It works on XP.

I can't build it on Win7 yet (have to wait the build fix)

See the comment on sporadic nature of the failure. It works for me too, but it fails sometimes on buildbot

The problem could be cause by really high and uneven load on the testing box. The test makes assumption that load on the box is more or less even and error caused of high load can't differ from calculated more then on 2 sec.

Other cause of the problem could be different implementation of sleep() in windows where parallel execution of sleep can lead to earlier ending of one of them or delaying other one.

Could you say something about sleep() implementation in windows? (see comment)

Sleep implementation on Windows (Item_func_sleep) is the same as anywhere else - it is using wait on a condition variable. The timers Windows uses internally are not high-performance timers, i.e their accuracy is around 10 ms. Parallel waiting on condition variables is of course possible

The test is not related to QC (sleep is not cached). It is somehow windows specific. My knowladge about windows limited and so I run out of ideas how to repeat it or why it happened on windows.

No, it is not only windows specific. The link shows Windows failure, because it happened on Windows. Yet, it is test specific, as I have seen it on freebsd, too. Now, we do not have freebsd running on builbot, but it is easily obtainable and installable on virtual machines

Also, I personally found "sleep" parts of tests in query_cache.test to be tricky to the extend they get incomprehensible and impenetrable. Maybe it is better to just remove them and replace with 1 simple test proving that sleep() is not cached?

did not fail for quite a long time. Will reopen if I see it failing..