Reuse UTF-8 scratch for DataTable dictionary decoding

[xiangfu0]

Decoding a V4 DataTable allocates a temporary UTF-8 byte array for each string dictionary entry. Reuse one method-local scratch array, growing to the required length only when necessary. Each returned String owns its contents. Empty entries, wire encoding, the standalone string decoder and malformed-input behavior remain unchanged. The change includes regression tests for buffer state, String ownership and complete V4 messages.

Reproduction and benchmark

Measure DataTableFactory.getDataTable on already encoded V4 messages, including source-buffer duplication and complete decoding, using heap, direct and read-only heap buffers. Input preparation/encoding, SQL parsing, scans, aggregation, JNI and transport are excluded. Each fork validates all decoded rows, schema, metadata and source-buffer position before timing.

Workloads: NUMERIC_5000 has 5,000 INT/LONG rows and no string dictionary; STRING_ONE has one entry; ASCII_5000 has 5,000 unique 64-byte strings; UTF8_5000 uses Japanese, supplementary-plane and accented characters; OUTLIER_5000 starts with one 262,152-byte string followed by 4,999 64-byte strings; GROWING_512 has 512 distinct strings increasing from 32 to 4,120 bytes in 8-byte steps, requiring a larger scratch array for every entry.

Measured against upstream 5771d6acea60cd72738116835111cf7c49965373 on an Apple M2 Max (12 cores, 32 GiB), macOS ARM64 and Temurin JDK 25. Three alternating 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. Ordinary shared-host activity remained; no task builds or other task benchmarks ran during timing.

Both arms used identical measurement code and a pinned runtime; only the DataTableImplV4 classes differed. Filename-sorted classpaths, runtime manifests and source hashes were checked before/after each arm. This is a controlled class comparison, not a packaged deployment comparison.

Results

Times and allocations are medians of three arm means; ratios are medians of paired baseline/candidate times, not ratios of displayed medians. Values above 1 favor the candidate; ranges retain all three pairs.

Scenario	Buffer	Baseline µs/op	Candidate µs/op	Paired ratio	Pair range	Allocation B/op before → after	Reduction
ASCII_5000	DIRECT	88.281	86.486	1.033×	0.988–1.110	980953 → 581033	40.8%
ASCII_5000	HEAP	87.289	85.614	1.028×	1.016–1.040	980945 → 581025	40.8%
ASCII_5000	READ_ONLY	88.503	85.481	1.029×	1.029–1.035	980945 → 581025	40.8%
GROWING_512	DIRECT	98.151	98.561	0.996×	0.975–1.007	2161593 → 2161593	0.0%
GROWING_512	HEAP	97.623	97.544	1.001×	0.992–1.013	2161585 → 2161585	0.0%
GROWING_512	READ_ONLY	98.533	98.186	0.991×	0.988–1.016	2161585 → 2161585	0.0%
NUMERIC_5000	DIRECT	2.309	2.297	1.004×	0.999–1.034	60952 → 60952	0.0%
NUMERIC_5000	HEAP	2.339	2.365	1.003×	0.989–1.013	60944 → 60944	0.0%
NUMERIC_5000	READ_ONLY	2.379	2.295	1.006×	1.005–1.048	60944 → 60944	0.0%
OUTLIER_5000	DIRECT	112.614	109.849	1.025×	1.003–1.103	1505130 → 1105210	26.6%
OUTLIER_5000	HEAP	111.067	113.448	0.968×	0.959–1.006	1505122 → 1105202	26.6%
OUTLIER_5000	READ_ONLY	111.548	114.548	0.974×	0.951–0.979	1505122 → 1105202	26.6%
STRING_ONE	DIRECT	0.154	0.150	1.021×	1.001–1.036	1152 → 1152	0.0%
STRING_ONE	HEAP	0.146	0.146	1.000×	0.984–1.001	1144 → 1144	0.0%
STRING_ONE	READ_ONLY	0.153	0.150	0.999×	0.968–1.021	1144 → 1144	0.0%
UTF8_5000	DIRECT	388.269	376.446	1.033×	1.022–1.046	2540957 → 2061053	18.9%
UTF8_5000	HEAP	388.805	372.515	1.047×	1.039–1.053	2540949 → 2061045	18.9%
UTF8_5000	READ_ONLY	389.744	374.906	1.040×	1.025–1.043	2540949 → 2061045	18.9%

ASCII/UTF-8 dictionary medians show modest latency gains and 40.8%/18.9% less allocation. Outlier-first dictionaries allocate 26.6% less, but the original heap/read-only median paired slowdowns were about 3.3%/2.7%. Growing-length and single-entry controls also include latency non-wins; numeric, single-entry and growing-length allocation is unchanged. Allocation here means bytes allocated per operation, not peak heap or RSS.

Separate longer outlier follow-up

The original read-only pair 2 was 109.718 → 115.326 µs/op, a +5.111% slowdown, flagged by benchmark review. Its original three-pair median paired slowdown was 2.689%. These findings remain part of the assessment.

Three fresh AB/BA/AB pairs used unchanged source/runtime, five 1-second warmups and ten 1-second measurements across all three outlier buffer types; other settings were unchanged. Source/runtime checks passed before/after each arm. These results are separate from the original protocol and do not replace it.

OUTLIER_5000 buffer	Baseline µs/op	Candidate µs/op	Median paired ratio	All pair ratios	Allocation reduction
DIRECT	116.079	108.123	1.062×	1.090, 1.062, 1.031	26.6%
HEAP	109.871	108.674	1.004×	1.004, 0.996, 1.011	26.6%
READ_ONLY	110.133	108.883	1.009×	1.006, 1.012, 1.009	26.6%

The longer protocol did not reproduce the >5% regression or explain its cause; one longer heap pair remained slower. The allocation reduction repeated. The original tradeoff remains: these shared-host component measurements establish neither universal latency improvement, query throughput, production capacity nor 10× performance.

Validation

114 tests passed: DataTableStringDictionaryTest (48), DataTableDictionarySerDeTest (8), DataTableSerDeTest (9), SelectionOperatorServiceTest (5) and InterSegmentGroupBySingleValueQueriesTest (44). Coverage includes empty entries/dictionaries, Unicode, malformed UTF-8 replacement, negative counts/lengths, truncation, buffer position/limit/mark, independence of earlier Strings and complete V4 scalar/MV/null/metadata/exception messages across heap/direct, read-only and sliced buffers. Independent correctness and testing source reviews found no actionable issues; the benchmark review's original slowdown remains disclosed above.

Recorded local checks passed: normal compilation/tests, packaging, Spotless, Checkstyle, license format/check and Apache RAT. Scoped lint retained the baseline’s two constructor this-escape warnings and four dependency-manifest path warnings, with no new source warnings; it excluded the ASF-header dangling-doc category. A prior same-base full-reactor Xlint run failed in unchanged Zstandard code on a missing JetBrains annotation and was not repeated.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.69%. Comparing base (5771d6a) to head (a42ffd9).

Additional details and impacted files

@@             Coverage Diff              @@
##             master   #19535      +/-   ##
============================================
- Coverage     67.72%   67.69%   -0.03%     
  Complexity     1450     1450              
============================================
  Files          3490     3490              
  Lines        225032   225042      +10     
  Branches      35527    35530       +3     
============================================
- Hits         152393   152346      -47     
- Misses        60612    60664      +52     
- Partials      12027    12032       +5     

Flag	Coverage Δ
integration	100.00% <ø> (ø)
integration1	100.00% <ø> (ø)
integration2	0.00% <ø> (ø)
java-25	67.69% <100.00%> (-0.03%)	⬇️
lane-a	100.00% <ø> (ø)
lane-b	0.00% <ø> (ø)
temurin	67.69% <100.00%> (-0.03%)	⬇️
unittests	67.69% <100.00%> (-0.03%)	⬇️
unittests1	57.81% <100.00%> (+<0.01%)	⬆️
unittests2	39.46% <0.00%> (-0.03%)	⬇️

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.