utkarsh39 opened a new pull request, #44321:
URL: https://github.com/apache/spark/pull/44321
<!--
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 type or message, please read the
guideline first in
'core/src/main/resources/error/README.md'.
-->
### What changes were proposed in this pull request?
`AccumulableInfo` is one of the top heap consumers in driver's heap dumps
for stages with many tasks. For a stage with a large number of tasks
(**_O(100k)_**), we saw **30%** of the heap usage stemming from
`TaskInfo.accumulables()`.
The `TaskSetManager` today keeps around the TaskInfo objects
([ref1](https://github.com/apache/spark/blob/c1ba963e64a22dea28e17b1ed954e6d03d38da1e/core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala#L134),
[ref2](https://github.com/apache/spark/blob/c1ba963e64a22dea28e17b1ed954e6d03d38da1e/core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala#L192)))
and in turn the task metrics (`AccumulableInfo`) for every task attempt until
the stage is completed. This means that for stages with a large number of
tasks, we keep metrics for all the tasks (`AccumulableInfo`) around even when
the task has completed and its metrics have been aggregated. Given a task has a
large number of metrics, stages with many tasks end up with a large heap usage
in the form of task metrics.
This PR reduces the driver's heap usage for stages with many tasks by no
longer referencing the task metrics of completed tasks. Once a task is
completed in `TaskSetManager`, we no longer keep its metrics around. Upon task
completion, we clone the `TaskInfo` object and empty out the metrics for the
clone. The cloned `TaskInfo` is retained by the `TaskSetManager` while the
original `TaskInfo` object with the metrics is sent over to the `DAGScheduler`
where the task metrics are aggregated. Thus for a completed task,
`TaskSetManager` holds a `TaskInfo` object with empty metrics. This reduces the
memory footprint by ensuring that the number of task metric objects is
proportional to the number of active tasks and not to the total number of tasks
in the stage.
### Config to gate changes
The changes in the PR are guarded with the Spark conf
`spark.scheduler.dropTaskInfoAccumulablesOnTaskCompletion.enabled` which can be
used for rollback or staged rollouts.
### 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.
-->
Reduce driver's heap usage, especially for stages with many tasks
## Benchmarking
On a cluster running a scan stage with 100k tasks, the TaskSetManager's heap
usage dropped from 1.1 GB to 37 MB. This **reduced the total driver's heap
usage by 38%**, down to 2 GB from 3.5 GB.
**BEFORE**
**WITH FIX**
<img width="1386" alt="image"
src="https://github.com/databricks/runtime/assets/10495099/b85129c8-dc10-4ee2-898d-61c8e7449616";>
### 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'.
-->
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.
If benchmark tests were added, please run the benchmarks in GitHub Actions
for the consistent environment, and the instructions could accord to:
https://spark.apache.org/developer-tools.html#github-workflow-benchmarks.
-->
Added new tests and did benchmarking on a cluster.
### Was this patch authored or co-authored using generative AI tooling?
<!--
If generative AI tooling has been used in the process of authoring this
patch, please include the
phrase: 'Generated-by: ' followed by the name of the tool and its version.
If no, write 'No'.
Please refer to the [ASF Generative Tooling
Guidance](https://www.apache.org/legal/generative-tooling.html) for details.
-->
Generated-by: Github Copilot
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
