otterc commented on code in PR #36601:
URL: https://github.com/apache/spark/pull/36601#discussion_r879790477
##########
core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala:
##########
@@ -2449,7 +2459,12 @@ private[spark] class DAGScheduler(
val currentEpoch = maybeEpoch.getOrElse(mapOutputTracker.getEpoch)
logDebug(s"Considering removal of executor $execId; " +
s"fileLost: $fileLost, currentEpoch: $currentEpoch")
- if (!executorFailureEpoch.contains(execId) || executorFailureEpoch(execId)
< currentEpoch) {
+ // Check if the execId is a shuffle push merger
+ // We do not remove the executor if it is,
+ // and only remove the outputs on that host/executor.
Review Comment:
Nit: `and only remove the outputs on the host`
##########
core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala:
##########
@@ -2461,7 +2476,7 @@ private[spark] class DAGScheduler(
clearCacheLocs()
}
if (fileLost) {
- val remove = if (ignoreShuffleFileLostEpoch) {
+ val remove = if (ignoreShuffleFileLostEpoch || isShuffleMerger) {
Review Comment:
`removeExecutorAndUnregisterOutputs` is being called from 2 places:
1. Fetch failure handling: So here `ignoreShuffleFileLostEpoch` is going to
be `true`. Then it doesn't matter if the execId is a shuffleMergerId or not.
2. Executor Lost handling: When it is called from here, then it is never
going to be a `shuffleMergerId`.
Given this, I think it might be better to remove this `isShuffleMerger`
check here and just add a comment that when the fetch failure is for a merged
shuffle chunk, `ignoreShuffleFileLostEpoch` is true, and so all the files will
be removed.
WDYT? cc. @mridulm
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
