Stop using `SlowImpactsEnum` for terms whose `docFreq` is less than 128. by jpountz · Pull Request #14017 · apache/lucene

jpountz · 2024-11-25T08:30:42Z

We currently use SlowImpactsEnum for terms whose docFreq is less than 128 because it's convenient as these terms don't have impacts anyway. But a recent slowdown on nightly benchmarks suggests that this contributes to making some hot calls more polymorphic than we'd like, so this PR moves such terms back to the regular impacts enums.

We currently use `SlowImpactsEnum` for terms whose `docFreq` is less than 128 because it's convenient as these terms don't have impacts anyway. But a recent slowdown on nightly benchmarks suggests that this contributes to making some hot calls more polymorphic than we'd like, so this PR moves such terms back to the regular impacts enums.

jpountz · 2024-11-25T09:05:01Z

With Combined tasks in the file (call is 3-polymorphic in the baseline, bimorphic in the modified version):

                            TaskQPS baseline      StdDevQPS my_modified_version      StdDev                Pct diff p-value
                        PKLookup      277.89      (2.3%)      274.07      (2.6%)   -1.4% (  -6% -    3%) 0.079
             FilteredOrStopWords       33.44      (2.3%)       33.42      (1.9%)   -0.1% (  -4% -    4%) 0.935
            FilteredAndStopWords       48.48      (2.6%)       48.48      (1.8%)   -0.0% (  -4% -    4%) 0.994
     FilteredAnd2Terms2StopWords      195.64      (1.8%)      195.89      (1.4%)    0.1% (  -3% -    3%) 0.806
             FilteredAndHighHigh       62.16      (2.2%)       62.25      (1.7%)    0.1% (  -3% -    4%) 0.813
                  FilteredPhrase       25.15      (2.7%)       25.22      (2.5%)    0.3% (  -4% -    5%) 0.720
                          OrMany       18.96      (4.5%)       19.02      (5.6%)    0.3% (  -9% -   10%) 0.853
                FilteredOr3Terms      149.27      (2.0%)      149.75      (1.6%)    0.3% (  -3% -    4%) 0.574
                    CombinedTerm       34.03      (2.8%)       34.17      (2.2%)    0.4% (  -4% -    5%) 0.623
              FilteredOrHighHigh       52.50      (2.3%)       52.71      (1.8%)    0.4% (  -3% -    4%) 0.531
               FilteredOrHighMed      114.91      (2.3%)      115.46      (1.7%)    0.5% (  -3% -    4%) 0.465
                  FilteredOrMany        9.91      (2.8%)        9.96      (3.4%)    0.6% (  -5% -    6%) 0.572
               FilteredAnd3Terms      189.21      (1.8%)      190.35      (1.6%)    0.6% (  -2% -    4%) 0.257
      FilteredOr2Terms2StopWords      109.76      (1.6%)      110.43      (1.2%)    0.6% (  -2% -    3%) 0.169
                    FilteredTerm      159.61      (2.9%)      160.69      (2.2%)    0.7% (  -4% -    5%) 0.401
             CombinedAndHighHigh       15.58      (1.0%)       15.81      (0.8%)    1.5% (   0% -    3%) 0.000
              CombinedAndHighMed       57.04      (1.3%)       57.92      (0.8%)    1.5% (   0% -    3%) 0.000
              FilteredAndHighMed      123.14      (2.9%)      125.29      (3.0%)    1.8% (  -4% -    7%) 0.060
                       And3Terms      162.00      (4.3%)      167.29      (4.5%)    3.3% (  -5% -   12%) 0.019
             And2Terms2StopWords      152.25      (4.1%)      158.02      (4.2%)    3.8% (  -4% -   12%) 0.004
               CombinedOrHighMed       74.69      (0.7%)       78.15      (2.3%)    4.6% (   1% -    7%) 0.000
              CombinedOrHighHigh       19.55      (0.7%)       20.60      (2.5%)    5.4% (   2% -    8%) 0.000
                      AndHighMed      116.33      (4.0%)      122.79      (1.5%)    5.6% (   0% -   11%) 0.000
                     AndHighHigh       39.43      (4.2%)       41.74      (1.9%)    5.9% (   0% -   12%) 0.000
                    AndStopWords       27.82      (5.9%)       29.51      (6.7%)    6.1% (  -6% -   19%) 0.002
                      OrHighRare      265.91      (7.3%)      284.27      (5.1%)    6.9% (  -5% -   20%) 0.000
                        Or3Terms      155.29      (3.8%)      168.35      (5.7%)    8.4% (  -1% -   18%) 0.000
              Or2Terms2StopWords      146.73      (3.2%)      160.57      (5.7%)    9.4% (   0% -   18%) 0.000
                     OrStopWords       28.40      (5.6%)       32.31      (9.5%)   13.8% (  -1% -   30%) 0.000
                       OrHighMed      168.44      (3.9%)      197.35      (1.6%)   17.2% (  11% -   23%) 0.000
                      OrHighHigh       44.17      (4.6%)       53.19      (1.7%)   20.4% (  13% -   28%) 0.000

Without combined tasks in the file (call is bimorphic in the baseline, monomorphic in the modified version):

                            TaskQPS baseline      StdDevQPS my_modified_version      StdDev                Pct diff p-value
             FilteredOrStopWords       33.95      (1.7%)       33.73      (1.8%)   -0.7% (  -4% -    2%) 0.245
                  FilteredOrMany       10.23      (2.7%)       10.22      (2.9%)   -0.1% (  -5% -    5%) 0.928
              FilteredOrHighHigh       53.10      (1.9%)       53.11      (2.0%)    0.0% (  -3% -    3%) 0.962
                FilteredOr3Terms      150.83      (2.0%)      150.98      (2.3%)    0.1% (  -4% -    4%) 0.879
               FilteredOrHighMed      116.00      (2.0%)      116.30      (2.1%)    0.3% (  -3% -    4%) 0.690
                    FilteredTerm      161.02      (2.7%)      161.57      (2.3%)    0.3% (  -4% -    5%) 0.668
      FilteredOr2Terms2StopWords      110.97      (1.3%)      111.36      (1.3%)    0.4% (  -2% -    3%) 0.399
                  FilteredPhrase       25.49      (3.3%)       25.59      (2.8%)    0.4% (  -5% -    6%) 0.693
                       OrHighMed      195.25      (2.5%)      196.56      (1.9%)    0.7% (  -3% -    5%) 0.336
                      OrHighHigh       52.92      (2.6%)       53.32      (2.2%)    0.7% (  -3% -    5%) 0.323
              Or2Terms2StopWords      165.87      (3.3%)      167.13      (3.3%)    0.8% (  -5% -    7%) 0.471
                        Or3Terms      177.06      (3.5%)      178.44      (3.5%)    0.8% (  -6% -    8%) 0.484
                     OrStopWords       34.25      (5.2%)       34.64      (5.2%)    1.1% (  -8% -   12%) 0.492
                        PKLookup      274.02      (1.4%)      277.21      (1.4%)    1.2% (  -1% -    4%) 0.009
                      OrHighRare      282.75      (6.6%)      286.27      (6.9%)    1.2% ( -11% -   15%) 0.559
             And2Terms2StopWords      167.40      (3.0%)      170.65      (3.6%)    1.9% (  -4% -    8%) 0.064
                          OrMany       19.64      (4.0%)       20.22      (3.7%)    2.9% (  -4% -   11%) 0.017
                      AndHighMed      131.41      (1.8%)      135.53      (1.8%)    3.1% (   0% -    6%) 0.000
                     AndHighHigh       44.99      (1.9%)       46.55      (2.2%)    3.5% (   0% -    7%) 0.000
            FilteredAndStopWords       50.36      (2.8%)       52.27      (1.9%)    3.8% (   0% -    8%) 0.000
     FilteredAnd2Terms2StopWords      202.89      (2.1%)      210.87      (1.5%)    3.9% (   0% -    7%) 0.000
                    AndStopWords       32.04      (4.8%)       33.33      (5.9%)    4.0% (  -6% -   15%) 0.018
                       And3Terms      178.11      (3.8%)      185.46      (4.2%)    4.1% (  -3% -   12%) 0.001
             FilteredAndHighHigh       65.05      (2.8%)       68.05      (1.9%)    4.6% (   0% -    9%) 0.000
              FilteredAndHighMed      132.31      (3.0%)      139.26      (3.0%)    5.2% (   0% -   11%) 0.000
               FilteredAnd3Terms      195.92      (1.8%)      209.14      (1.8%)    6.7% (   3% -   10%) 0.000

…28. (#14017) We currently use `SlowImpactsEnum` for terms whose `docFreq` is less than 128 because it's convenient as these terms don't have impacts anyway. But a recent slowdown on nightly benchmarks suggests that this contributes to making some hot calls more polymorphic than we'd like, so this PR moves such terms back to the regular impacts enums.

…28. (apache#14017) We currently use `SlowImpactsEnum` for terms whose `docFreq` is less than 128 because it's convenient as these terms don't have impacts anyway. But a recent slowdown on nightly benchmarks suggests that this contributes to making some hot calls more polymorphic than we'd like, so this PR moves such terms back to the regular impacts enums.

jpountz added this to the 10.1.0 milestone Nov 25, 2024

jpountz mentioned this pull request Nov 25, 2024

Make CombinedFieldQuery eligible for WAND/MAXSCORE. #13999

Merged

jpountz merged commit a5bf8a5 into apache:main Nov 25, 2024

jpountz deleted the stop_using_slow_impacts_enum_under_128_doc_freq branch November 25, 2024 14:48

jpountz mentioned this pull request Apr 7, 2025

Reduce the number of classes extending DocIdSetIterator #14450

Open

8 tasks

jpountz mentioned this pull request Apr 24, 2025

Provide better impacts for fields indexed with IndexOptions.DOCS #14511

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Stop using `SlowImpactsEnum` for terms whose `docFreq` is less than 128.#14017

Stop using `SlowImpactsEnum` for terms whose `docFreq` is less than 128.#14017
jpountz merged 1 commit intoapache:mainfrom
jpountz:stop_using_slow_impacts_enum_under_128_doc_freq

jpountz commented Nov 25, 2024

Uh oh!

jpountz commented Nov 25, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

jpountz commented Nov 25, 2024

Uh oh!

jpountz commented Nov 25, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant