linzebing opened a new pull request #27223: [SPARK-30511][CORE] Spark marks ended speculative tasks as pending leads to holding idle executors URL: https://github.com/apache/spark/pull/27223 ### What changes were proposed in this pull request? <!-- Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. See the examples below. 1. If you refactor some codes with changing classes, showing the class hierarchy will help reviewers. 2. If you fix some SQL features, you can provide some references of other DBMSes. 3. If there is design documentation, please add the link. 4. If there is a discussion in the mailing list, please add the link. --> Currently, when speculative tasks finished/failed/got killed, they are still considered as pending and count towards the calculation of number of needed executors. To be more accurate: `stageAttemptToNumSpeculativeTasks(stageAttempt)` is incremented on onSpeculativeTaskSubmitted, but never decremented. `stageAttemptToNumSpeculativeTasks -= stageAttempt` is performed on stage completion. **This means Spark is marking ended speculative tasks as pending, which leads to Spark to hold more executors that it actually needs!** This PR fixes this issue by updating `stageAttemptToSpeculativeTaskIndices` and `stageAttemptToNumSpeculativeTasks` on speculative tasks completion. This PR also addresses some other minor issues: scheduler behavior after receiving an intentionally killed task event; try to address [SPARK-28403](https://issues.apache.org/jira/browse/SPARK-28403). ### Why are the changes needed? <!-- Please clarify why the changes are needed. For instance, 1. If you propose a new API, clarify the use case for a new API. 2. If you fix a bug, you can clarify why it is a bug. --> This has caused resource wastage in our production with speculation enabled. With aggressive speculation, we found data skewed jobs can hold hundreds of idle executors with less than 10 tasks running. An easy repro of the issue (`--conf spark.speculation=true --conf spark.executor.cores=4 --conf spark.dynamicAllocation.maxExecutors=1000` in cluster mode): ``` val n = 4000 val someRDD = sc.parallelize(1 to n, n) someRDD.mapPartitionsWithIndex( (index: Int, it: Iterator[Int]) => { if (index < 300 && index >= 150) { Thread.sleep(index * 1000) // Fake running tasks } else if (index == 300) { Thread.sleep(1000 * 1000) // Fake long running tasks } it.toList.map(x => index + ", " + x).iterator }).collect ``` You will see when running the last task, we would be hold 38 executors (see below), which is exactly (152 + 3) / 4 = 38. ![image](https://user-images.githubusercontent.com/9404831/72469112-9a7fac00-3793-11ea-8f50-74d0ab7325a4.png) ### Does this PR introduce any user-facing change? <!-- If yes, please clarify the previous behavior and the change this PR proposes - provide the console output, description and/or an example to show the behavior difference if possible. If no, write 'No'. --> No ### How was this patch tested? <!-- If tests were added, say they were added here. Please make sure to add some test cases that check the changes thoroughly including negative and positive cases if possible. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> Added a comprehensive unit test. Test with the above repro shows that we are holding 2 executors at the end ![image](https://user-images.githubusercontent.com/9404831/72469177-bbe09800-3793-11ea-850f-4a2c67142899.png)
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org