apache
Specialize `filterCompetitiveHits` when have exact 2 clauses
Description
This PR propose to specialize function filterCompetitiveHits when we have exact 2 scorers, in order to reduce float calculation and potential function calls
Luceneutil result on wikimediumall with searchConcurrency=0, taskCountPerCat=5, taskRepeatCount=50 after 20 iterations
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
OrHighRare 116.59 (3.1%) 116.96 (2.7%) 0.3% ( -5% - 6%) 0.734
OrHighMed 87.75 (2.4%) 88.83 (2.6%) 1.2% ( -3% - 6%) 0.116
AndHighMed 67.91 (2.3%) 69.17 (2.2%) 1.9% ( -2% - 6%) 0.009
AndHighHigh 27.96 (1.4%) 28.63 (2.0%) 2.4% ( -1% - 5%) 0.000
OrHighHigh 26.16 (1.6%) 26.97 (1.6%) 3.1% ( 0% - 6%) 0.000
This PR does not have an entry in lucene/CHANGES.txt. Consider adding one. If the PR doesn't need a changelog entry, then add the skip-changelog label to it and you will stop receiving this reminder on future updates to the PR.
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
For what it's worth, the reason of this PR is that I find filterCompetitiveHits ocuppied about 13% of flamegraph on OrHighHigh query,
Also, filterCompetitiveHits calls MathUtil.sumUpperBound in a loop, seems repeatly calculate MathUtil.sumRelativeErrorBound(numValues), (numValues is constant within the loop), I tried to optimize this, but it shows no performance difference, maybe filterCompetitiveHits is no longer the bottleneck when numValues > 2
Nice speedup! I wonder if we can make it work in the general case with something like that (untested):
Thanks for your reply! These change totaly make sense to me. I've also tried to merge these two code paths, but got some terrifying output when running luceneutil :( (even after I explicitly set verifyCounts to False ). I run the suggested version of code and got similar output
Maybe there is something I'm not familiar with? Or should I just change the code according to your advice? :)
This PR does not have an entry in lucene/CHANGES.txt. Consider adding one. If the PR doesn't need a changelog entry, then add the skip-changelog label to it and you will stop receiving this reminder on future updates to the PR.
This suggests that the fined~2 query doesn't drop exactly the same hits by score, which in-turn means that my suggestion is not correct and isn't actually always the smallest float that meets the sumUpperBound(minRequiredScore + maxRemainingScore, numScorers) >= minCompetitiveScore condition, we should look into fixing it. :)
Got it, I'll dig into this, thanks for your explaination! It helps me a lot :)
I played a bit more with your change, and it looks like we could make things even further as a follow-up through vectorization by using (Int|Float)Vector#compress.
Thinking a bit more about this problem, finding the smallest float that meets the condition may not always be cheap. However, finding a good-enough float should be easy, and it's fine if the hit count changes.
This PR does not have an entry in lucene/CHANGES.txt. Consider adding one. If the PR doesn't need a changelog entry, then add the skip-changelog label to it and you will stop receiving this reminder on future updates to the PR.
This PR does not have an entry in lucene/CHANGES.txt. Consider adding one. If the PR doesn't need a changelog entry, then add the skip-changelog label to it and you will stop receiving this reminder on future updates to the PR.
This PR does not have an entry in lucene/CHANGES.txt. Consider adding one. If the PR doesn't need a changelog entry, then add the skip-changelog label to it and you will stop receiving this reminder on future updates to the PR.
Sorry for the late reply.
I've dug a little bit into this issue. There are two problems.
The first one is that docs NOT collected in baseline ARE collected under this patch, this is actually caused by the initial double to float cast, so I changed it from float to double
The second one is that docs collected in baseline are NOT collected under this patch, it turns out the original code actually let docs with score lower than minCompetitiveScore been collected, and this patch savely prunes it away.
Here is a example which can reproduce the two problems above:
package org.apache.lucene;
import org.apache.lucene.util.MathUtil;
public class Run {
public static void main(String[] args) {
{
// problem that doc collected in patch but not in baseline
System.out.println("problem that doc collected in patch but not in baseline");
float minCompetitiveScore = 3.5382755f;
double score = 2.201078414916992d;
double maxRemainingScore = 1.337196946144104d;
// false, means this can not be collected in baseline
System.out.println(
((float) MathUtil.sumUpperBound(score + maxRemainingScore, 2)) >= minCompetitiveScore);
{
float minRequiredScore = (float) (minCompetitiveScore - maxRemainingScore);
while ((float) MathUtil.sumUpperBound(minRequiredScore + maxRemainingScore, 2) > minCompetitiveScore) {
minRequiredScore = Math.nextDown(minRequiredScore);
}
// score=2.201078414916992d, minRequiredScore=2.2010784f
// true, means this will be collected in patch
System.out.println(score >= minRequiredScore);
}
{
double minRequiredScore = (minCompetitiveScore - maxRemainingScore);
while ((float) MathUtil.sumUpperBound(minRequiredScore + maxRemainingScore, 2) > minCompetitiveScore) {
minRequiredScore = Math.nextDown(minRequiredScore);
}
// false, score=2.201078414916992d, minRequiredScoreD=2.2010785341262817d
System.out.println(score >= minRequiredScore);
}
}
System.out.println();
{
// problem that doc collected in baseline but not in patch
System.out.println("problem that doc collected in baseline but not in patch");
float minCompetitiveScore = 7.638806f;
double score = 7.638805627822876d;
double maxRemainingScore = 0.0d;
int numScorers = 33;
double minRequiredScore = (minCompetitiveScore - maxRemainingScore);
while ((float) MathUtil.sumUpperBound(minRequiredScore + maxRemainingScore, numScorers) > minCompetitiveScore) {
minRequiredScore = Math.nextDown(minRequiredScore);
}
// false, means this can not be collected by current patch.
System.out.println(score >= minRequiredScore);
// true, means this can be collected in baseline
// MathUtil.sumUpperBound(score + maxRemainingScore, numScorers)=7.638805627822984d;
// (float) MathUtil.sumUpperBound(score + maxRemainingScore, numScorers)=7.638806f == minCompetitiveScore;
//
// The original double (before cast to float) is actually smaller than minCompetitiveScore
// which means we can actually prune this doc safely.
System.out.println(
((float) MathUtil.sumUpperBound(score + maxRemainingScore, numScorers)) >= minCompetitiveScore);
}
}
}
I did another quick one iteration full task luceneutils to verify the hit count, this time it still complains about the hit count, but all the diffs are one direction (patch less than basline), which I think is expected behavior:
And I run another luceneutil on wikimediumall with searchConcurrency=0, taskCountPerCat=5, taskRepeatCount=50 after 20 iterations:
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
OrHighRare 120.94 (2.7%) 122.37 (1.9%) 1.2% ( -3% - 5%) 0.109
OrHighMed 90.74 (3.6%) 92.40 (2.8%) 1.8% ( -4% - 8%) 0.073
AndHighMed 69.64 (2.8%) 71.32 (2.1%) 2.4% ( -2% - 7%) 0.002
AndHighHigh 28.49 (2.4%) 29.29 (2.1%) 2.8% ( -1% - 7%) 0.000
OrHighHigh 26.83 (2.5%) 27.60 (1.9%) 2.9% ( -1% - 7%) 0.000
The while loop containing call to Math.nextDown (which operate on a double rather than float now) doesn's seem to add much overhead. I locally added some simple print over the iteration count of the loop, they are almost entirely zero (for the queries above).
The while loop containing call to Math.nextDown (which operate on a double rather than float now) doesn's seem to add much overhead
I'll run another full task luceneutil to verify the performance impact tonight
I played a bit more with your change, and it looks like we could make things even further as a follow-up through vectorization by using
(Int|Float)Vector#compress.
Thanks for your advice! I think i understand what you mean, will try to find a proper way to utilize the vectorization these days.
luceneutil result under same setup with lastest code:
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
TermB1M 571.58 (4.4%) 560.31 (5.7%) -2.0% ( -11% - 8%) 0.219
TermB1M1P 571.38 (4.4%) 560.76 (5.6%) -1.9% ( -11% - 8%) 0.243
Term1M 570.90 (4.2%) 560.56 (5.8%) -1.8% ( -11% - 8%) 0.260
Term10K 570.64 (4.4%) 560.95 (5.9%) -1.7% ( -11% - 8%) 0.301
FilteredTerm 84.77 (4.2%) 83.39 (2.7%) -1.6% ( -8% - 5%) 0.146
Term 571.14 (4.3%) 561.95 (5.6%) -1.6% ( -10% - 8%) 0.305
Term100 569.96 (4.2%) 561.59 (6.0%) -1.5% ( -11% - 9%) 0.371
CountTerm 6833.17 (4.2%) 6736.75 (4.8%) -1.4% ( -10% - 7%) 0.323
FilteredOrHighMed 52.38 (3.5%) 51.76 (2.2%) -1.2% ( -6% - 4%) 0.200
OrHighRare 121.85 (3.7%) 120.51 (4.8%) -1.1% ( -9% - 7%) 0.421
IntSet 397.61 (3.6%) 393.40 (3.6%) -1.1% ( -7% - 6%) 0.349
DismaxTerm 624.07 (2.6%) 617.94 (2.6%) -1.0% ( -6% - 4%) 0.232
CombinedAndHighMed 27.95 (4.2%) 27.69 (2.6%) -0.9% ( -7% - 6%) 0.404
TermDTSort 184.68 (3.5%) 183.18 (2.3%) -0.8% ( -6% - 5%) 0.385
SloppyPhrase 1.42 (2.9%) 1.41 (4.5%) -0.8% ( -7% - 6%) 0.526
FilteredOrHighHigh 17.80 (1.6%) 17.67 (1.5%) -0.8% ( -3% - 2%) 0.124
CombinedOrHighMed 27.47 (5.5%) 27.26 (4.4%) -0.7% ( -10% - 9%) 0.641
TermMonthSort 2430.41 (3.5%) 2413.88 (3.0%) -0.7% ( -6% - 6%) 0.513
FilteredPhrase 12.52 (2.0%) 12.43 (1.8%) -0.7% ( -4% - 3%) 0.263
TermTitleSort 67.84 (4.2%) 67.41 (3.3%) -0.6% ( -7% - 7%) 0.591
FilteredOr2Terms2StopWords 65.16 (1.8%) 64.74 (1.3%) -0.6% ( -3% - 2%) 0.210
FilteredOr3Terms 58.20 (2.3%) 57.84 (1.8%) -0.6% ( -4% - 3%) 0.348
CountAndHighHigh 61.16 (1.6%) 60.79 (1.3%) -0.6% ( -3% - 2%) 0.194
CountFilteredOrHighMed 29.79 (1.7%) 29.62 (1.6%) -0.6% ( -3% - 2%) 0.286
TermDayOfYearSort 341.54 (2.0%) 339.65 (1.7%) -0.6% ( -4% - 3%) 0.343
Phrase 9.73 (2.9%) 9.68 (2.6%) -0.5% ( -5% - 5%) 0.548
Respell 43.85 (3.0%) 43.64 (1.9%) -0.5% ( -5% - 4%) 0.540
CountFilteredOrHighHigh 25.38 (1.6%) 25.27 (1.4%) -0.4% ( -3% - 2%) 0.357
DismaxOrHighMed 65.04 (2.3%) 64.80 (2.3%) -0.4% ( -4% - 4%) 0.610
CountFilteredIntNRQ 22.26 (1.7%) 22.18 (1.7%) -0.3% ( -3% - 3%) 0.545
Fuzzy1 49.96 (3.5%) 49.80 (2.4%) -0.3% ( -5% - 5%) 0.735
CombinedTerm 14.61 (3.7%) 14.57 (3.5%) -0.3% ( -7% - 7%) 0.785
IntNRQ 48.09 (1.6%) 47.95 (1.3%) -0.3% ( -3% - 2%) 0.501
FilteredIntNRQ 47.86 (1.6%) 47.71 (1.3%) -0.3% ( -3% - 2%) 0.515
CountOrHighMed 95.08 (2.7%) 94.83 (1.5%) -0.3% ( -4% - 4%) 0.706
Wildcard 58.21 (4.4%) 58.09 (3.8%) -0.2% ( -8% - 8%) 0.875
FilteredOrStopWords 10.96 (1.4%) 10.94 (1.9%) -0.2% ( -3% - 3%) 0.703
FilteredOrMany 5.04 (1.6%) 5.03 (1.3%) -0.2% ( -3% - 2%) 0.664
CountFilteredPhrase 11.59 (1.4%) 11.57 (1.4%) -0.2% ( -2% - 2%) 0.661
Fuzzy2 45.25 (3.3%) 45.17 (2.0%) -0.2% ( -5% - 5%) 0.851
CountFilteredOrMany 6.06 (1.2%) 6.05 (1.0%) -0.2% ( -2% - 2%) 0.668
CountAndHighMed 92.80 (2.4%) 92.69 (2.1%) -0.1% ( -4% - 4%) 0.870
CountOrHighHigh 63.38 (1.2%) 63.36 (1.2%) -0.0% ( -2% - 2%) 0.940
AndHighOrMedMed 18.16 (2.7%) 18.16 (2.4%) -0.0% ( -4% - 5%) 0.981
FilteredAndHighHigh 14.92 (2.3%) 14.94 (2.9%) 0.1% ( -5% - 5%) 0.918
Prefix3 98.99 (3.8%) 99.07 (3.5%) 0.1% ( -6% - 7%) 0.941
OrHighMed 88.50 (2.6%) 88.59 (2.1%) 0.1% ( -4% - 4%) 0.889
DismaxOrHighHigh 46.10 (1.7%) 46.16 (2.8%) 0.1% ( -4% - 4%) 0.867
FilteredAnd3Terms 131.08 (1.5%) 131.28 (1.5%) 0.2% ( -2% - 3%) 0.746
FilteredAndStopWords 11.78 (2.2%) 11.82 (2.8%) 0.4% ( -4% - 5%) 0.662
CombinedAndHighHigh 7.20 (2.1%) 7.24 (1.9%) 0.5% ( -3% - 4%) 0.415
FilteredPrefix3 92.06 (3.9%) 92.61 (3.2%) 0.6% ( -6% - 8%) 0.601
CountOrMany 6.37 (2.0%) 6.41 (1.9%) 0.7% ( -3% - 4%) 0.264
AndMedOrHighHigh 20.85 (1.6%) 21.01 (1.3%) 0.7% ( -2% - 3%) 0.106
CombinedOrHighHigh 7.04 (4.5%) 7.10 (4.9%) 1.0% ( -8% - 10%) 0.520
IntervalsOrdered 3.00 (3.7%) 3.03 (3.9%) 1.1% ( -6% - 8%) 0.372
CountPhrase 3.28 (4.3%) 3.32 (3.1%) 1.1% ( -6% - 8%) 0.367
AndHighMed 67.26 (2.4%) 68.09 (1.4%) 1.2% ( -2% - 5%) 0.049
SpanNear 3.07 (4.9%) 3.11 (3.4%) 1.2% ( -6% - 10%) 0.350
AndHighHigh 26.61 (2.8%) 26.94 (2.8%) 1.3% ( -4% - 7%) 0.156
FilteredAndHighMed 43.51 (1.8%) 44.08 (1.9%) 1.3% ( -2% - 5%) 0.027
OrHighHigh 25.58 (2.2%) 25.99 (2.3%) 1.6% ( -2% - 6%) 0.024
FilteredAnd2Terms2StopWords 73.91 (2.8%) 75.69 (1.4%) 2.4% ( -1% - 6%) 0.001
Or2Terms2StopWords 73.90 (3.3%) 76.12 (1.0%) 3.0% ( -1% - 7%) 0.000
And2Terms2StopWords 71.57 (3.1%) 73.79 (1.3%) 3.1% ( -1% - 7%) 0.000
And3Terms 85.24 (3.7%) 88.75 (1.3%) 4.1% ( 0% - 9%) 0.000
OrStopWords 9.89 (6.7%) 10.31 (5.6%) 4.2% ( -7% - 17%) 0.031
Or3Terms 77.31 (4.0%) 80.81 (1.5%) 4.5% ( -1% - 10%) 0.000
OrMany 5.53 (6.8%) 5.82 (3.0%) 5.1% ( -4% - 16%) 0.002
AndStopWords 9.17 (5.8%) 9.70 (3.5%) 5.7% ( -3% - 15%) 0.000
WARNING: cat=Fuzzy1: hit counts differ: 11543+ vs 11524+
WARNING: cat=Fuzzy2: hit counts differ: 12724+ vs 12716+
WARNING: cat=Or2Terms2StopWords: hit counts differ: 14986+ vs 14985+
FWIW, I was a little bit worried about the cost of Math.ulp (compared to Math.nextDown), and also not sure wheather we should calculate the Math.ulp first before the while loop, so I did some comparation based on luceneutil. Here is the result if it helps
final patch vs baseline
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
CountTerm 6770.90 (6.8%) 6583.92 (7.3%) -2.8% ( -15% - 12%) 0.218
FilteredOr2Terms2StopWords 65.75 (6.9%) 64.68 (6.3%) -1.6% ( -13% - 12%) 0.438
TermDTSort 188.30 (4.4%) 186.09 (4.5%) -1.2% ( -9% - 8%) 0.401
Term10K 577.54 (7.5%) 570.96 (7.5%) -1.1% ( -14% - 14%) 0.629
Term 577.73 (7.5%) 571.15 (7.4%) -1.1% ( -14% - 14%) 0.628
TermMonthSort 2442.28 (3.3%) 2414.48 (4.7%) -1.1% ( -8% - 7%) 0.372
CombinedAndHighMed 28.53 (5.7%) 28.21 (5.5%) -1.1% ( -11% - 10%) 0.524
TermB1M1P 578.27 (7.4%) 571.77 (7.5%) -1.1% ( -14% - 14%) 0.633
CombinedOrHighMed 27.94 (7.0%) 27.64 (6.6%) -1.1% ( -13% - 13%) 0.619
FilteredOrHighMed 53.15 (5.1%) 52.59 (4.9%) -1.1% ( -10% - 9%) 0.500
Term1M 577.12 (7.5%) 571.38 (7.4%) -1.0% ( -14% - 15%) 0.672
FilteredOr3Terms 58.49 (4.8%) 57.96 (4.5%) -0.9% ( -9% - 8%) 0.541
TermB1M 577.53 (7.3%) 572.52 (7.6%) -0.9% ( -14% - 15%) 0.712
IntSet 391.58 (4.5%) 388.40 (4.8%) -0.8% ( -9% - 8%) 0.582
FilteredAnd3Terms 129.69 (3.4%) 128.71 (2.6%) -0.8% ( -6% - 5%) 0.433
Term100 577.00 (7.4%) 572.67 (7.3%) -0.8% ( -14% - 15%) 0.748
TermTitleSort 69.16 (3.7%) 68.80 (4.9%) -0.5% ( -8% - 8%) 0.707
CombinedTerm 14.62 (4.9%) 14.55 (5.6%) -0.5% ( -10% - 10%) 0.770
CountOrHighMed 95.39 (2.2%) 94.93 (2.2%) -0.5% ( -4% - 3%) 0.488
Wildcard 58.16 (3.6%) 57.89 (3.5%) -0.5% ( -7% - 6%) 0.669
SpanNear 3.11 (4.0%) 3.10 (4.3%) -0.4% ( -8% - 8%) 0.750
FilteredPrefix3 91.64 (3.2%) 91.27 (3.2%) -0.4% ( -6% - 6%) 0.691
FilteredTerm 85.80 (5.3%) 85.48 (4.7%) -0.4% ( -9% - 10%) 0.813
IntervalsOrdered 2.98 (3.8%) 2.97 (4.1%) -0.4% ( -7% - 7%) 0.770
CountAndHighMed 92.52 (2.8%) 92.19 (2.6%) -0.3% ( -5% - 5%) 0.683
IntNRQ 48.55 (1.9%) 48.39 (1.7%) -0.3% ( -3% - 3%) 0.571
DismaxTerm 625.93 (6.1%) 624.14 (5.9%) -0.3% ( -11% - 12%) 0.880
FilteredOrHighHigh 17.88 (3.2%) 17.83 (3.3%) -0.3% ( -6% - 6%) 0.784
SloppyPhrase 1.45 (4.6%) 1.45 (4.0%) -0.3% ( -8% - 8%) 0.839
Prefix3 98.03 (3.5%) 97.78 (3.5%) -0.3% ( -7% - 6%) 0.816
CountOrHighHigh 63.25 (2.1%) 63.11 (2.5%) -0.2% ( -4% - 4%) 0.761
FilteredIntNRQ 48.33 (2.1%) 48.25 (1.7%) -0.2% ( -3% - 3%) 0.777
CountFilteredIntNRQ 22.23 (1.9%) 22.20 (1.7%) -0.2% ( -3% - 3%) 0.773
CountFilteredOrMany 6.02 (1.5%) 6.01 (1.8%) -0.2% ( -3% - 3%) 0.758
CountAndHighHigh 61.53 (1.3%) 61.45 (1.7%) -0.1% ( -3% - 2%) 0.783
AndMedOrHighHigh 20.70 (4.3%) 20.70 (4.1%) 0.0% ( -8% - 8%) 0.993
CountFilteredPhrase 11.57 (2.1%) 11.57 (2.1%) 0.0% ( -4% - 4%) 0.945
Phrase 9.82 (2.6%) 9.83 (2.3%) 0.1% ( -4% - 5%) 0.918
CountFilteredOrHighHigh 25.13 (1.2%) 25.16 (1.5%) 0.1% ( -2% - 2%) 0.829
FilteredOrMany 5.01 (2.9%) 5.01 (2.6%) 0.1% ( -5% - 5%) 0.916
CountFilteredOrHighMed 29.53 (1.3%) 29.57 (1.4%) 0.1% ( -2% - 2%) 0.766
AndHighOrMedMed 18.23 (3.9%) 18.26 (2.6%) 0.2% ( -6% - 6%) 0.882
DismaxOrHighMed 64.34 (8.7%) 64.46 (8.3%) 0.2% ( -15% - 18%) 0.943
OrHighRare 121.70 (4.4%) 121.95 (4.5%) 0.2% ( -8% - 9%) 0.882
Fuzzy1 49.44 (4.3%) 49.60 (4.0%) 0.3% ( -7% - 9%) 0.809
CountOrMany 6.36 (1.8%) 6.38 (2.5%) 0.3% ( -3% - 4%) 0.640
TermDayOfYearSort 343.74 (1.3%) 344.99 (1.6%) 0.4% ( -2% - 3%) 0.433
FilteredPhrase 12.55 (3.0%) 12.60 (2.1%) 0.4% ( -4% - 5%) 0.631
Fuzzy2 44.71 (2.8%) 44.88 (2.5%) 0.4% ( -4% - 5%) 0.637
CountPhrase 3.28 (4.6%) 3.30 (4.7%) 0.4% ( -8% - 10%) 0.770
OrHighMed 86.40 (13.5%) 86.84 (13.4%) 0.5% ( -23% - 31%) 0.904
CombinedAndHighHigh 7.17 (2.3%) 7.22 (2.2%) 0.6% ( -3% - 5%) 0.395
CombinedOrHighHigh 7.01 (4.7%) 7.05 (5.1%) 0.6% ( -8% - 10%) 0.683
FilteredAndStopWords 11.80 (3.8%) 11.89 (3.1%) 0.7% ( -5% - 7%) 0.510
Respell 43.59 (3.2%) 43.92 (1.4%) 0.8% ( -3% - 5%) 0.333
FilteredOrStopWords 10.92 (3.0%) 11.01 (2.2%) 0.8% ( -4% - 6%) 0.362
AndHighMed 66.88 (9.2%) 67.46 (9.1%) 0.9% ( -16% - 21%) 0.765
FilteredAndHighHigh 14.95 (3.5%) 15.10 (3.2%) 1.0% ( -5% - 7%) 0.353
DismaxOrHighHigh 45.06 (7.9%) 45.58 (7.6%) 1.2% ( -13% - 18%) 0.639
And2Terms2StopWords 71.10 (10.3%) 72.31 (10.3%) 1.7% ( -17% - 24%) 0.604
FilteredAnd2Terms2StopWords 72.77 (7.7%) 74.03 (7.5%) 1.7% ( -12% - 18%) 0.474
AndHighHigh 26.25 (10.1%) 26.71 (10.0%) 1.7% ( -16% - 24%) 0.581
FilteredAndHighMed 43.31 (4.5%) 44.07 (4.1%) 1.8% ( -6% - 10%) 0.199
OrHighHigh 24.55 (12.9%) 25.19 (13.2%) 2.6% ( -20% - 32%) 0.527
Or2Terms2StopWords 71.62 (12.1%) 73.89 (11.8%) 3.2% ( -18% - 30%) 0.403
And3Terms 83.39 (8.5%) 87.53 (7.7%) 5.0% ( -10% - 23%) 0.053
OrMany 5.42 (7.8%) 5.73 (5.0%) 5.6% ( -6% - 19%) 0.006
Or3Terms 73.96 (11.5%) 78.67 (11.4%) 6.4% ( -14% - 33%) 0.078
OrStopWords 9.31 (11.6%) 10.08 (12.2%) 8.3% ( -13% - 36%) 0.027
AndStopWords 8.85 (9.2%) 9.67 (8.4%) 9.2% ( -7% - 29%) 0.001
WARNING: cat=Fuzzy1: hit counts differ: 11543+ vs 11524+
WARNING: cat=Fuzzy2: hit counts differ: 12724+ vs 12716+
WARNING: cat=Or2Terms2StopWords: hit counts differ: 14986+ vs 14985+
cached Math.ulp (baseline) vs Math.nextDown (candidate)
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
OrHighRare 122.65 (4.9%) 120.37 (4.8%) -1.9% ( -11% - 8%) 0.229
CombinedTerm 14.67 (4.5%) 14.55 (3.9%) -0.8% ( -8% - 7%) 0.527
FilteredPhrase 12.73 (1.8%) 12.63 (2.0%) -0.8% ( -4% - 3%) 0.205
FilteredPrefix3 92.09 (2.2%) 91.41 (2.8%) -0.7% ( -5% - 4%) 0.356
CombinedAndHighHigh 7.26 (1.8%) 7.22 (2.8%) -0.6% ( -5% - 4%) 0.449
CountFilteredIntNRQ 22.37 (2.4%) 22.25 (2.2%) -0.5% ( -4% - 4%) 0.479
AndHighOrMedMed 18.60 (2.7%) 18.52 (3.3%) -0.4% ( -6% - 5%) 0.644
CountFilteredOrHighMed 29.92 (2.0%) 29.80 (1.7%) -0.4% ( -3% - 3%) 0.488
DismaxOrHighHigh 46.43 (5.9%) 46.27 (5.7%) -0.4% ( -11% - 11%) 0.843
CountFilteredOrHighHigh 25.45 (1.8%) 25.37 (1.6%) -0.3% ( -3% - 3%) 0.549
FilteredIntNRQ 48.27 (2.1%) 48.12 (2.4%) -0.3% ( -4% - 4%) 0.661
Respell 43.46 (2.1%) 43.33 (2.0%) -0.3% ( -4% - 3%) 0.646
SloppyPhrase 1.44 (4.6%) 1.44 (5.0%) -0.3% ( -9% - 9%) 0.847
CombinedOrHighHigh 7.08 (4.5%) 7.06 (4.6%) -0.3% ( -9% - 9%) 0.842
AndHighHigh 26.93 (9.2%) 26.85 (9.0%) -0.3% ( -16% - 19%) 0.922
CountOrMany 6.41 (2.9%) 6.40 (2.6%) -0.3% ( -5% - 5%) 0.768
Prefix3 98.30 (2.8%) 98.07 (3.7%) -0.2% ( -6% - 6%) 0.827
IntSet 398.07 (5.0%) 397.31 (4.7%) -0.2% ( -9% - 9%) 0.901
FilteredTerm 89.51 (4.3%) 89.37 (4.4%) -0.2% ( -8% - 9%) 0.908
TermTitleSort 69.24 (3.9%) 69.16 (4.0%) -0.1% ( -7% - 8%) 0.928
TermDayOfYearSort 343.02 (1.6%) 342.64 (1.0%) -0.1% ( -2% - 2%) 0.796
And3Terms 89.96 (6.1%) 89.89 (6.1%) -0.1% ( -11% - 12%) 0.964
Wildcard 58.53 (2.6%) 58.48 (3.5%) -0.1% ( -5% - 6%) 0.936
IntNRQ 48.59 (2.2%) 48.56 (2.0%) -0.1% ( -4% - 4%) 0.918
CountAndHighHigh 61.94 (1.5%) 61.90 (1.6%) -0.1% ( -3% - 3%) 0.911
OrHighHigh 25.88 (9.1%) 25.87 (9.0%) -0.0% ( -16% - 19%) 0.996
FilteredAndHighHigh 15.19 (2.3%) 15.19 (2.5%) 0.0% ( -4% - 4%) 0.995
DismaxTerm 642.57 (5.7%) 642.60 (5.7%) 0.0% ( -10% - 12%) 0.997
FilteredOrStopWords 11.18 (1.8%) 11.19 (2.6%) 0.0% ( -4% - 4%) 0.978
AndMedOrHighHigh 21.04 (3.5%) 21.04 (3.5%) 0.0% ( -6% - 7%) 0.985
FilteredAndStopWords 11.98 (2.5%) 11.98 (2.5%) 0.0% ( -4% - 5%) 0.970
OrHighMed 91.32 (8.8%) 91.36 (8.5%) 0.0% ( -15% - 19%) 0.987
CountFilteredOrMany 6.07 (1.8%) 6.07 (1.9%) 0.1% ( -3% - 3%) 0.922
AndStopWords 9.63 (6.2%) 9.63 (6.3%) 0.1% ( -11% - 13%) 0.976
Term1M 587.59 (7.6%) 588.02 (6.8%) 0.1% ( -13% - 15%) 0.975
CombinedAndHighMed 29.45 (4.2%) 29.47 (4.5%) 0.1% ( -8% - 9%) 0.956
IntervalsOrdered 2.93 (3.3%) 2.94 (2.4%) 0.1% ( -5% - 5%) 0.915
CountFilteredPhrase 11.49 (2.1%) 11.50 (1.8%) 0.1% ( -3% - 4%) 0.865
CountTerm 6878.31 (5.8%) 6886.78 (5.5%) 0.1% ( -10% - 12%) 0.945
OrStopWords 10.27 (7.9%) 10.28 (8.3%) 0.1% ( -14% - 17%) 0.960
DismaxOrHighMed 66.86 (5.8%) 66.96 (5.6%) 0.1% ( -10% - 12%) 0.937
AndHighMed 68.99 (8.0%) 69.09 (8.1%) 0.1% ( -14% - 17%) 0.955
CountOrHighHigh 63.98 (2.0%) 64.08 (2.1%) 0.2% ( -3% - 4%) 0.806
Or3Terms 81.49 (6.9%) 81.63 (6.8%) 0.2% ( -12% - 14%) 0.938
Term 586.86 (7.7%) 587.87 (7.0%) 0.2% ( -13% - 16%) 0.941
Phrase 9.85 (2.3%) 9.87 (3.4%) 0.2% ( -5% - 5%) 0.843
Term100 587.22 (7.6%) 588.35 (7.2%) 0.2% ( -13% - 16%) 0.934
TermB1M1P 586.68 (7.7%) 588.21 (7.0%) 0.3% ( -13% - 16%) 0.910
CombinedOrHighMed 28.78 (5.6%) 28.87 (5.1%) 0.3% ( -9% - 11%) 0.862
FilteredAndHighMed 44.81 (3.2%) 44.97 (3.1%) 0.4% ( -5% - 6%) 0.715
TermDTSort 193.95 (4.4%) 194.66 (4.2%) 0.4% ( -7% - 9%) 0.787
FilteredOrHighHigh 18.21 (2.6%) 18.29 (2.9%) 0.5% ( -4% - 6%) 0.593
TermB1M 586.34 (7.6%) 589.19 (7.0%) 0.5% ( -13% - 16%) 0.834
Fuzzy2 45.79 (2.2%) 46.01 (2.6%) 0.5% ( -4% - 5%) 0.510
SpanNear 3.04 (5.3%) 3.06 (5.1%) 0.5% ( -9% - 11%) 0.757
CountAndHighMed 92.07 (3.7%) 92.61 (3.4%) 0.6% ( -6% - 8%) 0.605
Fuzzy1 51.48 (3.5%) 51.78 (3.8%) 0.6% ( -6% - 8%) 0.610
Or2Terms2StopWords 79.27 (8.1%) 79.76 (7.7%) 0.6% ( -14% - 17%) 0.808
FilteredAnd2Terms2StopWords 77.97 (6.1%) 78.46 (5.8%) 0.6% ( -10% - 13%) 0.739
FilteredAnd3Terms 131.77 (2.9%) 132.61 (2.6%) 0.6% ( -4% - 6%) 0.464
And2Terms2StopWords 77.57 (8.4%) 78.06 (8.0%) 0.6% ( -14% - 18%) 0.806
FilteredOrHighMed 54.67 (4.1%) 55.02 (4.2%) 0.6% ( -7% - 9%) 0.622
Term10K 585.60 (7.6%) 589.42 (7.0%) 0.7% ( -12% - 16%) 0.777
FilteredOrMany 5.11 (2.6%) 5.14 (2.3%) 0.7% ( -4% - 5%) 0.370
TermMonthSort 2472.84 (3.7%) 2490.38 (4.5%) 0.7% ( -7% - 9%) 0.584
CountOrHighMed 95.51 (3.2%) 96.23 (2.9%) 0.8% ( -5% - 6%) 0.425
OrMany 5.77 (5.3%) 5.81 (4.9%) 0.8% ( -8% - 11%) 0.616
CountPhrase 3.24 (5.2%) 3.26 (4.2%) 0.8% ( -8% - 10%) 0.581
FilteredOr3Terms 60.14 (4.0%) 60.72 (3.8%) 1.0% ( -6% - 9%) 0.431
FilteredOr2Terms2StopWords 68.59 (5.7%) 69.35 (5.7%) 1.1% ( -9% - 13%) 0.537
cached Math.ulp (baseline) vs non-cached (candidate)
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
AndHighHigh 27.11 (3.5%) 26.72 (9.8%) -1.4% ( -14% - 12%) 0.538
SloppyPhrase 1.44 (4.0%) 1.43 (3.8%) -1.3% ( -8% - 6%) 0.310
CountFilteredPhrase 11.66 (2.0%) 11.54 (1.7%) -1.0% ( -4% - 2%) 0.079
CountPhrase 3.34 (1.7%) 3.31 (3.1%) -1.0% ( -5% - 3%) 0.226
AndStopWords 9.69 (4.3%) 9.61 (7.1%) -0.9% ( -11% - 11%) 0.647
AndMedOrHighHigh 20.97 (2.5%) 20.83 (4.6%) -0.7% ( -7% - 6%) 0.541
AndHighMed 68.38 (3.3%) 68.08 (9.3%) -0.4% ( -12% - 12%) 0.840
Phrase 9.72 (2.4%) 9.68 (2.1%) -0.4% ( -4% - 4%) 0.588
TermDayOfYearSort 343.57 (1.6%) 342.85 (1.8%) -0.2% ( -3% - 3%) 0.704
FilteredIntNRQ 48.19 (1.2%) 48.12 (1.8%) -0.1% ( -3% - 2%) 0.756
IntNRQ 48.40 (1.3%) 48.35 (1.8%) -0.1% ( -3% - 2%) 0.816
Respell 44.00 (2.4%) 43.97 (2.1%) -0.1% ( -4% - 4%) 0.907
CountFilteredOrMany 6.00 (2.1%) 5.99 (1.8%) -0.1% ( -3% - 3%) 0.923
SpanNear 3.14 (3.6%) 3.14 (4.0%) 0.0% ( -7% - 7%) 0.998
CombinedTerm 14.70 (3.6%) 14.70 (3.1%) 0.0% ( -6% - 6%) 0.995
CountOrHighHigh 62.87 (2.5%) 62.94 (2.4%) 0.1% ( -4% - 5%) 0.889
CombinedAndHighHigh 7.27 (1.7%) 7.28 (2.2%) 0.1% ( -3% - 4%) 0.858
And3Terms 88.51 (3.6%) 88.62 (7.9%) 0.1% ( -11% - 12%) 0.950
IntervalsOrdered 2.96 (3.1%) 2.97 (3.2%) 0.1% ( -5% - 6%) 0.888
OrStopWords 10.20 (9.6%) 10.22 (9.3%) 0.2% ( -17% - 21%) 0.956
OrMany 5.82 (3.4%) 5.83 (5.0%) 0.2% ( -7% - 8%) 0.899
CountAndHighHigh 61.27 (1.7%) 61.37 (1.6%) 0.2% ( -3% - 3%) 0.742
CountOrMany 6.33 (3.3%) 6.34 (2.3%) 0.2% ( -5% - 6%) 0.838
CountOrHighMed 95.26 (2.5%) 95.50 (2.2%) 0.2% ( -4% - 5%) 0.737
CountFilteredIntNRQ 22.12 (1.3%) 22.21 (1.6%) 0.4% ( -2% - 3%) 0.410
FilteredPhrase 12.44 (2.4%) 12.49 (2.0%) 0.4% ( -3% - 4%) 0.561
CountAndHighMed 92.69 (2.8%) 93.06 (2.4%) 0.4% ( -4% - 5%) 0.622
CountFilteredOrHighHigh 25.01 (1.8%) 25.13 (1.7%) 0.5% ( -3% - 4%) 0.409
CombinedOrHighHigh 7.18 (4.0%) 7.21 (3.4%) 0.5% ( -6% - 8%) 0.643
AndHighOrMedMed 18.40 (2.5%) 18.50 (3.1%) 0.5% ( -4% - 6%) 0.543
CountFilteredOrHighMed 29.41 (1.8%) 29.57 (2.2%) 0.6% ( -3% - 4%) 0.382
FilteredAnd3Terms 129.79 (2.6%) 130.56 (3.2%) 0.6% ( -5% - 6%) 0.526
OrHighHigh 25.48 (9.7%) 25.65 (9.7%) 0.6% ( -17% - 22%) 0.837
Term10K 571.75 (4.7%) 575.74 (8.0%) 0.7% ( -11% - 14%) 0.735
FilteredOrStopWords 11.01 (2.2%) 11.09 (2.5%) 0.8% ( -3% - 5%) 0.283
DismaxOrHighHigh 45.46 (6.0%) 45.83 (6.4%) 0.8% ( -10% - 14%) 0.681
FilteredAndHighMed 44.08 (3.0%) 44.45 (4.4%) 0.8% ( -6% - 8%) 0.492
TermMonthSort 2383.55 (3.6%) 2403.84 (5.0%) 0.9% ( -7% - 9%) 0.535
Term1M 571.35 (4.8%) 576.33 (7.9%) 0.9% ( -11% - 14%) 0.673
Term 572.59 (4.8%) 577.59 (8.1%) 0.9% ( -11% - 14%) 0.679
TermB1M 571.63 (4.6%) 577.08 (8.1%) 1.0% ( -11% - 14%) 0.646
Term100 571.33 (4.8%) 576.88 (8.1%) 1.0% ( -11% - 14%) 0.646
IntSet 386.36 (3.7%) 390.32 (4.2%) 1.0% ( -6% - 9%) 0.413
FilteredAndStopWords 11.79 (3.4%) 11.91 (3.3%) 1.1% ( -5% - 8%) 0.323
OrHighRare 117.89 (4.1%) 119.16 (4.5%) 1.1% ( -7% - 10%) 0.433
DismaxOrHighMed 64.77 (6.9%) 65.47 (7.1%) 1.1% ( -12% - 16%) 0.625
FilteredAndHighHigh 14.93 (3.1%) 15.10 (3.2%) 1.1% ( -5% - 7%) 0.264
TermB1M1P 570.76 (4.9%) 577.30 (8.0%) 1.1% ( -11% - 14%) 0.586
FilteredOrMany 5.04 (2.3%) 5.10 (2.4%) 1.2% ( -3% - 6%) 0.106
DismaxTerm 620.66 (4.3%) 628.51 (5.4%) 1.3% ( -8% - 11%) 0.412
FilteredOrHighHigh 17.78 (3.2%) 18.00 (3.1%) 1.3% ( -4% - 7%) 0.201
TermTitleSort 69.48 (5.1%) 70.40 (4.4%) 1.3% ( -7% - 11%) 0.380
Or3Terms 79.24 (8.1%) 80.29 (8.3%) 1.3% ( -13% - 19%) 0.608
Fuzzy2 45.38 (2.2%) 46.00 (2.8%) 1.4% ( -3% - 6%) 0.081
OrHighMed 87.94 (10.1%) 89.31 (10.4%) 1.6% ( -17% - 24%) 0.629
TermDTSort 186.32 (4.1%) 189.37 (3.9%) 1.6% ( -6% - 10%) 0.199
CountTerm 6527.34 (4.5%) 6639.20 (6.2%) 1.7% ( -8% - 12%) 0.315
Prefix3 98.71 (3.3%) 100.50 (3.8%) 1.8% ( -5% - 9%) 0.106
CombinedAndHighMed 28.34 (5.2%) 28.91 (5.5%) 2.0% ( -8% - 13%) 0.232
FilteredAnd2Terms2StopWords 74.35 (6.4%) 75.85 (7.7%) 2.0% ( -11% - 17%) 0.365
Wildcard 57.69 (3.8%) 58.85 (4.4%) 2.0% ( -5% - 10%) 0.117
FilteredOr3Terms 57.92 (4.6%) 59.17 (4.7%) 2.2% ( -6% - 12%) 0.143
Fuzzy1 50.18 (3.4%) 51.31 (4.5%) 2.3% ( -5% - 10%) 0.074
FilteredOrHighMed 52.48 (5.1%) 53.68 (4.8%) 2.3% ( -7% - 12%) 0.147
FilteredPrefix3 91.96 (3.0%) 94.13 (3.8%) 2.4% ( -4% - 9%) 0.030
FilteredTerm 85.14 (5.1%) 87.42 (5.7%) 2.7% ( -7% - 14%) 0.118
CombinedOrHighMed 27.90 (6.1%) 28.65 (5.7%) 2.7% ( -8% - 15%) 0.150
Or2Terms2StopWords 74.51 (9.6%) 76.63 (10.3%) 2.8% ( -15% - 25%) 0.366
And2Terms2StopWords 72.82 (8.2%) 74.91 (10.8%) 2.9% ( -14% - 23%) 0.343
FilteredOr2Terms2StopWords 64.55 (6.7%) 66.61 (6.9%) 3.2% ( -9% - 17%) 0.139
The result shows there is no significant difference between these subtle changes, so I think it's fine to just leave it as is. BTW, I repeat the new UT 100000 times, all passed
The change from ulp(double) to ulp(float) seems to help decrease the p-value compared to previous benchmark result, it's great!
luceneutil
TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value
IntNRQ 48.58 (1.4%) 48.14 (2.0%) -0.9% ( -4% - 2%) 0.090
Prefix3 98.75 (3.5%) 97.99 (4.3%) -0.8% ( -8% - 7%) 0.534
CountOrMany 6.41 (2.2%) 6.36 (2.6%) -0.7% ( -5% - 4%) 0.328
TermTitleSort 70.07 (5.0%) 69.57 (5.6%) -0.7% ( -10% - 10%) 0.669
CountPhrase 3.28 (3.2%) 3.27 (3.5%) -0.5% ( -6% - 6%) 0.616
FilteredIntNRQ 48.13 (1.6%) 47.88 (2.0%) -0.5% ( -4% - 3%) 0.363
FilteredPrefix3 92.07 (3.5%) 91.63 (4.0%) -0.5% ( -7% - 7%) 0.685
CountTerm 6724.13 (5.8%) 6693.36 (6.7%) -0.5% ( -12% - 12%) 0.817
CountFilteredOrMany 6.08 (1.6%) 6.05 (2.3%) -0.4% ( -4% - 3%) 0.548
FilteredPhrase 12.62 (1.4%) 12.58 (2.1%) -0.3% ( -3% - 3%) 0.554
CountOrHighHigh 63.73 (2.0%) 63.52 (2.5%) -0.3% ( -4% - 4%) 0.647
CountFilteredOrHighHigh 25.24 (1.3%) 25.18 (1.8%) -0.3% ( -3% - 2%) 0.610
CountAndHighHigh 61.82 (1.5%) 61.66 (1.7%) -0.3% ( -3% - 2%) 0.607
CountFilteredIntNRQ 22.20 (2.1%) 22.15 (2.2%) -0.2% ( -4% - 4%) 0.728
CountFilteredOrHighMed 29.69 (1.4%) 29.62 (1.8%) -0.2% ( -3% - 3%) 0.647
FilteredTerm 87.59 (5.9%) 87.40 (6.6%) -0.2% ( -11% - 13%) 0.912
Respell 43.52 (2.0%) 43.45 (2.4%) -0.2% ( -4% - 4%) 0.816
TermDayOfYearSort 344.16 (1.8%) 343.61 (1.2%) -0.2% ( -3% - 2%) 0.745
CountFilteredPhrase 11.50 (2.4%) 11.49 (2.1%) -0.1% ( -4% - 4%) 0.866
CountOrHighMed 96.68 (2.5%) 96.57 (3.3%) -0.1% ( -5% - 5%) 0.899
FilteredAndStopWords 11.98 (4.2%) 11.97 (4.2%) -0.1% ( -8% - 8%) 0.954
FilteredOrHighHigh 18.06 (3.3%) 18.05 (3.7%) -0.1% ( -6% - 7%) 0.962
FilteredAndHighHigh 15.22 (4.0%) 15.22 (3.9%) -0.0% ( -7% - 8%) 0.971
FilteredOrHighMed 54.08 (4.6%) 54.06 (5.6%) -0.0% ( -9% - 10%) 0.982
SloppyPhrase 1.44 (4.8%) 1.44 (3.9%) 0.0% ( -8% - 9%) 0.982
CountAndHighMed 93.74 (1.9%) 93.80 (3.4%) 0.1% ( -5% - 5%) 0.943
Wildcard 58.08 (3.5%) 58.13 (3.6%) 0.1% ( -6% - 7%) 0.937
Phrase 9.83 (2.0%) 9.85 (2.3%) 0.1% ( -4% - 4%) 0.839
DismaxTerm 634.49 (4.5%) 635.53 (6.3%) 0.2% ( -10% - 11%) 0.924
FilteredOr3Terms 59.59 (4.2%) 59.72 (5.2%) 0.2% ( -8% - 10%) 0.884
Fuzzy2 45.18 (3.2%) 45.29 (3.2%) 0.3% ( -6% - 6%) 0.802
FilteredOr2Terms2StopWords 67.27 (6.2%) 67.45 (7.2%) 0.3% ( -12% - 14%) 0.901
FilteredAnd3Terms 131.54 (2.7%) 131.95 (3.3%) 0.3% ( -5% - 6%) 0.747
CombinedOrHighMed 28.71 (5.4%) 28.81 (5.7%) 0.4% ( -10% - 12%) 0.837
TermMonthSort 2460.88 (4.2%) 2470.76 (5.0%) 0.4% ( -8% - 10%) 0.784
AndMedOrHighHigh 20.92 (1.5%) 21.02 (2.1%) 0.5% ( -3% - 4%) 0.412
CombinedAndHighMed 29.00 (5.5%) 29.14 (5.6%) 0.5% ( -10% - 12%) 0.776
TermDTSort 190.27 (4.4%) 191.39 (5.1%) 0.6% ( -8% - 10%) 0.695
Term10K 584.68 (5.0%) 588.32 (6.6%) 0.6% ( -10% - 12%) 0.738
DismaxOrHighMed 66.04 (4.2%) 66.49 (4.3%) 0.7% ( -7% - 9%) 0.621
FilteredOrMany 5.09 (2.0%) 5.12 (3.0%) 0.7% ( -4% - 5%) 0.402
Fuzzy1 50.40 (4.2%) 50.76 (4.7%) 0.7% ( -7% - 10%) 0.613
Term1M 584.67 (4.9%) 589.05 (6.9%) 0.7% ( -10% - 13%) 0.694
CombinedTerm 14.72 (4.4%) 14.84 (4.6%) 0.8% ( -7% - 10%) 0.580
AndHighOrMedMed 18.29 (3.5%) 18.44 (2.2%) 0.8% ( -4% - 6%) 0.381
TermB1M1P 583.58 (5.0%) 588.67 (6.7%) 0.9% ( -10% - 13%) 0.643
FilteredOrStopWords 10.94 (2.8%) 11.04 (3.3%) 0.9% ( -5% - 7%) 0.347
CombinedAndHighHigh 7.19 (2.0%) 7.26 (2.9%) 0.9% ( -3% - 5%) 0.234
Term100 582.93 (5.0%) 588.80 (6.8%) 1.0% ( -10% - 13%) 0.594
TermB1M 583.60 (4.8%) 589.94 (6.6%) 1.1% ( -9% - 13%) 0.552
CombinedOrHighHigh 7.10 (3.1%) 7.17 (4.1%) 1.1% ( -5% - 8%) 0.341
IntSet 384.28 (3.5%) 388.53 (4.8%) 1.1% ( -6% - 9%) 0.404
IntervalsOrdered 2.93 (3.7%) 2.96 (3.4%) 1.1% ( -5% - 8%) 0.309
SpanNear 3.06 (4.8%) 3.09 (3.9%) 1.2% ( -7% - 10%) 0.397
Term 582.79 (5.0%) 589.71 (6.8%) 1.2% ( -10% - 13%) 0.528
DismaxOrHighHigh 45.65 (3.4%) 46.21 (3.1%) 1.2% ( -5% - 8%) 0.235
OrHighRare 117.32 (6.4%) 118.79 (7.2%) 1.2% ( -11% - 15%) 0.562
OrHighMed 90.35 (4.4%) 91.60 (4.7%) 1.4% ( -7% - 10%) 0.337
AndHighMed 68.62 (3.3%) 69.82 (3.9%) 1.7% ( -5% - 9%) 0.126
FilteredAndHighMed 44.21 (3.8%) 44.99 (3.5%) 1.8% ( -5% - 9%) 0.123
AndHighHigh 26.53 (2.8%) 27.08 (2.4%) 2.1% ( -3% - 7%) 0.012
OrHighHigh 25.43 (2.8%) 26.16 (2.4%) 2.8% ( -2% - 8%) 0.001
FilteredAnd2Terms2StopWords 74.67 (6.7%) 77.28 (6.8%) 3.5% ( -9% - 18%) 0.101
And2Terms2StopWords 73.37 (8.9%) 76.53 (9.4%) 4.3% ( -12% - 24%) 0.137
Or2Terms2StopWords 75.05 (8.1%) 78.52 (8.3%) 4.6% ( -10% - 22%) 0.075
And3Terms 85.24 (6.1%) 90.46 (4.3%) 6.1% ( -3% - 17%) 0.000
Or3Terms 77.07 (6.1%) 82.11 (3.9%) 6.5% ( -3% - 17%) 0.000
OrMany 5.44 (7.2%) 5.84 (3.9%) 7.3% ( -3% - 19%) 0.000
OrStopWords 9.49 (7.8%) 10.21 (4.0%) 7.6% ( -3% - 21%) 0.000
AndStopWords 8.90 (7.7%) 9.64 (3.0%) 8.4% ( -2% - 20%) 0.000
Thank you, we may want to look into making this better in follow-ups, but this looks good enough for me so I merged. :+1:
The seed did not reproduce for me, but I think I understand the problem. The code assumes that if a + b > c then a - ε + b <= c (ε > 0). However this is not true with floating-point arithmetics, it may be that a + b and a - ε + b round to the same value (even though a - ε doesn't round to a). So the code assumed that docs had to have a score of at least a, but that would mistakenly filter out docs whose score was a - ε.
I pushed a fix.
I worked on improving the ScorerUtil test so that it would catch this problem. It helped me find another problem. I pushed directly. I think we're good now.
Nightly benchmarks confirmed the speedup: https://benchmarks.mikemccandless.com/OrStopWords.html