Repository navigation
perf: skip wasted per-batch work in TopK insert_batch - #25657
Merged
Merged
Conversation
Evaluate the dynamic threshold filter before computing sort keys, so a batch the filter rejects outright no longer pays for them. With an expression key such as `l_extendedprice + 0` those keys were otherwise computed twice per batch, once directly and once inside the filter. Also move the input batch into the heap entry instead of cloning it, encoding the last-row prefix needed for early completion before the heap takes ownership, and clone only the elapsed_compute timer rather than the whole BaselineMetrics in TopK and PartitionedTopK. TopK elapsed_compute on TPC-H SF1 lineitem, ORDER BY l_extendedprice + 0 DESC LIMIT 1, median of 5: 151.97 ms before, 93.58 ms after (-38%). sort-tpch --limit 10 shows no regression.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #25657 +/- ##
==========================================
- Coverage 82.55% 82.55% -0.01%
==========================================
Files 1141 1141
Lines 440209 440365 +156
Branches 440209 440365 +156
==========================================
+ Hits 363409 363533 +124
- Misses 54829 54837 +8
- Partials 21971 21995 +24 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
An empty batch goes through the rejected batch path and must not try to read a last row. Check that it leaves the heap and store alone.
neilconway
self-requested a review
September 27, 2026 16:50
neilconway
approved these changes
Sep 28, 2026
neilconway
left a comment
neilconway
left a comment
Contributor
There was a problem hiding this comment.
@dwsmith1983 Nice PR. Overall lgtm! One minor suggestion for consistency.
| value.into_array(batch.num_rows()) | ||
| }) | ||
| .collect::<Result<Vec<_>>>()?; | ||
| let elapsed_compute = self.metrics.baseline.elapsed_compute().clone(); |
Contributor
There was a problem hiding this comment.
Can we make the same fix in PartitionedTopKRank and PartitionedTopKDenseRank?
Contributor
Author
There was a problem hiding this comment.
Done in 2c25298: PartitionedTopKRank and PartitionedTopKDenseRank now take the same elapsed_compute handle as the other two insert paths. Also merged main.
PartitionedTopKRank and PartitionedTopKDenseRank still copied the whole BaselineMetrics set on every insert_batch call just to start a scoped timer. Take a handle to elapsed_compute alone, matching TopK and PartitionedTopK.
neilconway
enabled auto-merge
September 28, 2026 15:31
Contributor
|
Thanks @dwsmith1983 ! |
Omega359
pushed a commit
to Omega359/arrow-datafusion
that referenced
this pull request
Oct 11, 2026
## Which issue does this PR close? - Related to apache#17098. ## Rationale for this change Once the TopK heap is full, most input batches are rejected outright by the threshold filter. `insert_batch` still evaluated every sort expression over the whole batch before checking the filter, then evaluated the same expressions again inside the filter, and cloned the batch even when nothing from it was kept. With an expression sort key that doubles the per-batch cost for no output. ## What changes are included in this PR? - The threshold filter runs before the sort keys are evaluated; fully rejected batches skip sort-key evaluation. - The batch moves into the heap entry instead of being cloned. The last row's common prefix is encoded before the move so early completion still compares against the same boundary at the same point. - The per-batch timer clones one metric instead of the whole baseline set, here and in `PartitionedTopK`, `PartitionedTopKRank`, and `PartitionedTopKDenseRank`. `SELECT l_orderkey, l_extendedprice FROM lineitem ORDER BY l_extendedprice + 0 DESC LIMIT 1` on TPC-H SF1, TopK `elapsed_compute`, median of five: 152 ms before, 94 ms after. `sort-tpch --limit 10` shows no change. One nuance: a sort-key expression that errors only on rows the threshold already rejects no longer raises, because those batches skip sort-key evaluation. The filter evaluates the same expression over the batch first, so this can only differ for a secondary key under short-circuit evaluation. ## What is the testing strategy for this PR? New unit tests in `datafusion/physical-plan/src/topk`: `test_batch_rejected_by_threshold_leaves_heap_untouched` and `test_expression_sort_key_matches_column_sort_key`. Existing TopK tests and the `topk`, `window` and `limit` sqllogictest files pass unchanged. ## Are there any user-facing changes? No.
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
4 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?
ORDER BY time DESC LIMIT 1queries ( TopK or aggr rewrite) #17098.Rationale for this change
Once the TopK heap is full, most input batches are rejected outright by the threshold filter.
insert_batchstill evaluated every sort expression over the whole batch before checking the filter, then evaluated the same expressions again inside the filter, and cloned the batch even when nothing from it was kept. With an expression sort key that doubles the per-batch cost for no output.What changes are included in this PR?
PartitionedTopK,PartitionedTopKRank, andPartitionedTopKDenseRank.SELECT l_orderkey, l_extendedprice FROM lineitem ORDER BY l_extendedprice + 0 DESC LIMIT 1on TPC-H SF1, TopKelapsed_compute, median of five: 152 ms before, 94 ms after.sort-tpch --limit 10shows no change.One nuance: a sort-key expression that errors only on rows the threshold already rejects no longer raises, because those batches skip sort-key evaluation. The filter evaluates the same expression over the batch first, so this can only differ for a secondary key under short-circuit evaluation.
What is the testing strategy for this PR?
New unit tests in
datafusion/physical-plan/src/topk:test_batch_rejected_by_threshold_leaves_heap_untouchedandtest_expression_sort_key_matches_column_sort_key. Existing TopK tests and thetopk,windowandlimitsqllogictest files pass unchanged.Are there any user-facing changes?
No.