Ngone51 opened a new pull request #32385: URL: https://github.com/apache/spark/pull/32385
<!-- Thanks for sending a pull request! Here are some tips for you: 1. If this is your first time, please read our contributor guidelines: https://spark.apache.org/contributing.html 2. Ensure you have added or run the appropriate tests for your PR: https://spark.apache.org/developer-tools.html 3. If the PR is unfinished, add '[WIP]' in your PR title, e.g., '[WIP][SPARK-XXXX] Your PR title ...'. 4. Be sure to keep the PR description updated to reflect all changes. 5. Please write your PR title to summarize what this PR proposes. 6. If possible, provide a concise example to reproduce the issue for a faster review. 7. If you want to add a new configuration, please read the guideline first for naming configurations in 'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'. 8. If you want to add or modify an error message, please read the guideline first: https://spark.apache.org/error-message-guidelines.html --> ### 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. --> This PR proposes to add checksum support for shuffle blocks. The basic idea is: On the mapper side, we'll wrap a `CheckedOutputStream` upon the `FileOutputStream` to calculate the checksum (use the same checksum calculator `Adler32` with broadcast) for each shuffle block (a.k.a partition) at the same time when we writing map output files. And similar to the index file, we'll have a checksum file to save these checksums. On the reducer side, we'll also wrap a `CheckedInputStream` upon the `FileInputStream` to read the block. When block corruption is detected, we'll try to diagnose corruption for the cause: First, we'll use the `CheckedInputStream` to consume the remaining data of the corrupted block to calculate the checksum (`c1`); Second, the reducer send an RPC request called `DiagnoseCorruption` (which contains `c1`) to the server (where the reducer executed) Third, the server will read (using a very small memory) the corresponding block back from the disk and calculate the checksum (`c2`) again for it. And also read back the checksum(`c3`) of the block saved in the checksum file. Then, if `c2 != c3`, we'll suspect the corruption is caused by the disk issue. Otherwise, if `c1 != c3`, we'll suspect the corruption is caused by the network issue. Otherwise, the cause remains unknown. The server then will reply to the reducer with `CorruptionCause` containing the cause. Fourth, the reducer needs to take action after it receives the cause. If it's a disk issue or unknown, it will throw fetch failure directly. If it's a network issue, it will re-fetch the block later. Also note that, if the corruption happens inside `BufferReleasingInputStream`, the reducer will throw the fetch failure immediately no matter what the cause is since the data has been partially consumed by downstream RDDs. If corruption happens again after retry, the reducer will throw the fetch failure directly this time without the diagnosis. Overall, I think we don't introduce severe overhead with this proposal. In a normal case, the checksum is calculated in the streaming way as well as other streams, e.g., encryption, compression. And the major overhead here is that we need an extra data file traverse in the error case in order to calculate the checksum (`c2`). And the proposal in this PR is much simpler compared to the previous one https://github.com/apache/spark/pull/15894/ (abandoned due to complexity ), which introduce more overhead as it need to traverse the data file twice for every block. In that proposal, the checksum is appended to each block data, so it's also invasive to the existing code. ### 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. --> Shuffle data corruption is a long-standing issue in Spark. For example, in SPARK-18105, people continually reports corruption issue. However, data corruption is difficult to reproduce in most cases and even harder to tell the root cause. We don't know if it's a Spark issue or not. With the checksum support for the shuffle, Spark itself can at least distinguish the cause between disk and network, which is very important for users. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as the documentation fix. 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 possible, please also clarify if this is a user-facing change compared to the released Spark versions or within the unreleased branches such as master. If no, write 'No'. --> Yes. 1) Added a conf `spark.shuffle.checksum` to let user enables/disables the checksum (enabled by default) 2) With checksum enabled, users can know the possible cause of corruption rather than "Stream is corrupted" only. ### 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 an end-to-end unit test in `ShuffleSuite`. I'll add more tests if the community accepts the proposal. -- 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 --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org