[
https://issues.apache.org/jira/browse/MESOS-7506?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16225515#comment-16225515
]
Andrei Budnik commented on MESOS-7506:
--------------------------------------
Some tests (from {{SlaveTest}} and {{SlaveRecoveryTest}}) have a pattern [like
this|https://github.com/apache/mesos/blob/ff01d0c44251e2ffaa2f4f47b33c790594d194d9/src/tests/slave_tests.cpp#L393-L406],
where the clock is advanced by {{executor_registration_timeout}} and then it
waits in a loop until a task status update is sent. This loop is executing
while the container is being destroyed. At the same time, container destruction
consists of multiple steps, one of them waits for [cgroups
destruction|https://github.com/apache/mesos/blob/ff01d0c44251e2ffaa2f4f47b33c790594d194d9/src/slave/containerizer/mesos/linux_launcher.cpp#L567].
That means, we have a race between container destruction process and the loop
that advances the clock, leading to the following outcomes:
# Container completely destroyed, before clock advancing reaches timeout (e.g.
{{cgroups::DESTROY_TIMEOUT}}).
# Triggered timeout due to clock advancing, before container destruction
completes. That results in [leaving
orphaned|https://github.com/apache/mesos/blob/ff01d0c44251e2ffaa2f4f47b33c790594d194d9/src/slave/containerizer/mesos/containerizer.cpp#L2367-L2380]
containers that will be detected by [Slave
destructor|https://github.com/apache/mesos/blob/ff01d0c44251e2ffaa2f4f47b33c790594d194d9/src/tests/cluster.cpp#L559-L584]
in `tests/cluster.cpp`, so the test will fail.
The issue is easily reproduced by advancing the clocks by 60 seconds or more in
the loop, which waits for a status update.
> Multiple tests leave orphan containers.
> ---------------------------------------
>
> Key: MESOS-7506
> URL: https://issues.apache.org/jira/browse/MESOS-7506
> Project: Mesos
> Issue Type: Bug
> Components: containerization
> Environment: Ubuntu 16.04
> Fedora 23
> other Linux distros
> Reporter: Alexander Rukletsov
> Assignee: Andrei Budnik
> Labels: containerizer, flaky-test, mesosphere
>
> I've observed a number of flaky tests that leave orphan containers upon
> cleanup. A typical log looks like this:
> {noformat}
> ../../src/tests/cluster.cpp:580: Failure
> Value of: containers->empty()
> Actual: false
> Expected: true
> Failed to destroy containers: { da3e8aa8-98e7-4e72-a8fd-5d0bae960014 }
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
