chengxis-mdb commented on PR #16394:
URL: https://github.com/apache/lucene/pull/16394#issuecomment-4972322766
Benchmark results from luceneutil (wikimedium10m), run by my colleague
Tianxiao Wei, including a new `ConstInDisj` task that issues pure-should
disjunctions of constant-score term-set clauses (the shape described above):
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
ConstInDisj 8.28 (2.8%) 93.00
(89.8%) 1023.9% ( 906% - 1148%) 0.000
```
All other tasks are within noise (full run below). Note `ConstMSM2`,
`OrHighHigh`, `AndHighHigh` etc. are unchanged — the gate keeps every
non-disjunction-backed constant-score clause and all impact-scored clauses on
their existing paths.
<details><summary>Full luceneutil output</summary>
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
HighTermDayOfYearSort 750.80 (6.2%) 725.13
(4.2%) -3.4% ( -13% - 7%) 0.042
range 6520.02 (5.3%) 6341.16
(6.7%) -2.7% ( -14% - 9%) 0.152
BM25MSM2 188.93 (4.7%) 184.19
(9.4%) -2.5% ( -15% - 12%) 0.288
MedIntervalsOrdered 1235.81 (5.6%) 1206.70
(5.6%) -2.4% ( -12% - 9%) 0.184
Fuzzy1 268.96 (1.8%) 262.73
(2.8%) -2.3% ( -6% - 2%) 0.002
Prefix3 1639.66 (5.5%) 1603.03
(3.9%) -2.2% ( -11% - 7%) 0.138
BrowseMonthSSDVFacets 204.59 (8.6%) 200.24
(1.5%) -2.1% ( -11% - 8%) 0.274
AndHighHigh 1162.49 (4.3%) 1138.39
(4.8%) -2.1% ( -10% - 7%) 0.151
OrHighHigh 994.12 (3.4%) 976.73
(5.3%) -1.7% ( -10% - 7%) 0.217
LowSloppyPhrase 756.00 (5.9%) 743.84
(6.1%) -1.6% ( -12% - 11%) 0.396
HighSloppyPhrase 443.93 (7.3%) 436.81
(6.7%) -1.6% ( -14% - 13%) 0.466
LowSpanNear 861.32 (4.4%) 847.66
(4.3%) -1.6% ( -9% - 7%) 0.253
Respell 186.44 (2.3%) 183.73
(2.8%) -1.5% ( -6% - 3%) 0.071
HighPhrase 473.57 (4.2%) 467.95
(6.4%) -1.2% ( -11% - 9%) 0.489
MedSpanNear 627.57 (3.9%) 620.20
(6.0%) -1.2% ( -10% - 9%) 0.462
MedTerm 3226.22 (3.7%) 3193.18
(3.6%) -1.0% ( -7% - 6%) 0.371
OrHighLow 1583.34 (2.9%) 1568.06
(3.3%) -1.0% ( -6% - 5%) 0.321
LowIntervalsOrdered 1423.38 (3.1%) 1411.08
(3.4%) -0.9% ( -7% - 5%) 0.399
MedSloppyPhrase 97.97 (4.9%) 97.13
(4.7%) -0.9% ( -9% - 9%) 0.571
Wildcard 517.48 (2.8%) 513.67
(3.8%) -0.7% ( -7% - 6%) 0.484
AndHighMed 1582.62 (3.8%) 1572.13
(5.7%) -0.7% ( -9% - 9%) 0.664
PKLookup 467.78 (3.5%) 465.29
(3.7%) -0.5% ( -7% - 6%) 0.638
BrowseDateSSDVFacets 37.58 (1.8%) 37.43
(1.2%) -0.4% ( -3% - 2%) 0.393
HighIntervalsOrdered 269.55 (4.7%) 268.51
(5.5%) -0.4% ( -10% - 10%) 0.810
OrHighMed 1499.26 (2.5%) 1494.71
(4.4%) -0.3% ( -6% - 6%) 0.786
Fuzzy2 109.82 (1.8%) 109.51
(3.0%) -0.3% ( -4% - 4%) 0.711
HighTerm 2525.48 (4.3%) 2523.08
(4.5%) -0.1% ( -8% - 9%) 0.946
ConstMSM2 1746.68 (4.1%) 1745.22
(5.9%) -0.1% ( -9% - 10%) 0.959
BrowseMonthTaxoFacets 138.20 (5.9%) 138.34
(6.8%) 0.1% ( -11% - 13%) 0.961
IntNRQ 1982.74 (5.1%) 1984.90
(4.5%) 0.1% ( -9% - 10%) 0.943
BrowseDateTaxoFacets 160.59 (0.9%) 161.54
(0.6%) 0.6% ( 0% - 2%) 0.017
BrowseDayOfYearTaxoFacets 160.14 (0.8%) 161.17
(0.6%) 0.6% ( 0% - 2%) 0.006
LowTerm 3684.93 (4.4%) 3708.81
(6.5%) 0.6% ( -9% - 12%) 0.714
BrowseDayOfYearSSDVFacets 178.71 (3.6%) 179.93
(2.4%) 0.7% ( -5% - 6%) 0.487
BrowseRandomLabelTaxoFacets 141.68 (2.7%) 142.71
(2.0%) 0.7% ( -3% - 5%) 0.335
HighSpanNear 528.43 (11.0%) 533.07
(11.5%) 0.9% ( -19% - 26%) 0.805
LowPhrase 939.27 (5.8%) 949.72
(3.9%) 1.1% ( -8% - 11%) 0.474
AndHighLow 2840.67 (3.7%) 2878.15
(4.3%) 1.3% ( -6% - 9%) 0.296
BrowseRandomLabelSSDVFacets 127.32 (4.1%) 129.11
(4.8%) 1.4% ( -7% - 10%) 0.322
MedPhrase 1046.02 (4.4%) 1060.90
(4.0%) 1.4% ( -6% - 10%) 0.283
HighTermMonthSort 1434.77 (3.4%) 1455.43
(4.7%) 1.4% ( -6% - 9%) 0.267
ConstInDisj 8.28 (2.8%) 93.00
(89.8%) 1023.9% ( 906% - 1148%) 0.000
```
</details>
We also validated end-to-end on the production-like workload where we found
the regression: top-k disjunctions of dense constant-score term-set clauses
recover most of the slowdown introduced in 10.3 (avg latency 316ms → ~236ms
across replicated runs), while equals/point-based and impact-scored shapes are
unchanged. An earlier unconditional version of this patch (bulk path for every
constant-score clause) measurably regressed single-postings and bit-set-backed
clauses, which is what motivated the `DisjunctionDISIApproximation` gate.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
