Victor Sunderland created SPARK-60480:
-----------------------------------------
Summary: Avoid scanning healthy executor pods in
ExecutorPodsLifecycleManager
Key: SPARK-60480
URL: https://issues.apache.org/jira/browse/SPARK-60480
Project: Spark
Issue Type: Improvement
Components: Kubernetes
Affects Versions: 5.0.0
Reporter: Victor Sunderland
{{ExecutorPodsLifecycleManager.onNewSnapshots()}} scans the full executor pod
map of every snapshot twice: once to collect failed executors and once to
handle final and inactive pods. Healthy (pending or running) pods never need
lifecycle handling, so with many executors and frequent watch updates most of
this work is wasted.
Each {{ExecutorPodsSnapshot}} can keep an index of its final and inactive pods,
maintained incrementally in {{withUpdate()}}, so the manager only visits the
pods it acts on. Every snapshot is still processed, so failures seen in
intermediate snapshots and retries of failed pod deletions are preserved.
Missing-pod detection keeps using the full pod map.
In a local synthetic benchmark with 10,000 executor pods and 100 watch updates
per batch, snapshot updates plus lifecycle processing for healthy-pod batches
dropped from 14.9 ms to 0.011 ms.
This is similar to SPARK-55963, which optimized snapshot traversal in
{{ExecutorPodsAllocator}}.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
