped old etcd instances and removed their old data directories specified by the data-dir configuration parameter.
### Client Request Timeout
etcd sets different timeouts for various types of client requests. The timeout value is not tunable now, which will be improved soon (https://github.com/coreos/etcd/issues/2038).
#### Get requests
Timeout is not set for get requests, because etcd serves the result locally in a non-blocking way.
**Note**: QuorumGet request is a different type, which is mentioned in the following sections.
#### Watch requests
Timeout is not set for watch requests. etcd will not stop a watch request until client cancels it, or the connection is broken.
#### Delete, Put, Post, QuorumGet requests
The default timeout is 5 seconds. It should be large enough to allow all key modifications if the majority of cluster is functioning.
If the request times out, it indicates two possibilities:
1. the server the request sent to was not functioning at that time.
2. the majority of the cluster is not functioning.
If timeout happens several times continuously, administrators should check status of cluster and resolve it as soon as possible.
### Best Practices
#### Maximum OS threads
By default, etcd uses the default configuration of the Go 1.4 runtime, which means that at most one operating system thread will be used to execute code simultaneously. (Note that this default behavior [has changed in Go 1.5][golang1.5-runtime]).
When using etcd in heavy-load scenarios on machines with multiple cores it will usually be desirable to increase the number of threads that etcd can utilize. To do this, simply set the environment variable GOMAXPROCS to the desired number when starting etcd. For more information on this variable, see the [Go runtime documentation][golang-runtime].
-`X-Etcd-Index` is the current etcd index as explained above.
-`X-Raft-Index` is similar to the etcd index but is for the underlying raft protocol
-`X-Raft-Term` is an integer that will increase whenever an etcd master election happens in the cluster. If this number is increasing rapidly, you may need to tune the election timeout. See the [tuning][tuning] section for details.
[tuning]: #tuning
*`X-Etcd-Index` is the current etcd index as explained above. When request is a watch on key space, `X-Etcd-Index` is the current etcd index when the watch starts, which means that the watched event may happen after `X-Etcd-Index`.
*`X-Raft-Index` is similar to the etcd index but is for the underlying raft protocol.
*`X-Raft-Term` is an integer that will increase whenever an etcd master election happens in the cluster. If this number is increasing rapidly, you may need to tune the election timeout. See the [tuning][tuning] section for details.
We get the index is outdated response, since we miss the 1000 events kept in etcd.
```
{"errorCode":401,"message":"The event in requested index is outdated and cleared","cause":"the requested history has been cleared [1008/8]","index":2007}
```
To start watch, first we need to fetch the current state of key `/foo`:
@ -424,13 +542,15 @@ etcd can be used as a centralized coordination service in a cluster, and `Compar
This command will set the value of a key only if the client-provided conditions are equal to the current conditions.
*Note that `CompareAndSwap` does not work with [directories][directories]. If an attempt is made to `CompareAndSwap` a directory, a 102 "Not a file" error will be returned.*
The current comparable conditions are:
1. `prevValue` - checks the previous value of the key.
2. `prevIndex` - checks the previous modifiedIndex of the key.
3. `prevExist` - checks existence of the key: if `prevExist` is true, it is an `update` request; if prevExist is `false`, it is a `create` request.
3. `prevExist` - checks existence of the key: if `prevExist` is true, it is an `update` request; if `prevExist` is `false`, it is a `create` request.
Here is a simple example.
Let's create a key-value pair first: `foo=one`.
@ -509,6 +629,8 @@ We successfully changed the value from "one" to "two" since we gave the correct
This command will delete a key only if the client-provided conditions are equal to the current conditions.
*Note that `CompareAndDelete` does not work with [directories]. If an attempt is made to `CompareAndDelete` a directory, a 102 "Not a file" error will be returned.*
The current comparable conditions are:
1. `prevValue` - checks the previous value of the key.
@ -639,22 +761,22 @@ We should see the response as an array of items:
etcd is designed to reliably store infrequently updated data and provide reliable watch queries. etcd exposes previous versions of key-value pairs to support inexpensive snapshots and watch history events (“time travel queries”). A persistent, multi-version, concurrency-control data model is a good fit for these use cases.
etcd stores data in a multiversion [persistent][persistent-ds] key-value store. The persistent key-value store preserves the previous version of a key-value pair when its value is superseded with new data. The key-value store is effectively immutable; its operations do not update the structure in-place, but instead always generates a new updated structure. All past versions of keys are still accessible and watchable after modification. To prevent the data store from growing indefinitely over time from maintaining old versions, the store may be compacted to shed the oldest versions of superseded data.
### Logical View
The store’s logical view is a flat binary key space. The key space has a lexically sorted index on byte string keys so range queries are inexpensive.
The key space maintains multiple revisions. Each atomic mutative operation (e.g., a transaction operation may contain multiple operations) creates a new revision on the key space. All data held by previous revisions remains unchanged. Old versions of key can still be accessed through previous revisions. Likewise, revisions are indexed as well; ranging over revisions with watchers is efficient. If the store is compacted to recover space, revisions before the compact revision will be removed.
A key’s lifetime spans a generation. Each key may have one or multiple generations. Creating a key increments the generation of that key, starting at 1 if the key never existed. Deleting a key generates a key tombstone, concluding the key’s current generation. Each modification of a key creates a new version of the key. Once a compaction happens, any generation ended before the given revision will be removed and values set before the compaction revision except the latest one will be removed.
### Physical View
etcd stores the physical data as key-value pairs in a persistent [b+tree][b+tree]. Each revision of the store’s state only contains the delta from its previous revision to be efficient. A single revision may correspond to multiple keys in the tree.
The key of key-value pair is a 3-tuple (major, sub, type). Major is the store revision holding the key. Sub differentiates among keys within the same revision. Type is an optional suffix for special value (e.g., `t` if the value contains a tombstone). The value of the key-value pair contains the modification from previous revision, thus one delta from previous revision. The b+tree is ordered by key in lexical byte-order. Ranged lookups over revision deltas are fast; this enables quickly finding modifications from one specific revision to another. Compaction removes out-of-date keys-value pairs.
etcd also keeps a secondary in-memory [btree][btree] index to speed up range queries over keys. The keys in the btree index are the keys of the store exposed to user. The value is a pointer to the modification of the persistent b+tree. Compaction removes dead pointers.
## KV API Guarantees
etcd is a consistent and durable key value store with mini-transaction(TODO: link to txn doc when we have it) support. The key value store is exposed through the KV APIs. etcd tries to ensure the strongest consistency and durability guarantees for a distributed system. This specification enumerates the KV API guarantees made by etcd.
### APIs to consider
* Read APIs
* range
* watch
* Write APIs
* put
* delete
* Combination (read-modify-write) APIs
* txn
### etcd Specific Definitions
#### operation completed
An etcd operation is considered complete when it is committed through consensus, and therefore “executed” -- permanently stored -- by the etcd storage engine. The client knows an operation is completed when it receives a response from the etcd server. Note that the client may be uncertain about the status of an operation if it times out, or there is a network disruption between the client and the etcd member. etcd may also abort operations when there is a leader election. etcd does not send `abort` responses to clients’ outstanding requests in this event.
#### revision
An etcd operation that modifies the key value store is assigned with a single increasing revision. A transaction operation might modifies the key value store multiple times, but only one revision is assigned. The revision attribute of a key value pair that modified by the operation has the same value as the revision of the operation. The revision can be used as a logical clock for key value store. A key value pair that has a larger revision is modified after a key value pair with a smaller revision. Two key value pairs that have the same revision are modified by an operation "concurrently".
### Guarantees Provided
#### Atomicity
All API requests are atomic; an operation either completes entirely or not at all. For watch requests, all events generated by one operation will be in one watch response. Watch never observes partial events for a single operation.
#### Consistency
All API calls ensure [sequential consistency][seq_consistency], the strongest consistency guarantee available from distributed systems. No matter which etcd member server a client makes requests to, a client reads the same events in the same order. If two members complete the same number of operations, the state of the two members is consistent.
For watch operations, etcd guarantees to return the same value for the same key across all members for the same revision. For range operations, etcd has a similar guarantee for [linearized][Linearizability] access; serialized access may be behind the quorum state, so that the later revision is not yet available.
As with all distributed systems, it is impossible for etcd to ensure [strict consistency][strict_consistency]. etcd does not guarantee that it will return to a read the “most recent” value (as measured by a wall clock when a request is completed) available on any cluster member.
#### Isolation
etcd ensures [serializable isolation][serializable_isolation], which is the highest isolation level available in distributed systems. Read operations will never observe any intermediate data.
#### Durability
Any completed operations are durable. All accessible data is also durable data. A read will never return data that has not been made durable.
#### Linearizability
Linearizability (also known as Atomic Consistency or External Consistency) is a consistency level between strict consistency and sequential consistency.
For linearizability, suppose each operation receives a timestamp from a loosely synchronized global clock. Operations are linearized if and only if they always complete as though they were executed in a sequential order and each operation appears to complete in the order specified by the program. Likewise, if an operation’s timestamp precedes another, that operation must also precede the other operation in the sequence.
For example, consider a client completing a write at time point 1 (*t1*). A client issuing a read at *t2* (for *t2* > *t1*) should receive a value at least as recent as the previous write, completed at *t1*. However, the read might actually complete only by *t3*, and the returned value, current at *t2* when the read began, might be "stale" by *t3*.
etcd does not ensure linearizability for watch operations. Users are expected to verify the revision of watch responses to ensure correct ordering.
etcd ensures linearizability for all other operations by default. Linearizability comes with a cost, however, because linearized requests must go through the Raft consensus process. To obtain lower latencies and higher throughput for read requests, clients can configure a request’s consistency mode to `serializable`, which may access stale data with respect to quorum, but removes the performance penalty of linearized accesses' reliance on live consensus.
A user is an identity to be authenticated. Each user can have multiple roles. The user has a capability (such as reading or writing) on the resource if one of the roles has that capability.
A user named `root` is required before authentication can be enabled, and it always has the ROOT role. The ROOT role can be granted to multiple users, but `root` is required for recovery purposes.
#### Roles
Each role has exact one associated Permission List. An permission list exists for each permission on key-value resources.
The special static ROOT (named `root`) role has a full permissions on all key-value resources, the permission to manage user resources and settings resources. Only the ROOT role has the permission to manage user resources and modify settings resources. The ROOT role is built-in and does not need to be created.
There is also a special GUEST role, named 'guest'. These are the permissions given to unauthenticated requests to etcd. This role will be created automatically, and by default allows access to the full keyspace due to backward compatibility. (etcd did not previously authenticate any actions.). This role can be modified by a ROOT role holder at any time, to reduce the capabilities of unauthenticated users.
#### Permissions
There are two types of permissions, `read` and `write`. All management and settings require the ROOT role.
A Permission List is a list of allowed patterns for that particular permission (read or write). Only ALLOW prefixes are supported. DENY becomes more complicated and is TBD.
### Key-Value Resources
A key-value resource is a key-value pairs in the store. Given a list of matching patterns, permission for any given key in a request is granted if any of the patterns in the list match.
Only prefixes or exact keys are supported. A prefix permission string ends in `*`.
A permission on `/foo` is for that exact key or directory, not its children or recursively. `/foo*` is a prefix that matches `/foo` recursively, and all keys thereunder, and keys with that prefix (eg. `/foobar`. Contrast to the prefix `/foo/*`). `*` alone is permission on the full keyspace.
### Settings Resources
Specific settings for the cluster as a whole. This can include adding and removing cluster members, enabling or disabling authentication, replacing certificates, and any other dynamic configuration by the administrator (holder of the ROOT role).
## v2 Auth
### Basic Auth
We only support [Basic Auth][basic-auth] for the first version. Client needs to attach the basic auth to the HTTP Authorization Header.
### Authorization field for operations
Added to requests to /v2/keys, /v2/auth
Add code 401 Unauthorized to the set of responses from the v2 API
Authorization: Basic {encoded string}
### Future Work
Other types of auth can be considered for the future (eg, signed certs, public keys) but the `Authorization:` header allows for other such types
### Things out of Scope for etcd Permissions
* Pluggable AUTH backends like LDAP (other Authorization tokens generated by LDAP et al may be a possibility)
* Very fine-grained access controls (eg: users modifying keys outside work hours)
## API endpoints
An Error JSON corresponds to:
{
"name": "ErrErrorName",
"description" : "The longer helpful description of the error."
}
#### Enable and Disable Authentication
**Get auth status**
GET /v2/auth/enable
Sent Headers:
Possible Status Codes:
200 OK
200 Body:
{
"enabled": true
}
**Enable auth**
PUT /v2/auth/enable
Sent Headers:
Put Body: (empty)
Possible Status Codes:
200 OK
400 Bad Request (if root user has not been created)
409 Conflict (already enabled)
200 Body: (empty)
**Disable auth**
DELETE /v2/auth/enable
Sent Headers:
Authorization: Basic <RootAuthString>
Possible Status Codes:
200 OK
401 Unauthorized (if not a root user)
409 Conflict (already disabled)
200 Body: (empty)
#### Users
The User JSON object is formed as follows:
```
{
"user": "userName",
"password": "password",
"roles": [
"role1",
"role2"
],
"grant": [],
"revoke": []
}
```
Password is only passed when necessary.
**Get a List of Users**
GET/HEAD /v2/auth/users
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
200 Headers:
Content-type: application/json
200 Body:
{
"users": [
{
"user": "alice",
"roles": [
{
"role": "root",
"permissions": {
"kv": {
"read": ["/*"],
"write": ["/*"]
}
}
}
]
},
{
"user": "bob",
"roles": [
{
"role": "guest",
"permissions": {
"kv": {
"read": ["/*"],
"write": ["/*"]
}
}
}
]
}
]
}
**Get User Details**
GET/HEAD /v2/auth/users/alice
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
404 Not Found
200 Headers:
Content-type: application/json
200 Body:
{
"user" : "alice",
"roles" : [
{
"role": "fleet",
"permissions" : {
"kv" : {
"read": [ "/fleet/" ],
"write": [ "/fleet/" ]
}
}
},
{
"role": "etcd",
"permissions" : {
"kv" : {
"read": [ "/*" ],
"write": [ "/*" ]
}
}
}
]
}
**Create Or Update A User**
A user can be created with initial roles, if filled in. However, no roles are required; only the username and password fields
PUT /v2/auth/users/charlie
Sent Headers:
Authorization: Basic <BasicAuthString>
Put Body:
JSON struct, above, matching the appropriate name
* Starting password and roles when creating.
* Grant/Revoke/Password filled in when updating (to grant roles, revoke roles, or change the password).
Possible Status Codes:
200 OK
201 Created
400 Bad Request
401 Unauthorized
404 Not Found (update non-existent users)
409 Conflict (when granting duplicated roles or revoking non-existent roles)
200 Headers:
Content-type: application/json
200 Body:
JSON state of the user
**Remove A User**
DELETE /v2/auth/users/charlie
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
403 Forbidden (remove root user when auth is enabled)
404 Not Found
200 Headers:
200 Body: (empty)
#### Roles
A full role structure may look like this. A Permission List structure is used for the "permissions", "grant", and "revoke" keys.
```
{
"role" : "fleet",
"permissions" : {
"kv" : {
"read" : [ "/fleet/" ],
"write": [ "/fleet/" ]
}
},
"grant" : {"kv": {...}},
"revoke": {"kv": {...}}
}
```
**Get Role Details**
GET/HEAD /v2/auth/roles/fleet
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
404 Not Found
200 Headers:
Content-type: application/json
200 Body:
{
"role" : "fleet",
"permissions" : {
"kv" : {
"read": [ "/fleet/" ],
"write": [ "/fleet/" ]
}
}
}
**Get a list of Roles**
GET/HEAD /v2/auth/roles
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
200 Headers:
Content-type: application/json
200 Body:
{
"roles": [
{
"role": "fleet",
"permissions": {
"kv": {
"read": ["/fleet/"],
"write": ["/fleet/"]
}
}
},
{
"role": "etcd",
"permissions": {
"kv": {
"read": ["/*"],
"write": ["/*"]
}
}
},
{
"role": "quay",
"permissions": {
"kv": {
"read": ["/*"],
"write": ["/*"]
}
}
}
]
}
**Create Or Update A Role**
PUT /v2/auth/roles/rkt
Sent Headers:
Authorization: Basic <BasicAuthString>
Put Body:
Initial desired JSON state, including the role name for verification and:
* Starting permission set if creating
* Granted/Revoked permission set if updating
Possible Status Codes:
200 OK
201 Created
400 Bad Request
401 Unauthorized
404 Not Found (update non-existent roles)
409 Conflict (when granting duplicated permission or revoking non-existent permission)
200 Body:
JSON state of the role
**Remove A Role**
DELETE /v2/auth/roles/rkt
Sent Headers:
Authorization: Basic <BasicAuthString>
Possible Status Codes:
200 OK
401 Unauthorized
403 Forbidden (remove root)
404 Not Found
200 Headers:
200 Body: (empty)
## Example Workflow
Let's walk through an example to show two tenants (applications, in our case) using etcd permissions.
### Create root role
```
PUT /v2/auth/users/root
Put Body:
{"user" : "root", "password": "betterRootPW!"}
```
### Enable auth
```
PUT /v2/auth/enable
```
### Modify guest role (revoke write permission)
```
PUT /v2/auth/roles/guest
Headers:
Authorization: Basic <root:betterRootPW!>
Put Body:
{
"role" : "guest",
"revoke" : {
"kv" : {
"write": [
"/*"
]
}
}
}
```
### Create Roles for the Applications
Create the rkt role fully specified:
```
PUT /v2/auth/roles/rkt
Headers:
Authorization: Basic <root:betterRootPW!>
Body:
{
"role" : "rkt",
"permissions" : {
"kv": {
"read": [
"/rkt/*"
],
"write": [
"/rkt/*"
]
}
}
}
```
But let's make fleet just a basic role for now:
```
PUT /v2/auth/roles/fleet
Headers:
Authorization: Basic <root:betterRootPW!>
Body:
{
"role" : "fleet"
}
```
### Optional: Grant some permissions to the roles
Well, we finally figured out where we want fleet to live. Let's fix it.
(Note that we avoided this in the rkt case. So this step is optional.)
```
PUT /v2/auth/roles/fleet
Headers:
Authorization: Basic <root:betterRootPW!>
Put Body:
{
"role" : "fleet",
"grant" : {
"kv" : {
"read": [
"/rkt/fleet",
"/fleet/*"
]
}
}
}
```
### Create Users
Same as before, let's use rocket all at once and fleet separately
Authentication -- having users and roles in etcd -- was added in etcd 2.1. This guide will help you set up basic authentication in etcd.
etcd before 2.1 was a completely open system; anyone with access to the API could change keys. In order to preserve backward compatibility and upgradability, this feature is off by default.
For a full discussion of the RESTful API, see [the authentication API documentation][auth-api]
## Special Users and Roles
There is one special user, `root`, and there are two special roles, `root` and `guest`.
### User `root`
User `root` must be created before security can be activated. It has the `root` role and allows for the changing of anything inside etcd. The idea behind the `root` user is for recovery purposes -- a password is generated and stored somewhere -- and the root role is granted to the administrator accounts on the system. In the future, for troubleshooting and recovery, we will need to assume some access to the system, and future documentation will assume this root user (though anyone with the role will suffice).
### Role `root`
Role `root` cannot be modified, but it may be granted to any user. Having access via the root role not only allows global read-write access (as was the case before 2.1) but allows modification of the authentication policy and all administrative things, like modifying the cluster membership.
### Role `guest`
The `guest` role defines the permissions granted to any request that does not provide an authentication. This will be created on security activation (if it doesn't already exist) to have full access to all keys, as was true in etcd 2.0. It may be modified at any time, and cannot be removed.
## Working with users
The `user` subcommand for `etcdctl` handles all things having to do with user accounts.
A listing of users can be found with
```
$ etcdctl user list
```
Creating a user is as easy as
```
$ etcdctl user add myusername
```
And there will be prompt for a new password.
Roles can be granted and revoked for a user with
```
$ etcdctl user grant myusername -roles foo,bar,baz
$ etcdctl user revoke myusername -roles bar,baz
```
We can look at this user with
```
$ etcdctl user get myusername
```
And the password for a user can be changed with
```
$ etcdctl user passwd myusername
```
Which will prompt again for a new password.
To delete an account, there's always
```
$ etcdctl user remove myusername
```
## Working with roles
The `role` subcommand for `etcdctl` handles all things having to do with access controls for particular roles, as were granted to individual users.
A listing of roles can be found with
```
$ etcdctl role list
```
A new role can be created with
```
$ etcdctl role add myrolename
```
A role has no password; we are merely defining a new set of access rights.
Roles are granted access to various parts of the keyspace, a single path at a time.
Reading a path is simple; if the path ends in `*`, that key **and all keys prefixed with it**, are granted to holders of this role. If it does not end in `*`, only that key and that key alone is granted.
Access can be granted as either read, write, or both, as in the following examples:
```
# Give read access to keys under the /foo directory
$ etcdctl role grant myrolename -path '/foo/*' -read
# Give write-only access to the key at /foo/bar
$ etcdctl role grant myrolename -path '/foo/bar' -write
# Give full access to keys under /pub
$ etcdctl role grant myrolename -path '/pub/*' -readwrite
```
Beware that
```
# Give full access to keys under /pub??
$ etcdctl role grant myrolename -path '/pub*' -readwrite
```
Without the slash may include keys under `/publishing`, for example. To do both, grant `/pub` and `/pub/*`
To see what's granted, we can look at the role at any time:
```
$ etcdctl role get myrolename
```
Revocation of permissions is done the same logical way:
```
$ etcdctl role revoke myrolename -path '/foo/bar' -write
```
As is removing a role entirely
```
$ etcdctl role remove myrolename
```
## Enabling authentication
The minimal steps to enabling auth are as follows. The administrator can set up users and roles before or after enabling authentication, as a matter of preference.
Make sure the root user is created:
```
$ etcdctl user add root
New password:
```
And enable authentication
```
$ etcdctl auth enable
```
After this, etcd is running with authentication enabled. To disable it for any reason, use the reciprocal command:
```
$ etcdctl -u root:rootpw auth disable
```
It would also be good to check what guests (unauthenticated users) are allowed to do:
```
$ etcdctl -u root:rootpw role get guest
```
And modify this role appropriately, depending on your policies.
## Using `etcdctl` to authenticate
`etcdctl` supports a similar flag as `curl` for authentication.
```
$ etcdctl -u user:password get foo
```
or if you prefer to be prompted:
```
$ etcdctl -u user get foo
```
Otherwise, all `etcdctl` commands remain the same. Users and roles can still be created and modified, but require authentication by a user with the root role.
The main goal of etcd 2.0 release is to improve cluster safety around bootstrapping and dynamic reconfiguration. To do this, we deprecated the old error-prone APIs and provide a new set of APIs.
The other main focus of this release was a more reliable Raft implementation, but as this change is internal it should not have any notable effects to users.
## Command Line Flags Changes
The major flag changes are to mostly related to bootstrapping. The `initial-*` flags provide an improved way to specify the required criteria to start the cluster. The advertised URLs now support a list of values instead of a single value, which allows etcd users to gracefully migrate to the new set of IANA-assigned ports (2379/client and 2380/peers) while maintaining backward compatibility with the old ports.
-`-addr` is replaced by `-advertise-client-urls`.
-`-bind-addr` is replaced by `-listen-client-urls`.
-`-peer-addr` is replaced by `-initial-advertise-peer-urls`.
-`-peer-bind-addr` is replaced by `-listen-peer-urls`.
-`-peers` is replaced by `-initial-cluster`.
-`-peers-file` is replaced by `-initial-cluster`.
-`-peer-heartbeat-interval` is replaced by `-heartbeat-interval`.
-`-peer-election-timeout` is replaced by `-election-timeout`.
The documentation of new command line flags can be found at
The default data dir location has changed from {$hostname}.etcd to {name}.etcd.
## Key-Value API
### Read consistency flag
The consistent flag for read operations is removed in etcd 2.0.0. The normal read operations provides the same consistency guarantees with the 0.4.6 read operations with consistent flag set.
The read consistency guarantees are:
The consistent read guarantees the sequential consistency within one client that talks to one etcd server. Read/Write from one client to one etcd member should be observed in order. If one client write a value to an etcd server successfully, it should be able to get the value out of the server immediately.
Each etcd member will proxy the request to leader and only return the result to user after the result is applied on the local member. Thus after the write succeed, the user is guaranteed to see the value on the member it sent the request to.
Reads do not provide linearizability. If you want linearizable read, you need to set quorum option to true.
**Previous behavior**
We added an option for a consistent read in the old version of etcd since etcd 0.x redirects the write request to the leader. When the user get back the result from the leader, the member it sent the request to originally might not apply the write request yet. With the consistent flag set to true, the client will always send read request to the leader. So one client should be able to see its last write when consistent=true is enabled. There is no order guarantees among different clients.
## Standby
etcd 0.4’s standby mode has been deprecated. [Proxy mode][proxymode] is introduced to solve a subset of problems standby was solving.
Standby mode was intended for large clusters that had a subset of the members acting in the consensus process. Overall this process was too magical and allowed for operators to back themselves into a corner.
Proxy mode in 2.0 will provide similar functionality, and with improved control over which machines act as proxies due to the operator specifically configuring them. Proxies also support read only or read/write modes for increased security and durability.
[proxymode]: proxy.md
## Discovery Service
A size key needs to be provided inside a [discovery token][discoverytoken].
`v2/admin` on peer url and `v2/keys/_etcd` are unified under the new [v2/members API][members-api] to better explain which machines are part of an etcd cluster, and to simplify the keyspace for all your use cases.
[members-api]: members_api.md
## HTTP Key Value API
- The follower can now transparently proxy write requests to the leader. Clients will no longer see 307 redirections to the leader from etcd.
- Expiration time is in UTC instead of local time.
- 1x dedicated local SSD mounted under /var/lib/etcd
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
- etcd version 2.1.0 alpha
## etcd Cluster
3 etcd members, each runs on a single machine
## Testing
Bootstrap another machine and use the [boom HTTP benchmark tool][boom] to send requests to each etcd member. Check the [benchmark hacking guide][hack-benchmark] for detailed instructions.
## Performance
### reading one single key
| key size in bytes | number of clients | target etcd server | read QPS | 90th Percentile Latency (ms) |
- 1x dedicated local SSD mounted as etcd data directory
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
## etcd Cluster
3 etcd 2.2.0 members, each runs on a single machine.
Detailed versions:
```
etcd Version: 2.2.0
Git SHA: e4561dd
Go Version: go1.5
Go OS/Arch: linux/amd64
```
## Testing
Bootstrap another machine, outside of the etcd cluster, and run the [`boom` HTTP benchmark tool](https://github.com/rakyll/boom) with a connection reuse patch to send requests to each etcd cluster member. See the [benchmark instructions](../../hack/benchmark/) for the patch and the steps to reproduce our procedures.
The performance is calulated through results of 100 benchmark rounds.
## Performance
### Single Key Read Performance
| key size in bytes | number of clients | target etcd server | average read QPS | read QPS stddev | average 90th Percentile Latency (ms) | latency stddev |
- Because etcd now records metrics for each API call, read QPS performance seems to see a minor decrease in most scenarios. This minimal performance impact was judged a reasonable investment for the breadth of monitoring and debugging information returned.
- Write QPS to cluster leaders seems to be increased by a small margin. This is because the main loop and entry apply loops were decoupled in the etcd raft logic, eliminating several blocks between them.
- Write QPS to all members seems to be increased by a significant margin, because followers now receive the latest commit index sooner, and commit proposals more quickly.
- 1x dedicated local SSD mounted under /var/lib/etcd
- 1x dedicated slow disk for the OS
- 1.8 GB memory
- 2x CPUs
## etcd Cluster
3 etcd 2.2.0-rc members, each runs on a single machine.
Detailed versions:
```
etcd Version: 2.2.0-alpha.1+git
Git SHA: 59a5a7e
Go Version: go1.4.2
Go OS/Arch: linux/amd64
```
Also, we use 3 etcd 2.1.0 alpha-stage members to form cluster to get base performance. etcd's commit head is at [c7146bd5][c7146bd5], which is the same as the one that we use in [etcd 2.1 benchmark][etcd-2.1-benchmark].
## Testing
Bootstrap another machine and use the [boom HTTP benchmark tool][boom] to send requests to each etcd member. Check the [benchmark hacking guide][hack-benchmark] for detailed instructions.
## Performance
### reading one single key
| key size in bytes | number of clients | target etcd server | read QPS | 90th Percentile Latency (ms) |
- read QPS in most scenarios is decreased by 5~8%. The reason is that etcd records store metrics for each store operation. The metrics is important for monitoring and debugging, so this is acceptable.
- write QPS to leader is increased by 20~30%. This is because we decouple raft main loop and entry apply loop, which avoids them blocking each other.
- write QPS to all servers is increased by 30~80% because follower could receive latest commit index earlier and commit proposals faster.
- When etcd reaches data size threshold, it may trigger leader election easily and drop part of proposals.
- At most cases, etcd cluster should work smoothly if it doesn't hit the threshold. If it doesn't work well due to insufficient resources, you need to decrease its data size.
| value bytes | key number limitation | suggested data size threshold(MB) | consumed RSS(MB) |
*NOTE*: The watch features are under active development, and their memory usage may change as that development progresses. We do not expect it to significantly increase beyond the figures stated below.
A primary goal of etcd is supporting a very large number of watchers doing a massively large amount of watching. etcd aims to support O(10k) clients, O(100K) watch streams (O(10) streams per client) and O(10M) total watchings (O(100) watching per stream). The memory consumed by each individual watching accounts for the largest portion of etcd's overall usage, and is therefore the focus of current and future optimizations.
Three related components of etcd watch consume physical memory: each `grpc.Conn`, each watch stream, and each instance of the watching activity. `grpc.Conn` maintains the actual TCP connection and other gRPC connection state. Each `grpc.Conn` consumes O(10kb) of memory, and might have multiple watch streams attached.
Each watch stream is an independent HTTP2 connection which consumes another O(10kb) of memory.
Multiple watchings might share one watch stream.
Watching is the actual struct that tracks the changes on the key-value store. Each watching should only consume <O(1kb).
Two components of etcd storage consume physical memory. The etcd process allocates an *in-memory index* to speed key lookup. The process's *page cache*, managed by the operating system, stores recently-accessed data from disk for quick re-use.
The in-memory index holds all the keys in a [B-tree][btree] data structure, along with pointers to the on-disk data (the values). Each key in the B-tree may contain multiple pointers, pointing to different versions of its values. The theoretical memory consumption of the in-memory index can hence be approximated with the formula:
where `c1` is the key metadata overhead and `c2` is the version metadata overhead.
The graph shows the detailed structure of the in-memory index B-tree.
```
In mem index
+------------+
| key || ... |
+--------------+ | || |
| | +------------+
| | | v1 || ... |
| disk <----------------| || | Tree Node
| | +------------+
| | | v2 || ... |
| <----------------+ || |
| | +------------+
+--------------+ +-----+ | | |
| | | | |
| +------------+
|
|
^
------+
| ... |
| |
+-----+
| ... | Tree Node
| |
+-----+
| ... |
| |
------+
```
[Page cache memory][pagecache] is managed by the operating system and is not covered in detail in this document.
## Testing Environment
etcd version
- git head https://github.com/coreos/etcd/commit/776e9fb7be7eee5e6b58ab977c8887b4fe4d48db
GCE n1-standard-2 machine type
- 7.5 GB memory
- 2x CPUs
## In-memory index memory usage
In this test, we only benchmark the memory usage of the in-memory index. The goal is to find `c1` and `c2` mentioned above and to understand the hard limit of memory consumption of the storage.
We calculate the memory usage consumption via the Go runtime.ReadMemStats. We calculate the total allocated bytes difference before creating the index and after creating the index. It cannot perfectly reflect the memory usage of the in-memory index itself but can show the rough consumption pattern.
| N | versions | key size | memory usage |
|------|----------|----------|--------------|
| 100K | 1 | 64bytes | 22MB |
| 100K | 5 | 64bytes | 39MB |
| 1M | 1 | 64bytes | 218MB |
| 1M | 5 | 64bytes | 432MB |
| 100K | 1 | 256bytes | 41MB |
| 100K | 5 | 256bytes | 65MB |
| 1M | 1 | 256bytes | 409MB |
| 1M | 5 | 256bytes | 506MB |
Based on the result, we can calculate `c1=120bytes`, `c2=30bytes`. We only need two sets of data to calculate `c1` and `c2`, since they are the only unknown variable in the formula. The `c1=120bytes` and `c2=30bytes` are the average value of the 4 sets of `c1` and `c2` we calculated. The key metadata overhead is still relatively nontrivial (50%) for small key-value pairs. However, this is a significant improvement over the old store, which had at least 1000% overhead.
## Overall memory usage
The overall memory usage captures how much RSS etcd consumes with the storage. The value size should have very little impact on the overall memory usage of etcd, since we keep values on disk and only retain hot values in memory, managed by the OS page cache.
| N | versions | key size | value size | memory usage |
Based on the result, we know the value size does not significantly impact the memory consumption. There is some minor increase due to more data held in the OS page cache.
* New development occurs on the [master branch][master].
* Master branch should always have a green build!
* Backwards-compatible bug fixes should target the master branch and subsequently be ported to stable branches.
* Once the master branch is ready for release, it will be tagged and become the new stable branch.
The etcd team has adopted a *rolling release model* and supports one stable version of etcd.
### Master branch
The `master` branch is our development branch. All new features land here first.
If you want to try new features, pull `master` and play with it. Note that `master` may not be stable because new features may introduce bugs.
Before the release of the next stable version, feature PRs will be frozen. We will focus on the testing, bug-fix and documentation for one to two weeks.
### Stable branches
All branches with prefix `release-` are considered _stable_ branches.
After every minor release (http://semver.org/), we will have a new stable branch for that release. We will keep fixing the backwards-compatible bugs for the latest stable release, but not previous releases. The _patch_ release, incorporating any bug fixes, will be once every two weeks, given any patches.
Starting an etcd cluster statically requires that each member knows another in the cluster. In a number of cases, you might not know the IPs of your cluster members ahead of time. In these cases, you can bootstrap an etcd cluster with the help of a discovery service.
Once an etcd cluster is up and running, adding or removing members is done via [runtime reconfiguration][runtime-conf]. To better understand the design behind runtime reconfiguration, we suggest you read [the runtime configuration design document][runtime-reconf-design].
This guide will cover the following mechanisms for bootstrapping an etcd cluster:
* [Static](#static)
* [etcd Discovery](#etcd-discovery)
* [DNS Discovery](#dns-discovery)
Each of the bootstrapping mechanisms will be used to create a three machine etcd cluster with the following details:
|Name|Address|Hostname|
|------|---------|------------------|
|infra0|10.0.1.10|infra0.example.com|
|infra1|10.0.1.11|infra1.example.com|
|infra2|10.0.1.12|infra2.example.com|
## Static
As we know the cluster members, their addresses and the size of the cluster before starting, we can use an offline bootstrap configuration by setting the `initial-cluster` flag. Each machine will get either the following command line or environment variables:
Note that the URLs specified in `initial-cluster` are the _advertised peer URLs_, i.e. they should match the value of `initial-advertise-peer-urls` on the respective nodes.
If you are spinning up multiple clusters (or creating and destroying a single cluster) with same configuration for testing purpose, it is highly recommended that you specify a unique `initial-cluster-token` for the different clusters. By doing this, etcd can generate unique cluster IDs and member IDs for the clusters even if they otherwise have the exact same configuration. This can protect you from cross-cluster-interaction, which might corrupt your clusters.
etcd listens on [`listen-client-urls`][conf-listen-client] to accept client traffic. etcd member advertises the URLs specified in [`advertise-client-urls`][conf-adv-client] to other members, proxies, clients. Please make sure the `advertise-client-urls` are reachable from intended clients. A common mistake is setting `advertise-client-urls` to localhost or leave it as default when you want the remote clients to reach etcd.
On each machine you would start etcd with these flags:
The command line parameters starting with `--initial-cluster` will be ignored on subsequent runs of etcd. You are free to remove the environment variables or command line flags after the initial bootstrap process. If you need to make changes to the configuration later (for example, adding or removing members to/from the cluster), see the [runtime configuration][runtime-conf] guide.
### Error Cases
In the following example, we have not included our new host in the list of enumerated nodes. If this is a new cluster, the node _must_ be added to the list of initial cluster members.
etcd: infra1 not listed in the initial cluster config
exit 1
```
In this example, we are attempting to map a node (infra0) on a different address (127.0.0.1:2380) than its enumerated address in the cluster list (10.0.1.10:2380). If this node is to listen on multiple addresses, all addresses _must_ be reflected in the "initial-cluster" configuration directive.
etcd: conflicting cluster ID to the target cluster (c6ab534d07e8fcc4 != bc25ea2a74fb18b0). Exiting.
exit 1
```
## Discovery
In a number of cases, you might not know the IPs of your cluster peers ahead of time. This is common when utilizing cloud providers or when your network uses DHCP. In these cases, rather than specifying a static configuration, you can use an existing etcd cluster to bootstrap a new one. We call this process "discovery".
There two methods that can be used for discovery:
* etcd discovery service
* DNS SRV records
### etcd Discovery
To better understand the design about discovery service protocol, we suggest you read [this][discovery-proto].
#### Lifetime of a Discovery URL
A discovery URL identifies a unique etcd cluster. Instead of reusing a discovery URL, you should always create discovery URLs for new clusters.
Moreover, discovery URLs should ONLY be used for the initial bootstrapping of a cluster. To change cluster membership after the cluster is already running, see the [runtime reconfiguration][runtime-conf] guide.
#### Custom etcd Discovery Service
Discovery uses an existing cluster to bootstrap itself. If you are using your own etcd cluster you can create a URL like so:
```
$ curl -X PUT https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83/_config/size -d value=3
```
By setting the size key to the URL, you create a discovery URL with an expected cluster size of 3.
If you bootstrap an etcd cluster using discovery service with more than the expected number of etcd members, the extra etcd processes will [fall back][fall-back] to being [proxies][proxy] by default.
The URL you will use in this case will be `https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83` and the etcd members will use the `https://myetcd.local/v2/keys/discovery/6c007a14875d53d9bf0ef5a6fc0257c817f0fb83` directory for registration as they start.
**Each member must have a different name flag specified. `Hostname` or `machine-id` can be a good choice. Or discovery will fail due to duplicated name.**
Now we start etcd with those relevant flags for each member:
This will cause each member to register itself with the custom etcd discovery service and begin the cluster once all machines have been registered.
#### Public etcd Discovery Service
If you do not have access to an existing cluster, you can use the public discovery service hosted at `discovery.etcd.io`. You can create a private discovery URL using the "new" endpoint like so:
This will create the cluster with an initial expected size of 3 members. If you do not specify a size, a default of 3 will be used.
If you bootstrap an etcd cluster using discovery service with more than the expected number of etcd members, the extra etcd processes will [fall back][fall-back] to being [proxies][proxy] by default.
**Each member must have a different name flag specified. `Hostname` or `machine-id` can be a good choice. Or discovery will fail due to duplicated name.**
Now we start etcd with those relevant flags for each member:
etcd cluster members can listen on domain names or IP address, the bootstrap process will resolve DNS A records.
The resolved address in `--initial-advertise-peer-urls`*must match* one of the resolved addresses in the SRV targets. The etcd member reads the resolved address to find out if it belongs to the cluster defined in the SRV records.
DNS SRV records can also be used to configure the list of peers for an etcd server running in proxy mode:
```
$ etcd --proxy on --discovery-srv example.com
```
#### etcd client configuration
DNS SRV records can also be used to help clients discover the etcd cluster.
The official [etcd/client][client] supports [DNS Discovery][client-discoverer].
`etcdctl` also supports DNS Discovery by specifying the `--discovery-srv` option.
```
$ etcdctl --discovery-srv example.com set foo bar
```
#### Error Cases
You might see an error like `cannot find local etcd $name from SRV records.`. That means the etcd member fails to find itself from the cluster defined in SRV records. The resolved address in `--initial-advertise-peer-urls`*must match* one of the resolved addresses in the SRV targets.
# 0.4 to 2.0+ Migration Guide
In etcd 2.0 we introduced the ability to listen on more than one address and to advertise multiple addresses. This makes using etcd easier when you have complex networking, such as private and public networks on various cloud providers.
To make understanding this feature easier, we changed the naming of some flags, but we support the old flags to make the migration from the old to new version easier.
|-peer-addr |--initial-advertise-peer-urls |If specified, peer-addr will be used as the only peer URL. Error if both flags specified.|
|-addr |--advertise-client-urls |If specified, addr will be used as the only client URL. Error if both flags specified.|
|-peer-bind-addr |--listen-peer-urls |If specified, peer-bind-addr will be used as the only peer bind URL. Error if both flags specified.|
|-bind-addr |--listen-client-urls |If specified, bind-addr will be used as the only client bind URL. Error if both flags specified.|
|-peers |none |Deprecated. The --initial-cluster flag provides a similar concept with different semantics. Please read this guide on cluster startup.|
|-peers-file |none |Deprecated. The --initial-cluster flag provides a similar concept with different semantics. Please read this guide on cluster startup.|
etcd is configurable through command-line flags and environment variables. Options set on the command line take precedence over those from the environment.
The format of environment variable for flag `--my-flag` is `ETCD_MY_FLAG`. It applies to all flags.
The [official etcd ports][iana-ports] are 2379 for client requests, and 2380 for peer communication. Some legacy code and documentation still references ports 4001 and 7001, but all new etcd use and discussion should adopt the assigned ports.
To start etcd automatically using custom settings at startup in Linux, using a [systemd][systemd-intro] unit is highly recommended.
+ This value is referenced as this node's own entries listed in the `--initial-cluster` flag (Ex: `default=http://localhost:2380` or `default=http://localhost:2380,default=http://localhost:7001`). This needs to match the key used in the flag if you're using [static bootstrapping][build-cluster]. When using discovery, each member must have a unique name. `Hostname` or `machine-id` can be a good choice.
### --data-dir
+ Path to the data directory.
+ default: "${name}.etcd"
+ env variable: ETCD_DATA_DIR
### --wal-dir
+ Path to the dedicated wal directory. If this flag is set, etcd will write the WAL files to the walDir rather than the dataDir. This allows a dedicated disk to be used, and helps avoid io competition between logging and other IO operations.
+ default: ""
+ env variable: ETCD_WAL_DIR
### --snapshot-count
+ Number of committed transactions to trigger a snapshot to disk.
+ default: "10000"
+ env variable: ETCD_SNAPSHOT_COUNT
### --heartbeat-interval
+ Time (in milliseconds) of a heartbeat interval.
+ default: "100"
+ env variable: ETCD_HEARTBEAT_INTERVAL
### --election-timeout
+ Time (in milliseconds) for an election to timeout. See [tuning.md](tuning.md#time-parameters) for details.
+ default: "1000"
+ env variable: ETCD_ELECTION_TIMEOUT
### --listen-peer-urls
+ List of URLs to listen on for peer traffic. This flag tells the etcd to accept incoming requests from its peers on the specified scheme://IP:port combinations. Scheme can be either http or https.If 0.0.0.0 is specified as the IP, etcd listens to the given port on all interfaces. If an IP address is given as well as a port, etcd will listen on the given port and interface. Multiple URLs may be used to specify a number of addresses and ports to listen on. The etcd will respond to requests from any of the listed addresses and ports.
+ invalid example: "http://example.com:2380" (domain name is invalid for binding)
### --listen-client-urls
+ List of URLs to listen on for client traffic. This flag tells the etcd to accept incoming requests from the clients on the specified scheme://IP:port combinations. Scheme can be either http or https. If 0.0.0.0 is specified as the IP, etcd listens to the given port on all interfaces. If an IP address is given as well as a port, etcd will listen on the given port and interface. Multiple URLs may be used to specify a number of addresses and ports to listen on. The etcd will respond to requests from any of the listed addresses and ports.
+ invalid example: "http://example.com:2379" (domain name is invalid for binding)
### --max-snapshots
+ Maximum number of snapshot files to retain (0 is unlimited)
+ default: 5
+ env variable: ETCD_MAX_SNAPSHOTS
+ The default for users on Windows is unlimited, and manual purging down to 5 (or your preference for safety) is recommended.
### --max-wals
+ Maximum number of wal files to retain (0 is unlimited)
+ default: 5
+ env variable: ETCD_MAX_WALS
+ The default for users on Windows is unlimited, and manual purging down to 5 (or your preference for safety) is recommended.
### --cors
+ Comma-separated white list of origins for CORS (cross-origin resource sharing).
+ default: none
+ env variable: ETCD_CORS
## Clustering Flags
`--initial` prefix flags are used in bootstrapping ([static bootstrap][build-cluster], [discovery-service bootstrap][discovery] or [runtime reconfiguration][reconfig]) a new member, and ignored when restarting an existing member.
`--discovery` prefix flags need to be set when using [discovery service][discovery].
### --initial-advertise-peer-urls
+ List of this member's peer URLs to advertise to the rest of the cluster. These addresses are used for communicating etcd data around the cluster. At least one must be routable to all cluster members. These URLs can contain domain names.
+ The key is the value of the `--name` flag for each node provided. The default uses `default` for the key because this is the default for the `--name` flag.
### --initial-cluster-state
+ Initial cluster state ("new" or "existing"). Set to `new` for all members present during initial static or DNS bootstrapping. If this option is set to `existing`, etcd will attempt to join the existing cluster. If the wrong value is set, etcd will attempt to start but fail safely.
+ default: "new"
+ env variable: ETCD_INITIAL_CLUSTER_STATE
[static bootstrap]: clustering.md#static
### --initial-cluster-token
+ Initial cluster token for the etcd cluster during bootstrap.
+ default: "etcd-cluster"
+ env variable: ETCD_INITIAL_CLUSTER_TOKEN
### --advertise-client-urls
+ List of this member's client URLs to advertise to the rest of the cluster. These URLs can contain domain names.
+ Be careful if you are advertising URLs such as http://localhost:2379 from a cluster member and are using the proxy feature of etcd. This will cause loops, because the proxy will be forwarding requests to itself until its resources (memory, file descriptors) are eventually depleted.
### --discovery
+ Discovery URL used to bootstrap the cluster.
+ default: none
+ env variable: ETCD_DISCOVERY
### --discovery-srv
+ DNS srv domain used to bootstrap the cluster.
+ default: none
+ env variable: ETCD_DISCOVERY_SRV
### --discovery-fallback
+ Expected behavior ("exit" or "proxy") when discovery services fails.
+ default: "proxy"
+ env variable: ETCD_DISCOVERY_FALLBACK
### --discovery-proxy
+ HTTP proxy to use for traffic to discovery service.
+ default: none
+ env variable: ETCD_DISCOVERY_PROXY
### --strict-reconfig-check
+ Reject reconfiguration requests that would cause quorum loss.
+ default: false
+ env variable: ETCD_STRICT_RECONFIG_CHECK
## Proxy Flags
`--proxy` prefix flags configures etcd to run in [proxy mode][proxy].
### --proxy
+ Proxy mode setting ("off", "readonly" or "on").
+ default: "off"
+ env variable: ETCD_PROXY
### --proxy-failure-wait
+ Time (in milliseconds) an endpoint will be held in a failed state before being reconsidered for proxied requests.
+ default: 5000
+ env variable: ETCD_PROXY_FAILURE_WAIT
### --proxy-refresh-interval
+ Time (in milliseconds) of the endpoints refresh interval.
+ default: 30000
+ env variable: ETCD_PROXY_REFRESH_INTERVAL
### --proxy-dial-timeout
+ Time (in milliseconds) for a dial to timeout or 0 to disable the timeout
+ default: 1000
+ env variable: ETCD_PROXY_DIAL_TIMEOUT
### --proxy-write-timeout
+ Time (in milliseconds) for a write to timeout or 0 to disable the timeout.
+ default: 5000
+ env variable: ETCD_PROXY_WRITE_TIMEOUT
### --proxy-read-timeout
+ Time (in milliseconds) for a read to timeout or 0 to disable the timeout.
+ Don't change this value if you use watches because they are using long polling requests.
+ default: 0
+ env variable: ETCD_PROXY_READ_TIMEOUT
## Security Flags
The security flags help to [build a secure etcd cluster][security].
### --ca-file [DEPRECATED]
+ Path to the client server TLS CA file. `--ca-file ca.crt` could be replaced by `--trusted-ca-file ca.crt --client-cert-auth` and etcd will perform the same.
+ default: none
+ env variable: ETCD_CA_FILE
### --cert-file
+ Path to the client server TLS cert file.
+ default: none
+ env variable: ETCD_CERT_FILE
### --key-file
+ Path to the client server TLS key file.
+ default: none
+ env variable: ETCD_KEY_FILE
### --client-cert-auth
+ Enable client cert authentication.
+ default: false
+ env variable: ETCD_CLIENT_CERT_AUTH
### --trusted-ca-file
+ Path to the client server TLS trusted CA key file.
+ default: none
+ env variable: ETCD_TRUSTED_CA_FILE
### --peer-ca-file [DEPRECATED]
+ Path to the peer server TLS CA file. `--peer-ca-file ca.crt` could be replaced by `--peer-trusted-ca-file ca.crt --peer-client-cert-auth` and etcd will perform the same.
+ default: none
+ env variable: ETCD_PEER_CA_FILE
### --peer-cert-file
+ Path to the peer server TLS cert file.
+ default: none
+ env variable: ETCD_PEER_CERT_FILE
### --peer-key-file
+ Path to the peer server TLS key file.
+ default: none
+ env variable: ETCD_PEER_KEY_FILE
### --peer-client-cert-auth
+ Enable peer client cert authentication.
+ default: false
+ env variable: ETCD_PEER_CLIENT_CERT_AUTH
### --peer-trusted-ca-file
+ Path to the peer server TLS trusted CA file.
+ default: none
+ env variable: ETCD_PEER_TRUSTED_CA_FILE
## Logging Flags
### --debug
+ Drop the default log level to DEBUG for all subpackages.
+ default: false (INFO for all packages)
+ env variable: ETCD_DEBUG
### --log-package-levels
+ Set individual etcd subpackages to specific log levels. An example being `etcdserver=WARNING,security=DEBUG`
+ default: none (INFO for all packages)
+ env variable: ETCD_LOG_PACKAGE_LEVELS
## Unsafe Flags
Please be CAUTIOUS when using unsafe flags because it will break the guarantees given by the consensus protocol.
For example, it may panic if other members in the cluster are still alive.
Follow the instructions when using these flags.
### --force-new-cluster
+ Force to create a new one-member cluster. It commits configuration changes forcing to remove all existing members in the cluster and add itself. It needs to be set to [restore a backup][restore].
+ default: false
+ env variable: ETCD_FORCE_NEW_CLUSTER
## Experimental Flags
### --experimental-v3demo
+ Enable experimental [v3 demo API][rfc-v3].
+ default: false
+ env variable: ETCD_EXPERIMENTAL_V3DEMO
## Miscellaneous Flags
### --version
+ Print the version and exit.
+ default: false
## Profiling flags
### --enable-pprof
+ Enable runtime profiling data via HTTP server. Address is at client URL + "/debug/pprof"
The guide talks about how to release a new version of etcd.
The procedure includes some manual steps for sanity checking but it can probably be further scripted. Please keep this document up-to-date if you want to make changes to the release process.
## Prepare Release
Set desired version as environment variable for following steps. Here is an example to release 2.1.3:
```
export VERSION=v2.1.3
export PREV_VERSION=v2.1.2
```
All releases version numbers follow the format of [semantic versioning 2.0.0](http://semver.org/).
### Major, Minor Version Release, or its Pre-release
- Ensure the relevant milestone on GitHub is complete. All referenced issues should be closed, or moved elsewhere.
- Remove this release from [roadmap](https://github.com/coreos/etcd/blob/master/ROADMAP.md), if necessary.
- Ensure the latest upgrade documentation is available.
- Bump [hardcoded MinClusterVerion in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L29), if necessary.
- Add feature capability maps for the new version, if necessary.
### Patch Version Release
- Discuss about commits that are backported to the patch release. The commits should not include merge commits.
- Cherry-pick these commits starting from the oldest one into stable branch.
## Write Release Note
- Write introduction for the new release. For example, what major bug we fix, what new features we introduce or what performance improvement we make.
- Write changelog for the last release. ChangeLog should be straightforward and easy to understand for the end-user.
- Put `[GH XXXX]` at the head of change line to reference Pull Request that introduces the change. Moreover, add a link on it to jump to the Pull Request.
## Tag Version
- Bump [hardcoded Version in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L30) to the latest version `${VERSION}`.
- Ensure all tests on CI system are passed.
- Manually check etcd is buildable in Linux, Darwin and Windows.
- Manually check upgrade etcd cluster of previous minor version works well.
- Manually check new features work well.
- Add a signed tag through `git tag -s ${VERSION}`.
- Sanity check tag correctness through `git show tags/$VERSION`.
- Push the tag to GitHub through `git push origin tags/$VERSION`. This assumes `origin` corresponds to "https://github.com/coreos/etcd".
## Build Release Binaries and Images
- Ensure `actool` is available, or installing it through `go get github.com/appc/spec/actool`.
- Ensure `docker` is available.
Run release script in root directory:
```
./scripts/release.sh ${VERSION}
```
It generates all release binaries and images under directory ./release.
## Sign Binaries and Images
Choose appropriate private key to sign the generated binaries and images.
The following commands are used for public release sign:
```
cd release
# personal GPG is okay for now
for i in etcd-*{.zip,.tar.gz}; do gpg --sign ${i}; done
# use `CoreOS ACI Builder <release@coreos.com>` secret key
gpg -u 88182190 -a --output etcd-${VERSION}-linux-amd64.aci.asc --detach-sig etcd-${VERSION}-linux-amd64.aci
```
## Publish Release Page in GitHub
- Set release title as the version name.
- Follow the format of previous release pages.
- Attach the generated binaries, aci image and signatures.
- Select whether it is a pre-release.
- Publish the release!
## Publish Docker Image in Quay.io
- Push docker image:
```
docker login quay.io
docker push quay.io/coreos/etcd:${VERSION}
```
- Add `latest` tag to the new image on [quay.io](https://quay.io/repository/coreos/etcd?tag=latest&tab=tags) if this is a stable release.
## Announce to etcd-dev Googlegroup
- Follow the format of [previous release emails](https://groups.google.com/forum/#!forum/etcd-dev).
- Make sure to include a list of authors that contributed since the previous release - something like the following might be handy:
- Create new stable branch through `git push origin ${VERSION_MAJOR}.${VERSION_MINOR}` if this is a major stable release. This assumes `origin` corresponds to "https://github.com/coreos/etcd".
- Bump [hardcoded Version in the repository](https://github.com/coreos/etcd/blob/master/version/version.go#L30) to the version `${VERSION}+git`.
Discovery service protocol helps new etcd member to discover all other members in cluster bootstrap phase using a shared discovery URL.
Discovery service protocol is _only_ used in cluster bootstrap phase, and cannot be used for runtime reconfiguration or cluster monitoring.
The protocol uses a new discovery token to bootstrap one _unique_ etcd cluster. Remember that one discovery token can represent only one etcd cluster. As long as discovery protocol on this token starts, even if it fails halfway, it must not be used to bootstrap another etcd cluster.
The rest of this article will walk through the discovery process with examples that correspond to a self-hosted discovery cluster. The public discovery service, discovery.etcd.io, functions the same way, but with a layer of polish to abstract away ugly URLs, generate UUIDs automatically, and provide some protections against excessive requests. At its core, the public discovery service still uses an etcd cluster as the data store as described in this document.
## The Protocol Workflow
The idea of discovery protocol is to use an internal etcd cluster to coordinate bootstrap of a new cluster. First, all new members interact with discovery service and help to generate the expected member list. Then each new member bootstraps its server using this list, which performs the same functionality as -initial-cluster flag.
In the following example workflow, we will list each step of protocol in curl format for ease of understanding.
By convention the etcd discovery protocol uses the key prefix `_etcd/registry`. If `http://example.com` hosts an etcd cluster for discovery service, a full URL to discovery keyspace will be `http://example.com/v2/keys/_etcd/registry`. We will use this as the URL prefix in the example.
### Creating a New Discovery Token
Generate a unique token that will identify the new cluster. This will be used as a unique prefix in discovery keyspace in the following steps. An easy way to do this is to use `uuidgen`:
```
UUID=$(uuidgen)
```
### Specifying the Expected Cluster Size
You need to specify the expected cluster size for this discovery token. The size is used by the discovery service to know when it has found all members that will initially form the cluster.
```
curl -X PUT http://example.com/v2/keys/_etcd/registry/${UUID}/_config/size -d value=${cluster_size}
```
Usually the cluster size is 3, 5 or 7. Check [optimal cluster size][cluster-size] for more details.
### Bringing up etcd Processes
Now that you have your discovery URL, you can use it as `-discovery` flag and bring up etcd processes. Every etcd process will follow this next few steps internally if given a `-discovery` flag.
### Registering itself
The first thing for etcd process is to register itself into the discovery URL as a member. This is done by creating member ID as a key in the discovery URL.
```
curl -X PUT http://example.com/v2/keys/_etcd/registry/${UUID}/${member_id}?prevExist=false -d value="${member_name}=${member_peer_url_1}&${member_name}=${member_peer_url_2}"
```
### Checking the Status
It checks the expected cluster size and registration status in discovery URL, and decides what the next action is.
```
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}/_config/size
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}
```
If registered members are still not enough, it will wait for left members to appear.
If the number of registered members is bigger than the expected size N, it treats the first N registered members as the member list for the cluster. If the member itself is in the member list, the discovery procedure succeeds and it fetches all peers through the member list. If it is not in the member list, the discovery procedure finishes with the failure that the cluster has been full.
In etcd implementation, the member may check the cluster status even before registering itself. So it could fail quickly if the cluster has been full.
### Waiting for All Members
The wait process is described in detail in the [etcd API documentation][api].
```
curl -X GET http://example.com/v2/keys/_etcd/registry/${UUID}?wait=true&waitIndex=${current_etcd_index}
```
It keeps waiting until finding all members.
## Public Discovery Service
CoreOS Inc. hosts a public discovery service at https://discovery.etcd.io/ , which provides some nice features for ease of use.
### Mask Key Prefix
Public discovery service will redirect `https://discovery.etcd.io/${UUID}` to etcd cluster behind for the key at `/v2/keys/_etcd/registry`. It masks register key prefix for short and readable discovery url.
### Get new token
```
GET /new
Sent query:
size=${cluster_size}
Possible status codes:
200 OK
400 Bad Request
200 Body:
generated discovery url
```
The generation process in the service follows the steps from [Creating a New Discovery Token][new-discovery-token] to [Specifying the Expected Cluster Size][expected-cluster-size].
### Check Discovery Status
```
GET /${UUID}
```
You can check the status for this discovery token, including the machines that have been registered, by requesting the value of the UUID.
### Open-source repository
The repository is located at https://github.com/coreos/discovery.etcd.io. You could use it to build your own public discovery service.
## 1) Why can an etcd client read an old version of data when a majority of the etcd cluster members are down?
In situations where a client connects to a minority, etcd
favors by default availability over consistency. This means that even though
data might be “out of date”, it is still better to return something versus
nothing.
In order to confirm that a read is up to date with a majority of the cluster,
the client can use the `quorum=true` parameter on reads of keys. This means
that a majority of the cluster is checked on reads before returning the data,
otherwise the read will timeout and fail.
## 2) With quorum=false, doesn’t this mean that if my client switched the member it was connected to, that it could experience a logical ordering where the cluster goes backwards in time?
Yes, but this could be handled at the etcd client implementation via
remembering the last seen index. The “index” is the cluster's single
irrevocable sequence of the entire modification history. The client could
remember the last seen index, and determine via comparing the index returned on
the GET whether or not the state of the key-value pair is before or after its
last seen state.
## 3) What happens if a watch is registered on a minority member?
The watch will stay untriggered, even as modifications are occurring in the
majority quorum. This is an open issue, and is being addressed in v3. There are
multiple ways to work around the watch trigger not firing.
1) build a signaling mechanism independent of etcd. This could be as simple as
a “pulse” to the client to reissue a GET with quorum=true for the most recent
version of the data.
2) poll on the `/v2/keys` endpoint and check that the raft-index is increasing every
timeout.
## 4) What is a proxy used for?
A proxy is a redirection server to the etcd cluster. The proxy handles the
redirection of a client to the current configuration of the etcd cluster. A
typical use case is to start a proxy on a machine, and on first boot up of the
proxy specify both the `--proxy` flag and the `--initial-cluster` flag.
From there, any etcdctl client that starts up automatically speaks to the local
proxy and the proxy redirects operations to the current configuration of the
cluster it was originally paired with.
In the v2 spec of etcd, proxies cannot be promoted to members of the cluster.
They also cannot be promoted to followers or at any point become part of the
replication of the etcd cluster itself.
## 5) How is cluster membership and health handled in etcd v2?
The design goal of etcd is that reconfiguration is simply an API, and health
monitoring and addition/removal of members is up to the individual application
and their integration with the reconfiguration API.
Thus, a member that is down, even infinitely, will never be automatically
removed from the etcd cluster member list.
This makes sense because it's usually an application level / administrative
action to determine whether a reconfiguration should happen based on health.
For more information, refer to the [runtime reconfiguration design document][runtime-reconf-design].
## 6) how does --endpoint work with etcdctl?
The `--endpoint` flag can specify any number of etcd cluster members in a comma
separated list. This list might be a subset, equal to, or more than the actual
etcd cluster member list itself.
If only one peer is specified via the `--endpoint` flag, the etcdctl discovers the
rest of the cluster via the member list of that one peer, and then it randomly
chooses a member to use. Again, the client can use the `quorum=true` flag on
reads, which will always fail when using a member in the minority.
If peers from multiple clusters are specified via the `--endpoint` flag, etcdctl
will randomly choose a peer, and the request will simply get routed to one of
the clusters. This is probably not what you want.
Note: --peers flag is now deprecated and --endpoint should be used instead,
as it might confuse users to give etcdctl a peerURL.
This document defines the various terms used in etcd documentation, command line and source code.
## Node
Node is an instance of raft state machine.
It has a unique identification, and records other nodes' progress internally when it is the leader.
## Member
Member is an instance of etcd. It hosts a node, and provides service to clients.
## Cluster
Cluster consists of several members.
The node in each member follows raft consensus protocol to replicate logs. Cluster receives proposals from members, commits them and apply to local store.
## Peer
Peer is another member of the same cluster.
## Proposal
A proposal is a request (for example a write request, a configuration change request) that needs to go through raft protocol.
## Client
Client is a caller of the cluster's HTTP API.
## Machine (deprecated)
The alternative of Member in etcd before 2.0
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
