Details
-
Bug
-
Status: Open
-
Minor
-
Resolution: Unresolved
-
3.0.0
-
None
-
None
Description
A class extending ResourceAllocator (WorkerResourceInfo), has some potential issues (raised here: https://github.com/apache/spark/pull/26078#discussion_r340342820)
The WorkerInfo class is calling into availableAddrs and assignedAddrs but those calls appear to be coming from the UI (looking at the resourcesInfo* functions), e.g. JsonProtocol and MasterPage call this. Since the datastructures in ResourceAllocator are not concurrent, we could end up with bad data or potentially crashes, depending on when the calls are made.
Note that there are other calls to the resourceInfo* functions, but those are from the event loop.
Attachments
Issue Links
- relates to
-
SPARK-24615 SPIP: Accelerator-aware task scheduling for Spark
- Resolved