tree 0ecf2eb2a7c4908e43911573394bcf58fff37416
parent a32513a1fc6a149b30f04721f866e3cbb6003661
author Charles-Francois Natali <cf.natali@gmail.com> 1588380069 +0100
committer Andrei Budnik <abudnik@apache.org> 1588850732 +0200

Keep retrying to remove cgroup on EBUSY.

This is a follow-up to MESOS-10107, which introduced retries when
calling `rmdir` on a seemingly empty cgroup fails with `EBUSY`
because of various kernel bugs.
At the time, the fix introduced a bounded number of retries, using an
exponential backoff summing up to slightly over 1s. This was done
because it was similar to what Docker does, and worked during testing.
However, after 1 month without seeing this error in our cluster at work,
we finally experienced one case where the 1s timeout wasn't enough.
It could be because the machine was busy at the time, or some other
random factor.
So instead of only trying for 1s, I think it might make sense to just
keep retrying, until the top-level container destruction timeout - set
at 1 minute - kicks in.
This actually makes more sense, and avoids having a magical timeout in
the cgroup code.
We just need to ensure that when the destroyer is finalized, it discards
the future in charge of doing the periodic remove.

This closes #362