[SPARK-46957][CORE] Decommission migrated shuffle files should be able to cleanup from executor #47037
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@dongjoon-hyun @holdenk @mridulm @jiangxb1987 @bozhang2820 Could you help take a look? Thanks! |
|
@dongjoon-hyun Do you think it is the same issue for |
jiangxb1987
approved these changes
Jun 24, 2024
|
@attilapiros @LuciferYang @warrenzhu25 Anyone else want to take a look at this fix? |
|
I'm going to merge this PR by the end of day if there is no objection. |
LuciferYang
approved these changes
Jun 27, 2024
Ngone51
added a commit
that referenced
this pull request
Jun 27, 2024
…e to cleanup from executor ### What changes were proposed in this pull request? This PR uses `SortShuffleManager#taskIdMapsForShuffle` to track the migrated shuffle files on the destination executor. ### Why are the changes needed? This is a long-standing bug in decommission where the migrated shuffle files can't be cleaned up from the executor. Normally, the shuffle files are tracked by `taskIdMapsForShuffle` during the map task execution. Upon receiving the `RemoveShuffle(shuffleId)` request from driver, executor can clean up those shuffle files by searching `taskIdMapsForShuffle`. However, for the migrated shuffle files by decommission, they lose the track in the destination executor's `taskIdMapsForShuffle` and can't be deleted as a result. Note this bug only affects shuffle removal on the executor. For shuffle removal on the external shuffle service (when `spark.shuffle.service.removeShuffle` enabled and the executor stores the shuffle files has gone), we don't rely on `taskIdMapsForShuffle` but using the specific shuffle block id to locate the shuffle file directly. So it won't be an issue there. ### Does this PR introduce _any_ user-facing change? No. (Common users won't see the difference underlying.) ### How was this patch tested? Add unit test. ### Was this patch authored or co-authored using generative AI tooling? No. Closes #47037 from Ngone51/SPARK-46957. Authored-by: Yi Wu <yi.wu@databricks.com> Signed-off-by: Yi Wu <yi.wu@databricks.com> (cherry picked from commit b5a55e4) Signed-off-by: Yi Wu <yi.wu@databricks.com>
Ngone51
added a commit
that referenced
this pull request
Jun 27, 2024
…e to cleanup from executor ### What changes were proposed in this pull request? This PR uses `SortShuffleManager#taskIdMapsForShuffle` to track the migrated shuffle files on the destination executor. ### Why are the changes needed? This is a long-standing bug in decommission where the migrated shuffle files can't be cleaned up from the executor. Normally, the shuffle files are tracked by `taskIdMapsForShuffle` during the map task execution. Upon receiving the `RemoveShuffle(shuffleId)` request from driver, executor can clean up those shuffle files by searching `taskIdMapsForShuffle`. However, for the migrated shuffle files by decommission, they lose the track in the destination executor's `taskIdMapsForShuffle` and can't be deleted as a result. Note this bug only affects shuffle removal on the executor. For shuffle removal on the external shuffle service (when `spark.shuffle.service.removeShuffle` enabled and the executor stores the shuffle files has gone), we don't rely on `taskIdMapsForShuffle` but using the specific shuffle block id to locate the shuffle file directly. So it won't be an issue there. ### Does this PR introduce _any_ user-facing change? No. (Common users won't see the difference underlying.) ### How was this patch tested? Add unit test. ### Was this patch authored or co-authored using generative AI tooling? No. Closes #47037 from Ngone51/SPARK-46957. Authored-by: Yi Wu <yi.wu@databricks.com> Signed-off-by: Yi Wu <yi.wu@databricks.com> (cherry picked from commit b5a55e4) Signed-off-by: Yi Wu <yi.wu@databricks.com>
|
Thanks, merged to master/3.5/3.4! |
|
There are compliaton failures against Java 8 in branch-3.5/3.4. I'll send a followup fix soon. |
Ngone51
added a commit
to Ngone51/spark
that referenced
this pull request
Jun 27, 2024
…e to cleanup from executor ### What changes were proposed in this pull request? This PR uses `SortShuffleManager#taskIdMapsForShuffle` to track the migrated shuffle files on the destination executor. ### Why are the changes needed? This is a long-standing bug in decommission where the migrated shuffle files can't be cleaned up from the executor. Normally, the shuffle files are tracked by `taskIdMapsForShuffle` during the map task execution. Upon receiving the `RemoveShuffle(shuffleId)` request from driver, executor can clean up those shuffle files by searching `taskIdMapsForShuffle`. However, for the migrated shuffle files by decommission, they lose the track in the destination executor's `taskIdMapsForShuffle` and can't be deleted as a result. Note this bug only affects shuffle removal on the executor. For shuffle removal on the external shuffle service (when `spark.shuffle.service.removeShuffle` enabled and the executor stores the shuffle files has gone), we don't rely on `taskIdMapsForShuffle` but using the specific shuffle block id to locate the shuffle file directly. So it won't be an issue there. ### Does this PR introduce _any_ user-facing change? No. (Common users won't see the difference underlying.) ### How was this patch tested? Add unit test. ### Was this patch authored or co-authored using generative AI tooling? No. Closes apache#47037 from Ngone51/SPARK-46957. Authored-by: Yi Wu <yi.wu@databricks.com> Signed-off-by: Yi Wu <yi.wu@databricks.com>
|
FYI, created a separate backport PR along with the compliation error fix: #47122 |
attilapiros
pushed a commit
to attilapiros/spark
that referenced
this pull request
Oct 4, 2024
…e to cleanup from executor ### What changes were proposed in this pull request? This PR uses `SortShuffleManager#taskIdMapsForShuffle` to track the migrated shuffle files on the destination executor. ### Why are the changes needed? This is a long-standing bug in decommission where the migrated shuffle files can't be cleaned up from the executor. Normally, the shuffle files are tracked by `taskIdMapsForShuffle` during the map task execution. Upon receiving the `RemoveShuffle(shuffleId)` request from driver, executor can clean up those shuffle files by searching `taskIdMapsForShuffle`. However, for the migrated shuffle files by decommission, they lose the track in the destination executor's `taskIdMapsForShuffle` and can't be deleted as a result. Note this bug only affects shuffle removal on the executor. For shuffle removal on the external shuffle service (when `spark.shuffle.service.removeShuffle` enabled and the executor stores the shuffle files has gone), we don't rely on `taskIdMapsForShuffle` but using the specific shuffle block id to locate the shuffle file directly. So it won't be an issue there. ### Does this PR introduce _any_ user-facing change? No. (Common users won't see the difference underlying.) ### How was this patch tested? Add unit test. ### Was this patch authored or co-authored using generative AI tooling? No. Closes apache#47037 from Ngone51/SPARK-46957. Authored-by: Yi Wu <yi.wu@databricks.com> Signed-off-by: Yi Wu <yi.wu@databricks.com>
| .get | ||
|
|
||
| val newShuffleFiles = shuffleFiles.diff(existingShuffleFiles) | ||
| assert(newShuffleFiles.size >= shuffleBlockUpdates.size) |
There was a problem hiding this comment.
Unfortunately, this seems to introduce a flakiness.
[info] - SPARK-46957: Migrated shuffle files should be able to cleanup from executor *** FAILED *** (36 seconds, 137 milliseconds)
[info] 0 was not greater than or equal to 6 (BlockManagerDecommissionIntegrationSuite.scala:423)
There was a problem hiding this comment.
It must be a race condition on JDK21 but I did not managed to reproduce it locally. Neither on arm64 nor on a x86_64 (using docker).
There was a problem hiding this comment.
As I see a jira already exists for this: https://issues.apache.org/jira/browse/SPARK-49297
I am working on a possible fix.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
This PR uses
SortShuffleManager#taskIdMapsForShuffleto track the migrated shuffle files on the destination executor.Why are the changes needed?
This is a long-standing bug in decommission where the migrated shuffle files can't be cleaned up from the executor. Normally, the shuffle files are tracked by
taskIdMapsForShuffleduring the map task execution. Upon receiving theRemoveShuffle(shuffleId)request from driver, executor can clean up those shuffle files by searchingtaskIdMapsForShuffle. However, for the migrated shuffle files by decommission, they lose the track in the destination executor'staskIdMapsForShuffleand can't be deleted as a result.Note this bug only affects shuffle removal on the executor. For shuffle removal on the external shuffle service (when
spark.shuffle.service.removeShuffleenabled and the executor stores the shuffle files has gone), we don't rely ontaskIdMapsForShufflebut using the specific shuffle block id to locate the shuffle file directly. So it won't be an issue there.Does this PR introduce any user-facing change?
No. (Common users won't see the difference underlying.)
How was this patch tested?
Add unit test.
Was this patch authored or co-authored using generative AI tooling?
No.