Repository navigation
perf: reuse cached dictionary value hashes in vectorized_append - #25185
Merged
jayzhan211 merged 1 commit intoSep 15, 2026
Merged
Conversation
LiaCastaneda
marked this pull request as ready for review
September 11, 2026 12:19
LiaCastaneda
force-pushed
the
perf/dict-group-values-cache-hashes
branch
from
September 11, 2026 12:21
f0bcd4f to
e33ef31
Compare
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #25185 +/- ##
========================================
Coverage 81.92% 81.92%
========================================
Files 1132 1132
Lines 421192 421360 +168
Branches 421192 421360 +168
========================================
+ Hits 345041 345190 +149
- Misses 55756 55769 +13
- Partials 20395 20401 +6 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Contributor
|
run benchmark dictionary_group_values |
1 similar comment
Contributor
|
run benchmark dictionary_group_values |
13 of 19 tasks
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing perf/dict-group-values-cache-hashes (e33ef31) to 0da2151 (merge-base) diff Run configurationrun benchmark dictionary_group_valuesResults will be posted here when complete File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing perf/dict-group-values-cache-hashes (e33ef31) to 0da2151 (merge-base) diff Run configurationrun benchmark dictionary_group_valuesResults will be posted here when complete File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: Comparing perf/dict-group-values-cache-hashes (e33ef31) to 0da2151 (merge-base) diff Run configurationrun benchmark dictionary_group_valuesCPU Details (lscpu)Details
Resource Usagedictionary_group_values — base (merge-base)
dictionary_group_values — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: Comparing perf/dict-group-values-cache-hashes (e33ef31) to 0da2151 (merge-base) diff Run configurationrun benchmark dictionary_group_valuesCPU Details (lscpu)Details
Resource Usagedictionary_group_values — base (merge-base)
dictionary_group_values — branch
File an issue against this benchmark runner |
Contributor
|
there is no perf difference entirely due to how the benchmarks are set up. in the benchmarks the values array are created fresh for each if maintainers are interested in seeing what the performance gains this would mean for parquet files we may need to wait for #24227 to be over the finish line so we can run @adriangbot with the flag enabled. |
Contributor
Author
|
dictionary value arrays are commonly re-used in prod workloads, this inneficiency is something we noticed in our service, we saw the following scenario: with becuase there is no group ordering the partial aggregate goes through improvements are visible with this benchmark |
jayzhan211
approved these changes
Sep 12, 2026
jayzhan211
left a comment
Contributor
There was a problem hiding this comment.
Thanks @LiaCastaneda! I think this PR is good to merge. I also noticed another potential improvement, so I've filed a follow-up issue #25219
Contributor
Author
|
Thanks @jayzhan211, if we could also merge #25198 to showcase the results, it would be awsome |
This was referenced Sep 14, 2026
Contributor
|
I will merge those 2 PRs |
github-merge-queue
Bot
removed this pull request from the merge queue due to failed status checks
Sep 15, 2026
Contributor
Author
|
🤔 which check failed? |
lyne7-sc
pushed a commit
to lyne7-sc/datafusion
that referenced
this pull request
Sep 15, 2026
…#25198) ## Which issue does this PR close? [Related to apache#24822. No issue closed.](apache#25185) ## Rationale for this change The benchmarks in this file build a fresh values array per batch and cap cardinality at the batch size. Neither shape matches a real plan: `take` and `filter` clone the values `Arc` and rewrite only the keys, so batches downstream of a repartition share one values array, and because it is never compacted its cardinality can far exceed the rows in a batch. Per-batch work that scales with dictionary cardinality is invisible under the current parameters. ## What changes are included in this PR? Adds `bench_shared_values_arc`: 32 batches of 8192 rows sharing one values array, at cardinalities 8192, 100k and 500k. ## What is the testing strategy for this PR? Benchmark-only. Ran the bench locally and `cargo clippy --all-targets -D warnings`. ## Are there any user-facing changes? No
Contributor
|
Sometimes it needs a retry — not sure what's going on. |
This was referenced Sep 16, 2026
LiaCastaneda
added a commit
to LiaCastaneda/datafusion
that referenced
this pull request
Sep 16, 2026
…he#25185) ## Which issue does this PR close? - Closes #. ## Rationale for this change `vectorized_append` is the path taken when we have non streaming hash aggregations (when there is no ordering in the `group by`), right now it re hashed the entire dictionary values on every batch unlike append_val which that only does this if the values array was already seen before. ## What changes are included in this PR? `vectorized_append` now reuses the cached value hashes when a batch carries the same dictionary values array as the previous one (this is the common case of a repartition or filter), instead of re hashing the whole dictionary and re-initialising the map on every batch. This extends the existing `Arc::ptr_eq` cache from apache#24418, which only covered the scalar `append_val` path. ## What is the testing strategy for this PR? All dict tests pass, I have a PR for a benchmark that exercises this path -> apache#25198 ## Are there any user-facing changes? no, this is a pure perf PR (cherry picked from commit a0631ed)
LiaCastaneda
added a commit
to DataDog/datafusion
that referenced
this pull request
Sep 16, 2026
…he#25185) (#185) ## Which issue does this PR close? - Closes #. ## Rationale for this change `vectorized_append` is the path taken when we have non streaming hash aggregations (when there is no ordering in the `group by`), right now it re hashed the entire dictionary values on every batch unlike append_val which that only does this if the values array was already seen before. ## What changes are included in this PR? `vectorized_append` now reuses the cached value hashes when a batch carries the same dictionary values array as the previous one (this is the common case of a repartition or filter), instead of re hashing the whole dictionary and re-initialising the map on every batch. This extends the existing `Arc::ptr_eq` cache from apache#24418, which only covered the scalar `append_val` path. ## What is the testing strategy for this PR? All dict tests pass, I have a PR for a benchmark that exercises this path -> apache#25198 ## Are there any user-facing changes? no, this is a pure perf PR (cherry picked from commit a0631ed)
neilconway
pushed a commit
to neilconway/datafusion
that referenced
this pull request
Sep 30, 2026
…che#25652) ## Which issue does this PR close? - Closes apache#25219. ## Rationale for this change Follow-up to apache#25185. `vectorized_equal_to` was left untouched and still rebuilt the same `val_idx -> inner_slot` map from scratch on every batch. ## What changes are included in this PR? - Added `lookup_inner_slot`, which resolves a single `val_idx` to its slot in `inner` and memorizes the result in `val_to_inner`. - Changed `vectorized_equal_to` to call `sync_value_cache` and resolve slots lazily per row. - Removed `build_lookup_table`, the `rhs_rows.len() < num_distinct` guard and `equal_to_per_row`. ## What is the testing strategy for this PR? Covered by the existing `dictionary.rs` unit tests. ## Are there any user-facing changes? No. This is an internal performance change with no public API or behaviour change.
Omega359
pushed a commit
to Omega359/arrow-datafusion
that referenced
this pull request
Oct 11, 2026
…#25198) ## Which issue does this PR close? [Related to apache#24822. No issue closed.](apache#25185) ## Rationale for this change The benchmarks in this file build a fresh values array per batch and cap cardinality at the batch size. Neither shape matches a real plan: `take` and `filter` clone the values `Arc` and rewrite only the keys, so batches downstream of a repartition share one values array, and because it is never compacted its cardinality can far exceed the rows in a batch. Per-batch work that scales with dictionary cardinality is invisible under the current parameters. ## What changes are included in this PR? Adds `bench_shared_values_arc`: 32 batches of 8192 rows sharing one values array, at cardinalities 8192, 100k and 500k. ## What is the testing strategy for this PR? Benchmark-only. Ran the bench locally and `cargo clippy --all-targets -D warnings`. ## Are there any user-facing changes? No
Omega359
pushed a commit
to Omega359/arrow-datafusion
that referenced
this pull request
Oct 11, 2026
…he#25185) ## Which issue does this PR close? - Closes #. ## Rationale for this change `vectorized_append` is the path taken when we have non streaming hash aggregations (when there is no ordering in the `group by`), right now it re hashed the entire dictionary values on every batch unlike append_val which that only does this if the values array was already seen before. ## What changes are included in this PR? `vectorized_append` now reuses the cached value hashes when a batch carries the same dictionary values array as the previous one (this is the common case of a repartition or filter), instead of re hashing the whole dictionary and re-initialising the map on every batch. This extends the existing `Arc::ptr_eq` cache from apache#24418, which only covered the scalar `append_val` path. ## What is the testing strategy for this PR? All dict tests pass, I have a PR for a benchmark that exercises this path -> apache#25198 ## Are there any user-facing changes? no, this is a pure perf PR
Omega359
pushed a commit
to Omega359/arrow-datafusion
that referenced
this pull request
Oct 11, 2026
…che#25652) ## Which issue does this PR close? - Closes apache#25219. ## Rationale for this change Follow-up to apache#25185. `vectorized_equal_to` was left untouched and still rebuilt the same `val_idx -> inner_slot` map from scratch on every batch. ## What changes are included in this PR? - Added `lookup_inner_slot`, which resolves a single `val_idx` to its slot in `inner` and memorizes the result in `val_to_inner`. - Changed `vectorized_equal_to` to call `sync_value_cache` and resolve slots lazily per row. - Removed `build_lookup_table`, the `rhs_rows.len() < num_distinct` guard and `equal_to_per_row`. ## What is the testing strategy for this PR? Covered by the existing `dictionary.rs` unit tests. ## Are there any user-facing changes? No. This is an internal performance change with no public API or behaviour change.
kosiew
added a commit
to kumarUjjawal/datafusion
that referenced
this pull request
Oct 14, 2026
## Which issue does this PR close? - Part of apache#24117. ## Rationale for this change The existing dictionary benchmark only measures pre-encoded dictionaries with cardinality expressed as a fraction of row count. That is not enough to derive a safe automatic encoding policy: native dictionaries, per-batch dictionaries, `Utf8View`, plain strings, and runtime encoding have materially different costs, and the crossover also depends on absolute NDV, value width, and batch layout. I rebased this PR onto current `main` after apache#25185 merged and reran the relevant matrix, so the evidence includes the new cached dictionary-value-hash path. The crossover remains clear; the cache optimization does not make dictionary grouping universally profitable. Current-main results (`700ad9cf2`, release/LTO, Apple M5 Max, Rust 1.98.1; 10M rows, 8,192-row input batches, one target partition; median of iterations 2-6 after two warmups; output normalized back to `Utf8`): | value bytes | NDV | `Utf8View` | dict/shared | vs view | dict/per-batch | vs view | |---:|---:|---:|---:|---:|---:|---:| | 16 | 1,000 | 50.29 ms | 43.48 ms | -13.5% | 43.52 ms | -13.5% | | 16 | 10,000 | 54.24 ms | 91.61 ms | +68.9% | 94.03 ms | +73.3% | | 64 | 1,000 | 72.65 ms | 47.02 ms | -35.3% | 47.69 ms | -34.4% | | 64 | 10,000 | 79.69 ms | 115.85 ms | +45.4% | 115.04 ms | +44.4% | These are local measurements, not a proposed universal threshold. They show why apache#24117 needs a cost decision based on absolute NDV and value width rather than a blanket dictionary rewrite. Runtime encoding also lost in every earlier measured case, which is why the planned optimizer work targets native Parquet dictionary reads rather than inserting a runtime cast. ## What changes are included in this PR? - Compare pre-encoded dictionaries, `Utf8View`, plain strings, and runtime dictionary casts. - Configure absolute NDV, value width, and input batch size. - Model shared dictionary values and independent per-batch values. - Optionally normalize dictionary group keys back to `Utf8` in the output so output-type preservation is included in the measurement. - Preserve the existing default benchmark behavior and case names. - Reject invalid benchmark configurations instead of silently changing the requested workload. - Verify that all encodings produce identical grouped counts. This PR only establishes the benchmark surface needed to evaluate apache#24117; it does not add an optimizer rule or select a universal threshold. ## What is the testing strategy for this PR? On rebased head `5fb605471`: - `cargo test -p datafusion-benchmarks --lib` (153 passed) - `cargo test -p datafusion-benchmarks --lib --no-default-features` (153 passed) - `cargo clippy -p datafusion-benchmarks --all-targets --no-deps -- -D warnings` - `cargo clippy -p datafusion-benchmarks --all-targets --all-features --no-deps -- -D warnings` - `cargo fmt --all -- --check` - Current-main CLI benchmark matrix above; every case returned exactly the requested NDV ## Are there any user-facing changes? No DataFusion runtime or public API behavior changes. This only adds benchmark CLI controls. --------- Co-authored-by: kosiew <kosiew@gmail.com>
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
6 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
vectorized_appendis the path taken when we have non streaming hash aggregations (when there is no ordering in thegroup by), right now it re hashed the entire dictionary values on every batch unlike append_val which that only does this if the values array was already seen before.What changes are included in this PR?
vectorized_appendnow reuses the cached value hashes when a batch carries the same dictionary values array as the previous one (this is the common case of a repartition or filter), instead of re hashing the whole dictionary and re-initialising the map on every batch. This extends the existingArc::ptr_eqcache from #24418, which only covered the scalarappend_valpath.What is the testing strategy for this PR?
All dict tests pass, I have a PR for a benchmark that exercises this path -> #25198
Are there any user-facing changes?
no, this is a pure perf PR