tgravescs opened a new pull request #27053: [WIP][SPARK-27495][Core][YARN][k8s] Stage Level Scheduling code for reference URL: https://github.com/apache/spark/pull/27053 <!-- 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. --> - ### 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 is all the code for stage level scheduling feature - except for the UI changes. This is meant to be for a reference when reviewing as I'm splitting this into mulitple prs with the intention its easier to review. Note that only YARN currently supports this and it requires dynamic allocation to be enabled because currently we get new executors that match the profile exactly. We do not try to fit tasks into executors that were acquired for a different profile. At a high level in order to support having different stages with different ResourceProfiles the changes required include: - Add a ResourceProfileManager that tracks the profiles and is used to map a Resource Profile Id to the actual Resource Profile. This allows us to pass around and store the id rather then the entire profile. - Introduce the concept of a default profile. This is essentially the profile you get today without stage level scheduling from the application level configs. - ImmutableResourceProfile - this is the actual resource profile used internally to Spark that is immutable. This is to allow the user to create and change a ResourceProfile in their code but as soon as they associated the profile with an RDD then spark internally uses the Immutable version so that it doesn't change. - YARN cluster manager updated to handle the profiles and request the correct containers from YARN. I had to introduce using priorities here because YARN doesn't allow you to create containers with different resources within the same priority. Now we have the priority = ResourceProfile Id and its easy to match the container we get from Yarn to what ResourceProfile we requested it for. - ExecutorAllocationManager, ExecutorMonitor, CoarseGrainedExecutorBackend - Updated to handle tracking the executors per ResourceProfile. - Scheduler - updated to handle the ResourceProfile associated with an RDD. It creates the Stage with the appropriate ResourceProfile. It has logic for handling conflicting ResourceProfiles when multiple RDD are put into the same stage that have different ResourceProfiles. The default behavior is to throw an exception, but there is a config that will allow the scheduler to merge the profiles using the max value of each resource. The task scheduler was updated to make sure the resources of each executor meet the task resources for that profile and to assign them out properly. - I updated all the locations that used the hardcoded task cpus or other global configs to use the ResourceProfile based configs. - RDD api added and ResourceProfile, ExecutorResourceRequests, and TaskResourceRequests made public. ### 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. --> Allow for different stages to use different executor/task resources ### 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'. --> Yes the RDD.withResources and ResourceProfile, ExecutorResourceRequest, TaskResourceRequest apis ### 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. --> Unit tests and manually.
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
