[MESOS-8826] ReaperProcess depends on clock manipulation.

• Type: Task
• Status: Accepted
• Priority: Major
• Resolution: Unresolved
• Component/s: libprocess
• Labels: integration mesosphere
• Epic Link: test-harness
• Assignee: Unassigned
• Reporter: Andrei Budnik
• Created: 24/Apr/18 12:23
• Updated: 29/Apr/19 09:26

Description

After making composing containerizer as a default containerizer in tests by setting a value `mesos,docker` for `containerizers` flag in `MesosTest::CreateSlaveFlags()`, some tests started to endlessly wait for recovery completion after starting slave with paused clocks.
Recovery process for a docker c'zer uses `subprocess()` to spawn `docker ps -a` subprocess and subscribes for its termination. As a reaper process uses `delay()`, this leads to a hanging recovery process for the docker c'zer, when the libprocess' clocks are paused.

A possible solution might be using a single dedicated thread that doesn't use libprocess clocks and periodically iterates over pids and calls `waitpid` for each of them.
In addition, we can introduce one more dedicated thread that calls `waitpid(-1)` to wait for termination of children processes, so that we can get rid of

while (someEvent.isPending()) {
  Clock::advance(process::MAX_REAP_INTERVAL());
  Clock::settle();
}

pattern in tests.

Issue Links

• relates to: MESOS-9139 Use composing containerizer by default in tests.
• relates to: MESOS-8732 Use composing containerizer in some agent tests.