[
https://issues.apache.org/jira/browse/MAPREDUCE-7474?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17838340#comment-17838340
]
ASF GitHub Bot commented on MAPREDUCE-7474:
-------------------------------------------
steveloughran commented on code in PR #6716:
URL: https://github.com/apache/hadoop/pull/6716#discussion_r1569264943
##########
hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapreduce/lib/output/committer/manifest/stages/CleanupJobStage.java:
##########
@@ -142,64 +154,93 @@ protected Result executeStage(
}
Outcome outcome = null;
- IOException exception;
+ IOException exception = null;
+ boolean baseDirDeleted = false;
// to delete.
LOG.info("{}: Deleting job directory {}", getName(), baseDir);
if (args.deleteTaskAttemptDirsInParallel) {
- // Attempt to do a parallel delete of task attempt dirs;
- // don't overreact if a delete fails, but stop trying
- // to delete the others, and fall back to deleting the
- // job dir.
- Path taskSubDir
- = getStageConfig().getJobAttemptTaskSubDir();
- try (DurationInfo info = new DurationInfo(LOG,
- "parallel deletion of task attempts in %s",
- taskSubDir)) {
- RemoteIterator<FileStatus> dirs =
- RemoteIterators.filteringRemoteIterator(
- listStatusIterator(taskSubDir),
- FileStatus::isDirectory);
- TaskPool.foreach(dirs)
- .executeWith(getIOProcessors())
- .stopOnFailure()
- .suppressExceptions(false)
- .run(this::rmTaskAttemptDir);
- getIOStatistics().aggregate((retrieveIOStatistics(dirs)));
-
- if (getLastDeleteException() != null) {
- // one of the task attempts failed.
- throw getLastDeleteException();
+
+ // parallel delete of task attempt dirs.
+
+ if (args.parallelDeleteAttemptBaseDeleteFirst) {
+ // attempt to delete the base dir first.
+ // This can reduce ABFS delete load but may time out
+ // (which the fallback to parallel delete will handle).
+ // on GCS it is slow.
+ try (DurationInfo info = new DurationInfo(LOG, true,
+ "Initial delete of %s", baseDir)) {
+ exception = deleteOneDir(baseDir);
+ if (exception == null) {
+ // success: record this as the outcome, which
+ // will skip the parallel delete.
+ outcome = Outcome.DELETED;
+ baseDirDeleted = true;
+ } else {
+ // failure: log and continue
+ LOG.warn("{}: Exception on initial attempt at deleting base dir
{}\n"
+ + "attempting parallel delete",
+ getName(), baseDir, exception);
+ }
+ }
+ }
+ if (!baseDirDeleted) {
Review Comment:
note that the next stage will fail immediately on the list() call, which is
caught as a FNFE and treated as success
> [ABFS] Improve commit resilience and performance in Manifest Committer
> ----------------------------------------------------------------------
>
> Key: MAPREDUCE-7474
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-7474
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: client
> Affects Versions: 3.4.0, 3.3.6
> Reporter: Steve Loughran
> Assignee: Steve Loughran
> Priority: Major
> Labels: pull-request-available
>
> * Manifest committer is not resilient to rename failures on task commit
> without HADOOP-18012 rename recovery enabled.
> * large burst of delete calls noted: are they needed?
> relates to HADOOP-19093 but takes a more minimal approach with goal of
> changes in manifest committer only.
> Initial proposed changes
> * retry recovery on task commit rename, always (repeat save, delete, rename)
> * audit delete use and see if it can be pruned
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
