Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
The scheduler has a bug where acker memory is not considered in the scheduling estimate. The case I found was where a topology should fit on two supervisors, but the cluster has 1 available and 2 blacklisted. The scheduler thinks the topology should fit on one supervisor and fails to schedule, but also fails to release a supervisor from the blacklist, resulting in the topology never getting scheduled.
With this fix, the scheduler properly detects the topology will need to be scheduled on two supervisors and releases one from the blacklist and schedules successfully.
Switched some scheduling logs from trace to debug to make debugging scheduling issues easier.