Optimize muti-column grouping with StringView/ByteView #19364
Closed
Conversation
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
alamb
commented
Dec 16, 2025
|
|
||
| fn vectorized_equal_to_inner( | ||
| /// Comparison when there are no nulls in array | ||
| fn vectorized_equal_to_no_nulls( |
Contributor
Author
There was a problem hiding this comment.
Change 1 is to create a second copy of the loop when there are no nulls (to avoid the null check and give LLVM a better chance to optimize)
| array: &GenericByteViewArray<B>, | ||
| rhs_row: usize, | ||
| ) -> bool { | ||
| // SAFETY: the row indexes passed to vectorized_equal are in bounds |
Contributor
Author
There was a problem hiding this comment.
Optimization 2: skip bounds check on data access
| ) | ||
| }; | ||
| exist_inline == input_inline | ||
| // the views are inlined and the lengths are equal, so just |
Contributor
Author
There was a problem hiding this comment.
Optimization 3: just compare the view directly rather than breaking it into parts first
Contributor
Author
|
run benchmarks |
Contributor
Author
|
run benchmark tpch |
Contributor
|
run benchmarks |
|
🤖 |
Contributor
Author
|
My runner script keeps dying when there is a problem with the scripts. I am working on a way to keep it going (make it more resilent to failures) |
|
🤖: Benchmark completed Details
|
|
🤖 |
|
🤖: Benchmark completed Details
|
|
🤖 |
Contributor
Author
|
TPCH Q1 does have the pattern that is optimized in this PR (multiple group by columns) so it is plausible that the benefits are measured there datafusion/benchmarks/queries/q1.sql Lines 1 to 21 in 32951c3 |
Contributor
Author
I added some error checking here: alamb/datafusion-benchmarking@64ebd3a |
|
🤖: Benchmark completed Details
|
alamb
force-pushed
the
alamb/multi_byte_view_comparison
branch
from
39690eb to
da34c4c
Compare
Contributor
Author
|
Marking back to draft as I try out some other ideas from @Dandandan |
Contributor
|
run benchmark tpcds |
|
🤖 |
|
🤖: Benchmark completed Details
|
Contributor
Author
|
I benchmarked with/without the no buffers optimization and it seems to be worth around 1% (10ms out of 990ms) hyperfine --warmup 3 " ./datafusion-cli-starting -c \"select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;\" "Starting (before da34c4c) After da34c4c It seems reproducable: |
Contributor
Author
|
I have a better version here: |
github-merge-queue bot
pushed a commit
that referenced
this pull request
Dec 20, 2025
…5% faster (#19413) ## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> - Part of #18411 - Closes #19344 - Closes #19364 Note this is an alternate to #19364 ## Rationale for this change @camuel found a query where DuckDB's raw grouping is is faster. I looked into it and much of the difference can be explained by better vectorization in the comparisons and short string optimizations ## What changes are included in this PR? Optimize (will comment inline) ## Are these changes tested? By CI. See also benchmark results below. I tested manually as well Create Data: ```shell nice tpchgen-cli --tables=lineitem --format=parquet --scale-factor 100 ``` Run query: ```shell hyperfine --warmup 3 " datafusion-cli -c \"select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;\" " ``` Before (main): 1.368s ```shell Benchmark 1: datafusion-cli -c "select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;" Time (mean ± σ): 1.393 s ± 0.020 s [User: 16.778 s, System: 0.688 s] Range (min … max): 1.368 s … 1.438 s 10 runs ``` After (this PR) 1.022s ```shell Benchmark 1: ./datafusion-cli-multi-gby-try2 -c "select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;" Time (mean ± σ): 1.022 s ± 0.015 s [User: 11.685 s, System: 0.644 s] Range (min … max): 1.005 s … 1.052 s 10 runs ``` I have a PR that improves string view hashing performance too, see - #19374 ## Are there any user-facing changes? Faster performance
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.
Which issue does this PR close?
Rationale for this change
@camuel found a query where DuckDB's raw grouping is is faster.
I looked into it and much of the difference can be explained by better vectorization in the comparisons and short string optimizations
What changes are included in this PR?
Optimize (will comment inline)
Are these changes tested?
By CI. See also benchmark results below. I tested manually as well
Create Data:
Run query:
Before (main)
After (this PR)
I have some thoughts to improve string view hashing performance too -- will make as a separate PR
Are there any user-facing changes?
Faster performance