It makes more sense to record the latency of successes (or all attempts) than of only a particular failure case.