Details
-
Wish
-
Status: Resolved
-
Minor
-
Resolution: Incomplete
-
2.2.3, 2.3.3, 2.4.0
-
None
Description
Spark in standalone mode currently only supports FIFO (first-in-first-out) scheduler across applications.
It will be great that a fair scheduler is supported. A fair scheduler across applications, not in a application.
Use case (for example with the integration of zeppelin)
At certain moment, user A submits a heavy application and consumes all the resources of the spark cluster.
At a later moment, user B submits a second application.
No matter how many work nodes you added now, all the resources go to user A due to the FIFO. User B will never get any resource until user A release its allocated resources.
A fair scheduler should distribute extra resources in a fair way on all running applications, which demands resources.