[MESOS-762] Revert the use of the soft limit and memory threshold notifications.

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.14.1
• Fix Version/s: 0.14.2, 0.15.0
• Component/s: containerization
• Assignee: Benjamin Mahler
• Reporter: Benjamin Mahler
• Created: 23/Oct/13 01:41
• Updated: 26/Apr/17 16:53
• Resolved: 23/Oct/13 19:25

Description

The use of the cgroup soft limit and threshold notifications to induce our own OOMs does not behave the same as relying on the hard limit and oom notifications.

We previously used the cgroup hard limit which ensures a reclaim of cache memory is performed when the hard limit is reached. The latest version 0.14.1 switched to using the cgroup memory soft limit, which does not perform a reclaim when reached. The result is that executors that perform substantial I/O will build up a large amount of cache memory, eventually reaching the soft limit and killed as a result.

The fix is that we're changing back to using the hard limit, to get back to the previous OOM behavior in Mesos.