Details
-
Improvement
-
Status: Resolved
-
Normal
-
Resolution: Fixed
-
None
Description
Some environments have ridiculous outliers for disk syncing. 20 seconds ridiculous.
Unit tests aren't testing crash safety so it is a pointless exercise.
Instead we could intercept calls to sync files and check whether it looks like the sync would succeed. Check that the things are not null, mapped, closed etc. Outside of units tests it can go straight to the regular sync call.
I would also like to have the disks for unit and dtests mounted with barrier=0,noatime,nodiratime to further reduce susceptibility to outliers. We aren't going to recover these nodes if they crash/restart.