Details
-
Sub-task
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
3.1.0
-
None
Description
When an executor is decommissioned, we would like to know if its shuffle data is truly going to be lost. In the case of external shuffle service, this means knowing that the worker (or the node that the executor is on) is also going to be lost.
( I don't think we need to worry about disaggregated remote shuffle storage at present since those are only used in a couple of web companies – but when there is remote shuffle then yes the shuffle won't be lost with a decommissioned executor )
We know for sure that a worker is being decommissioned when the Master is asked to decommission a worker. In case of other schedulers:
- Yarn support for decommissioning isn't implemented yet. But the idea would be for Yarn preeemption to not mark that the worker is being lost, but machine level decommissioning (like for kernel upgrades) to do mark such.
- K8s isn't quite working with external shuffle service as yet, so when the executor is lost, the worker isn't quite lost with it.
Attachments
Issue Links
- blocks
-
SPARK-32199 Clear shuffle state when decommissioned nodes/executors are finally lost
- Resolved
- links to