Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Duplicate
-
0.9.0, 1.0.0
-
None
-
None
Description
The yarn alpha version of calculateAMMemory takes into account the minimum resource capability and also subtracts out YarnAllocationHandler.MEMORY_OVERHEAD.
The yarn stable version just sets the -Xmx to whatever is passed in by the user.
The 2 of these should be the same.
Personally I also think its weird how spark currently takes whatever user passes in for memory and adds YarnAllocationHandler.MEMORY_OVERHEAD to request to RM. This can be confusing to users. We should revisit all of this and commonize the stable/alpha code where possible.
Attachments
Issue Links
- duplicates
-
SPARK-2140 yarn stable client doesn't properly handle MEMORY_OVERHEAD for AM
- Resolved