Reuse sort-key storage across rejected GROUP BY candidates

[xiangfu0]

GROUP BY top-K trimming allocates sort-key arrays and intermediate wrappers for rejected candidates. Reuse that storage across consecutive rejections; create a wrapper with the current key/record when a candidate is admitted. Immediate admissions retain the existing path, including comparator, tie handling and output order. An admitted row's sort-key array is never subsequently overwritten.

Reproduction and benchmark

Select top K from a deterministic prepared map of numeric group results. Small/large K retain 20/5,000 of 8,192 groups; the large-N case retains 5,000 of 100,000. Best-first rejects every later candidate; worst-first admits every later candidate. Full-sort/no-trim retain all 8,192 with/without sorting, and unsorted top-K retains 5,000 without ordering the output. Timing includes construction of TableResizer, ORDER BY extraction, heap maintenance and optional output sorting; it excludes input-map preparation, SQL parsing, scanning, aggregation, JNI and transport.

Measured against upstream 5771d6acea60cd72738116835111cf7c49965373 on an Apple M2 Max (12 cores, 32 GiB), macOS ARM64 and Temurin JDK 25. Three AB/BA/AB pairs, one fresh JMH fork per case/arm, one thread, 512 MiB heap, three 500 ms warmups and five 500 ms measurements, with GC profiling. Both arms use the same pinned runtime and measurement code; only the TableResizer class family differs. Sorted classpath order and runtime hashes were checked. The workstation was shared, with no task builds or other task benchmarks during timing.

Results

Times and allocations are medians of three arm means. Ratios are medians of paired baseline/candidate time ratios; ranges retain all three pairs. Values above 1 favor the candidate.

Scenario	Baseline us/op	Candidate us/op	Paired ratio	Pair range	Allocation B/op before -> after	Reduction
FULL_SORT	1357.213	1348.913	1.002x	0.968-1.019	491307 -> 491307	0.0%
LARGE_K_BEST_FIRST	684.138	675.338	1.012x	1.001-1.013	357073 -> 280489	21.4%
LARGE_K_RANDOM	1456.872	1457.745	0.995x	0.990-1.016	433692 -> 410004	5.5%
LARGE_K_WORST_FIRST	1091.692	1080.635	1.010x	0.997-1.053	433687 -> 433687	0.0%
LARGE_N_RANDOM	7392.171	6354.504	1.160x	1.156-1.163	4840558 -> 1234591	74.5%
NO_TRIM	0.091	0.092	0.998x	0.923-1.034	384 -> 384	0.0%
SMALL_K_BEST_FIRST	78.667	63.169	1.245x	1.158-1.316	197705 -> 1601	99.2%
SMALL_K_RANDOM	188.672	84.294	2.259x	2.217-2.310	393835 -> 10697	97.3%
SMALL_K_WORST_FIRST	288.432	297.038	0.971x	0.966-0.994	393836 -> 393836	0.0%
UNSORTED_TOP_K	613.472	605.108	1.027x	0.978-1.084	433681 -> 409992	5.5%

The small-K all-admitted case is about 3% slower with unchanged allocation; large-K random has a 0.995× median ratio. The no-trim control includes a 0.923× individual pair despite its 0.998× median. These non-wins remain part of the result. This is a controlled component comparison, not packaged deployment throughput or proof that every query benefits.

Validation

Recorded local validation: 81 tests passed—13 table-resizer, 9 indexed-table, 3 deterministic-indexed-table, 12 sorted-combine and 44 inter-segment GROUP BY tests. Regression coverage includes rejection runs, immediate admissions, exact identities/order, ties, both resize directions, nulls/NaN/signed zero and aggregate-state ownership. Every measurement fork validated complete result membership, uniqueness, ordering, record identity and input preservation against an independent oracle before timing.

Normal compilation/tests, Spotless, Checkstyle and license format/check passed. Independent correctness, testing and performance reviews found no actionable issues. Scoped lint reported no source warnings and four existing dependency-manifest path warnings. Full-reactor Xlint was not repeated after the same-base attempt failed in unchanged Zstandard code on a missing JetBrains annotation.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.72%. Comparing base (5771d6a) to head (e597d09).

Additional details and impacted files

@@            Coverage Diff            @@
##             master   #19534   +/-   ##
=========================================
  Coverage     67.72%   67.72%           
  Complexity     1450     1450           
=========================================
  Files          3490     3490           
  Lines        225032   225043   +11     
  Branches      35527    35530    +3     
=========================================
+ Hits         152393   152406   +13     
+ Misses        60612    60608    -4     
- Partials      12027    12029    +2     

Flag	Coverage Δ
integration	100.00% <ø> (ø)
integration1	100.00% <ø> (ø)
integration2	0.00% <ø> (ø)
java-25	67.72% <100.00%> (+<0.01%)	⬆️
lane-a	100.00% <ø> (ø)
lane-b	0.00% <ø> (ø)
temurin	67.72% <100.00%> (+<0.01%)	⬆️
unittests	67.72% <100.00%> (+<0.01%)	⬆️
unittests1	57.83% <100.00%> (+0.02%)	⬆️
unittests2	39.48% <0.00%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.