[SPARK-58189][PYTHON] Refactor SQL_GROUPED_MAP_PANDAS_UDF_WITH_STATE#57339
Closed
Yicong-Huang wants to merge 2 commits into
Closed
[SPARK-58189][PYTHON] Refactor SQL_GROUPED_MAP_PANDAS_UDF_WITH_STATE#57339Yicong-Huang wants to merge 2 commits into
Yicong-Huang wants to merge 2 commits into
Conversation
HyukjinKwon
approved these changes
Jul 19, 2026
uros-b
approved these changes
Jul 20, 2026
uros-b
left a comment
uros-b
left a comment
Member
There was a problem hiding this comment.
Thank you @Yicong-Huang and @HyukjinKwon!
Yicong-Huang
added a commit
that referenced
this pull request
Jul 20, 2026
### What changes were proposed in this pull request? This PR refactors `SQL_GROUPED_MAP_PANDAS_UDF_WITH_STATE` (applyInPandasWithState) so that the worker uses the plain `ArrowStreamSerializer` for pure Arrow stream I/O, moving the per-eval-type logic from `ApplyInPandasWithStateSerializer` into `read_udfs()` in `worker.py`: reconstructing each `GroupState` from the trailing state-info column, slicing the data columns by each chunk's `(startOffset, numRows)`, regrouping chunks by grouping key (via the reused state object), invoking the UDF once per key with a lazy iterator of pandas DataFrames and its state, validating the returned DataFrames, and serializing the output plus updated state back into count/data/state Arrow RecordBatches bounded by `arrow_max_records_per_batch`. This mirrors the earlier refactor of `SQL_TRANSFORM_WITH_STATE_PANDAS_UDF` in [SPARK-57400](https://issues.apache.org/jira/browse/SPARK-57400). The `ApplyInPandasWithStateSerializer` class is kept for now and deleted in a follow-up cleanup. ### Why are the changes needed? Part of [SPARK-55388](https://issues.apache.org/jira/browse/SPARK-55388). Keeping serializers as pure Arrow stream I/O and concentrating eval-type-specific logic in `worker.py` makes the per-eval-type data flow explicit. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Existing tests. No behavior change. ASV comparison (`bench_eval_type.ApplyInPandasWithStateUDFTimeBench` / `ApplyInPandasWithStateUDFPeakmemBench`, `-a repeat=3`): before = `upstream/master`, after = this PR. Values are from one representative run per side; the conclusion is consistent across runs. All deltas are within run-to-run noise (the larger swings track high before-side variance, e.g. `few_groups_sm/count` before `131ms+-228ms`), showing no regression. Peak memory is unchanged. ```text time_worker scenario udf before after diff --------------- ------------- -------------- -------------- ------ few_groups_sm identity_udf 1.05s+-25.2ms 1.04s+-13.8ms -1.0% few_groups_sm sort_udf 1.07s+-37.3ms 1.08s+-7.33ms +0.9% few_groups_sm count_udf 131ms+-228ms 129ms+-255ms -1.5% few_groups_lg identity_udf 7.21s+-78.4ms 7.19s+-49.9ms -0.3% few_groups_lg sort_udf 7.57s+-75.1ms 7.33s+-9.87ms -3.2% few_groups_lg count_udf 800ms+-8.91ms 781ms+-8.37ms -2.4% many_groups_sm identity_udf 7.12s+-106ms 7.00s+-34.9ms -1.7% many_groups_sm sort_udf 7.85s+-58.7ms 7.83s+-24.1ms -0.3% many_groups_sm count_udf 3.26s+-16.5ms 3.24s+-32.8ms -0.6% many_groups_lg identity_udf 4.83s+-136ms 4.75s+-72.3ms -1.7% many_groups_lg sort_udf 5.33s+-85ms 5.00s+-42.5ms -6.2% many_groups_lg count_udf 951ms+-6.36ms 935ms+-2.69ms -1.7% wide_cols identity_udf 6.13s+-30ms 5.94s+-29.8ms -3.1% wide_cols sort_udf 6.36s+-404ms 6.30s+-78.9ms -0.9% wide_cols count_udf 913ms+-25.6ms 940ms+-22.3ms +3.0% mixed_cols identity_udf 4.57s+-72.5ms 4.84s+-49.7ms +5.9% mixed_cols sort_udf 4.91s+-86.6ms 4.81s+-68.8ms -2.0% mixed_cols count_udf 739ms+-20.2ms 709ms+-5.46ms -4.1% nested_struct identity_udf 9.62s+-183ms 9.27s+-22.4ms -3.6% nested_struct sort_udf 12.1s+-62.1ms 11.0s+-154ms -9.1% nested_struct count_udf 3.95s+-21ms 3.53s+-27.1ms -10.6% ``` ```text peakmem_worker scenario udf before after --------------- ------------- ------- ------- few_groups_sm identity_udf 19M 19M few_groups_sm sort_udf 19M 19M few_groups_sm count_udf 2M 2M few_groups_lg identity_udf 156M 156M few_groups_lg sort_udf 156M 156M few_groups_lg count_udf 3M 3M many_groups_sm identity_udf 62M 63M many_groups_sm sort_udf 62M 63M many_groups_sm count_udf 20M 20M many_groups_lg identity_udf 61M 61M many_groups_lg sort_udf 61M 61M many_groups_lg count_udf 6M 6M wide_cols identity_udf 163M 163M wide_cols sort_udf 163M 163M wide_cols count_udf 6M 6M mixed_cols identity_udf 66M 66M mixed_cols sort_udf 66M 66M mixed_cols count_udf 4M 4M nested_struct identity_udf 103M 103M nested_struct sort_udf 103M 103M nested_struct count_udf 10M 10M ``` ### Was this patch authored or co-authored using generative AI tooling? No. Closes #57339 from Yicong-Huang/refactor-apply-in-pandas-with-state. Authored-by: Yicong Huang <17627829+Yicong-Huang@users.noreply.github.com> Signed-off-by: Yicong-Huang <17627829+Yicong-Huang@users.noreply.github.com> (cherry picked from commit ab513d6) Signed-off-by: Yicong-Huang <17627829+Yicong-Huang@users.noreply.github.com>
Contributor
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
This PR refactors
SQL_GROUPED_MAP_PANDAS_UDF_WITH_STATE(applyInPandasWithState) so that the worker uses the plainArrowStreamSerializerfor pure Arrow stream I/O, moving the per-eval-type logic fromApplyInPandasWithStateSerializerintoread_udfs()inworker.py: reconstructing eachGroupStatefrom the trailing state-info column, slicing the data columns by each chunk's(startOffset, numRows), regrouping chunks by grouping key (via the reused state object), invoking the UDF once per key with a lazy iterator of pandas DataFrames and its state, validating the returned DataFrames, and serializing the output plus updated state back into count/data/state Arrow RecordBatches bounded byarrow_max_records_per_batch. This mirrors the earlier refactor ofSQL_TRANSFORM_WITH_STATE_PANDAS_UDFin SPARK-57400. TheApplyInPandasWithStateSerializerclass is kept for now and deleted in a follow-up cleanup.Why are the changes needed?
Part of SPARK-55388. Keeping serializers as pure Arrow stream I/O and concentrating eval-type-specific logic in
worker.pymakes the per-eval-type data flow explicit.Does this PR introduce any user-facing change?
No.
How was this patch tested?
Existing tests. No behavior change.
ASV comparison (
bench_eval_type.ApplyInPandasWithStateUDFTimeBench/ApplyInPandasWithStateUDFPeakmemBench,-a repeat=3): before =upstream/master, after = this PR. Values are from one representative run per side; the conclusion is consistent across runs. All deltas are within run-to-run noise (the larger swings track high before-side variance, e.g.few_groups_sm/countbefore131ms+-228ms), showing no regression. Peak memory is unchanged.Was this patch authored or co-authored using generative AI tooling?
No.