Coordinated omission has nothing to do with Java in particular, but its a more general problem that comes up when measuring latency (response time) of a system over a period of time. Background In most latency sensitive applications, eg a trading system, given a stream of requests over a given period of time, usually requirements for latency are described as follows "99% of trades should be processed within 0.5 secs , 99.99% of trade should be processed in under a second and no trade should take longer than 2 seconds to be processed." Average response times does not provide any useful information, because you it is possible to have a low average response time but have worst case or 1% of the requests could be orders of magnitude higher. The best way to measure latency is by percentile distribution. How to measure The common way to do this is have a tool to send a send a requests continually and record and log the response time for each request, and build a histogram out of the recorded response time.
Problem (Coordinated Ommission) The above approach works well but makes an assumption, it does not take into account the stalling of the system under test, (all systems will stall at some point). Suppose our test tool sends 10 requests/sec and our system process each request within 1 millisecond, but after sometime say our system stalls for 10 seconds (for sake of simplicity say due to garbage collection), during this '10 seconds stall' time our test tool has only send one request, as it is also stalled waiting for the response. This long operation has only been recorded once. But in actuality during this 10 sec stall the system would have delayed 10 req/sec x 10 sec = 100 requests, but this was not recorded by the testing tool as the testing tool was waiting for the first request.
