-
Notifications
You must be signed in to change notification settings - Fork 28.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[SPARK-19868] conflict TasksetManager lead to spark stopped #17208
Conversation
CC @kayousterhout or @squito |
This looks like the right change. In fact, I could have sworn we had recently merged in something like this -- maybe there is another pr still in flight which includes this? @jinxing64 perhaps this is in one of your open prs? The description needs to be updated, and we really should have a unit test (though with a very quick look I don't see a good way to test, I'll need to think about that part). Here is my suggestion for the description:
|
Jenkins, ok to test |
Test build #74212 has finished for PR 17208 at commit
|
Looks good. Expanding on Imran's comment, how about: We must set the taskset to zombie before the DAGScheduler handles the taskEnded event. It's possible the taskEnded event will cause the DAGScheduler to launch a new stage attempt (this happens when map output data was lost), and if this happens before the taskSet has been set to zombie, it will appear that we have conflicting task sets. |
@squito |
@liujianhuiouc do you have time to update the comment here? It would be great to get this in soon. |
to be clear, I agree with Kay's rewording (in particular, I meant stage attempt, not task attempt). Also I think its worth including a test. You can use this: squito@aac8d98 I know its very narrowly focused but it seems worth including. |
ok, I will update that. |
@liujianhuiouc have you had time to fix this up yet? |
@kayousterhout I have already update the comments, and fix this issue, do you mean i should merge the test case by squito |
Yes can you also merge @squito's test case? |
@kayousterhout Done |
lgtm assuming tests pass |
Test build #75290 has finished for PR 17208 at commit
|
@squito tests fails |
Test build #75299 has started for PR 17208 at commit |
@squito update the no-args ManualClock constructor �with initialized time |
Jenkins, retest this please |
Looks like the tests were manually killed (-9). Thanks for catching that and fixing @liujianhuiouc |
Test build #75312 has finished for PR 17208 at commit
|
LGTM merged this to master |
I think |
hmm I think you're right @zsxwing that we should be updating |
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. #17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. #21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, #21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. #22806 and #23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, #21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). #22806 and #23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes #23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. #17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. #21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, #21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. #22806 and #23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, #21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). #22806 and #23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes #23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. apache#17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. apache#21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, apache#21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. apache#22806 and apache#23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, apache#21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). apache#22806 and apache#23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes apache#23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. apache#17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. apache#21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, apache#21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. apache#22806 and apache#23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, apache#21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). apache#22806 and apache#23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes apache#23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. apache#17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. apache#21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, apache#21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. apache#22806 and apache#23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, apache#21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). apache#22806 and apache#23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes apache#23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. apache/spark#17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. apache/spark#21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, #21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. #22806 and #23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, #21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). #22806 and #23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes #23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit 7df5aa6)
…a stage ## What changes were proposed in this pull request? This is another attempt to fix the more-than-one-active-task-set-managers bug. apache#17208 is the first attempt. It marks the TSM as zombie before sending a task completion event to DAGScheduler. This is necessary, because when the DAGScheduler gets the task completion event, and it's for the last partition, then the stage is finished. However, if it's a shuffle stage and it has missing map outputs, DAGScheduler will resubmit it(see the [code](https://github.com/apache/spark/blob/v2.4.0/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1416-L1422)) and create a new TSM for this stage. This leads to more than one active TSM of a stage and fail. This fix has a hole: Let's say a stage has 10 partitions and 2 task set managers: TSM1(zombie) and TSM2(active). TSM1 has a running task for partition 10 and it completes. TSM2 finishes tasks for partitions 1-9, and thinks he is still active because he hasn't finished partition 10 yet. However, DAGScheduler gets task completion events for all the 10 partitions and thinks the stage is finished. Then the same problem occurs: DAGScheduler may resubmit the stage and cause more than one actice TSM error. apache#21131 fixed this hole by notifying all the task set managers when a task finishes. For the above case, TSM2 will know that partition 10 is already completed, so he can mark himself as zombie after partitions 1-9 are completed. However, apache#21131 still has a hole: TSM2 may be created after the task from TSM1 is completed. Then TSM2 can't get notified about the task completion, and leads to the more than one active TSM error. apache#22806 and apache#23871 are created to fix this hole. However the fix is complicated and there are still ongoing discussions. This PR proposes a simple fix, which can be easy to backport: mark all existing task set managers as zombie when trying to create a new task set manager. After this PR, apache#21131 is still necessary, to avoid launching unnecessary tasks and fix [SPARK-25250](https://issues.apache.org/jira/browse/SPARK-25250 ). apache#22806 and apache#23871 are its followups to fix the hole. ## How was this patch tested? existing tests. Closes apache#23927 from cloud-fan/scheduler. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Imran Rashid <irashid@cloudera.com> (cherry picked from commit cb20fbc) Signed-off-by: Imran Rashid <irashid@cloudera.com>
What changes were proposed in this pull request?
We must set the taskset to zombie before the DAGScheduler handles the taskEnded event. It's possible the taskEnded event will cause the DAGScheduler to launch a new stage attempt (this happens when map output data was lost), and if this happens before the taskSet has been set to zombie, it will appear that we have conflicting task sets.