c21 opened a new pull request #32210:
URL: https://github.com/apache/spark/pull/32210
<!--
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'.
-->
### 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.
-->
A major pain point for spark users to stay away from using shuffled hash
join is out of memory issue. Shuffled hash join tends to have OOM issue because
it allocates in-memory hashed relation (`UnsafeHashedRelation` or
`LongHashedRelation`) for build side, and there's no recovery (e.g.
fallback/spill) once the size of hashed relation grows and cannot fit in
memory. On the other hand, shuffled hash join is more CPU and IO efficient than
sort merge join when joining one large table and a small table (but small table
is too large to be broadcasted), as SHJ does not sort the large table, but SMJ
needs to do that. See historical discussion in
https://github.com/apache/spark/pull/11788 for evidence.
To improve the reliability of shuffled hash join, a fallback mechanism can
be introduced to avoid shuffled hash join OOM issue automatically. Similarly we
already have a fallback to sort-based aggregation for hash aggregate. The idea
is:
* Build hashed relation as current, but stop adding rows to hashed relation
if there's no enough memory. Do not throw exception, and do not fail the
task/query.
* Sort stream side and build side on join keys if necessary. Note here we
need to read all build rows in hashed relation back and destruct the hashed
relation on the fly to free memory.
* Execute sort merge join on sorted stream & build side.
Note:
(1).the fallback is automatic and happened per task, which means task 0 can
incur the fallback e.g. if it has a big build side, but task 1,2 don't need to
incur the fallback depending on the size of hashed relation.
(2).there's no major code change for SHJ and SMJ. Major change is around
`HashedRelation` to introduce some new methods, e.g.
`HashedRelation.destructiveValues()` to return an Iterator of build side rows
in hashed relation and destruct hashed relation on the fly.
(3).Per this PR, a new config
`spark.sql.join.enableShuffledHashJoinFallback` is introduced to enable/disable
this feature (disable by default). This PR only supports fallback for
non-code-gen execution path. Fallback for code-gen will be added in followup
PRs, as it will depend on sort merge join code-gen work (SPARK-34705).
### 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.
-->
Shuffled hash join OOM is a huge pain point for users and developers. This
is the major reason why people stay away from shuffled hash join. This can
improve reliability for shuffled hash join quite a bit.
### 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.
-->
Added unit test in `JoinSuite.scala`.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
