Details
-
Sub-task
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
This is a problem after YARN-2008, queue's current-max-limit (it was called maxAvailCapacity in YARN-2008) will be updated when queue is traversed by scheduler, so application can get headroom based on staled current-max-limit. For example:
r / \ a b / \ a1 b1 ^ ^ | | app1 app2
We have above queue structure, two apps in the cluster: app1 in a1, and app2 in b1. If scheduler always travers b1 only, and continously allocate containers in b1, app1's headroom will not get updated.