[MESOS-9282] StatusUpdateManager does not sync checkpointed data to disk.

• Type: Bug
• Status: Accepted
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.5.1, 1.6.1, 1.7.0
• Component/s: agent
• Labels: foundations
• Assignee: Unassigned
• Reporter: Jie Yu
• Created: 01/Oct/18 23:22
• Updated: 05/Sep/19 08:52

Description

This is related to MESOS-9281, which we observed in a testing environment.

The status update manager used to open the checkpoint file using O_SYNC, which will guarantee that each write will be persisted to the disk (similar to calling fsync() after each write()).

This was removed due to some performance issue
https://reviews.apache.org/r/50635/

However, the assumption in the patch is no longer true after we allow the re-use the same agent ID after machine reboot. This will likely cause issues.

Issue Links

• relates to: MESOS-9281 SLRP gets a stale checkpoint after system crash.
• relates to: MESOS-5944 Remove `O_SYNC` from StatusUpdateManager logs