zacharymorn commented on PR #972:
URL: https://github.com/apache/lucene/pull/972#issuecomment-1163861983
Thanks @jpountz for looking into this! I did further experiments on this and
the result seems to suggest it may be caused by bug / caching in the util or
lucene itself.
What I did was I first only kept 1 query per pure disjunction task, and
remove the rest of the tasks like below.
```
OrHighHigh: several following # freq=436129 freq=416515
OrHighMed: international chris # freq=418261 freq=85523
OrHighLow: 2005 valois # freq=835460 freq=2277
```
and got this result:
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
PKLookup 189.27 (23.2%) 220.18
(16.1%) 16.3% ( -18% - 72%) 0.010
OrHighHigh 10.37 (41.8%) 20.92
(94.9%) 101.8% ( -24% - 410%) 0.000
OrHighMed 21.43 (54.3%) 56.18
(138.5%) 162.2% ( -19% - 777%) 0.000
OrHighLow 138.14 (26.2%) 368.50
(91.9%) 166.7% ( 38% - 385%) 0.000
```
However, when I added back the rest of the tasks but still kept 1 query for
each of the three disjunction tasks, I got vastly different results:
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
OrHighHigh 48.31 (7.8%) 38.21
(5.6%) -20.9% ( -31% - -8%) 0.000
OrHighMed 152.23 (8.5%) 140.16
(8.5%) -7.9% ( -23% - 9%) 0.003
BrowseDayOfYearSSDVFacets 21.67 (12.4%) 20.84
(7.1%) -3.8% ( -20% - 17%) 0.231
MedPhrase 103.08 (5.3%) 102.43
(9.1%) -0.6% ( -14% - 14%) 0.790
TermDTSort 162.97 (10.6%) 162.10
(6.4%) -0.5% ( -15% - 18%) 0.847
HighSloppyPhrase 62.12 (7.3%) 61.99
(5.7%) -0.2% ( -12% - 13%) 0.921
OrHighNotMed 1216.05 (4.1%) 1216.03
(3.2%) -0.0% ( -7% - 7%) 0.999
HighTerm 2088.45 (4.2%) 2091.84
(3.5%) 0.2% ( -7% - 8%) 0.895
BrowseMonthSSDVFacets 23.34 (10.0%) 23.49
(11.0%) 0.6% ( -18% - 24%) 0.845
MedTerm 3189.76 (3.5%) 3215.19
(3.9%) 0.8% ( -6% - 8%) 0.497
AndHighHigh 59.64 (6.9%) 60.14
(6.3%) 0.8% ( -11% - 15%) 0.688
MedSpanNear 46.86 (5.9%) 47.26
(7.8%) 0.9% ( -12% - 15%) 0.692
HighTermTitleBDVSort 125.65 (7.5%) 126.83
(10.1%) 0.9% ( -15% - 20%) 0.738
LowPhrase 21.25 (4.2%) 21.62
(4.4%) 1.8% ( -6% - 10%) 0.194
BrowseRandomLabelSSDVFacets 15.38 (6.6%) 15.66
(10.1%) 1.8% ( -13% - 19%) 0.509
HighSpanNear 20.48 (5.6%) 20.86
(5.1%) 1.9% ( -8% - 13%) 0.270
HighTermDayOfYearSort 187.51 (7.6%) 191.07
(9.4%) 1.9% ( -14% - 20%) 0.482
OrHighNotLow 1505.18 (10.5%) 1535.43
(4.4%) 2.0% ( -11% - 18%) 0.431
LowSloppyPhrase 233.02 (6.9%) 237.95
(8.4%) 2.1% ( -12% - 18%) 0.383
MedIntervalsOrdered 18.37 (5.0%) 18.77
(5.3%) 2.2% ( -7% - 13%) 0.177
OrNotHighMed 1310.81 (4.0%) 1342.33
(5.1%) 2.4% ( -6% - 11%) 0.096
MedSloppyPhrase 40.20 (6.2%) 41.18
(5.7%) 2.4% ( -8% - 15%) 0.190
AndHighHighDayTaxoFacets 13.90 (5.6%) 14.25
(3.7%) 2.6% ( -6% - 12%) 0.090
AndHighMed 566.84 (5.8%) 582.07
(5.7%) 2.7% ( -8% - 15%) 0.138
OrNotHighHigh 1976.63 (5.0%) 2030.44
(3.5%) 2.7% ( -5% - 11%) 0.044
Fuzzy2 50.72 (10.3%) 52.17
(7.9%) 2.9% ( -13% - 23%) 0.325
MedTermDayTaxoFacets 79.53 (6.7%) 81.86
(7.5%) 2.9% ( -10% - 18%) 0.192
OrHighNotHigh 1169.61 (6.6%) 1204.25
(3.8%) 3.0% ( -6% - 14%) 0.080
HighPhrase 400.95 (5.5%) 413.11
(2.2%) 3.0% ( -4% - 11%) 0.022
OrHighMedDayTaxoFacets 13.10 (5.0%) 13.50
(5.6%) 3.1% ( -7% - 14%) 0.066
AndHighMedDayTaxoFacets 52.75 (6.4%) 54.42
(6.3%) 3.2% ( -8% - 16%) 0.115
OrNotHighLow 2842.51 (3.5%) 2935.45
(5.5%) 3.3% ( -5% - 12%) 0.025
LowTerm 3032.44 (3.7%) 3140.83
(3.0%) 3.6% ( -3% - 10%) 0.001
HighTermMonthSort 210.58 (11.4%) 218.63
(11.3%) 3.8% ( -16% - 29%) 0.286
Fuzzy1 135.95 (8.1%) 141.35
(9.6%) 4.0% ( -12% - 23%) 0.158
IntNRQ 365.48 (17.1%) 380.05
(12.1%) 4.0% ( -21% - 40%) 0.395
Prefix3 72.71 (11.4%) 75.69
(9.5%) 4.1% ( -15% - 28%) 0.217
Wildcard 333.86 (9.4%) 347.59
(10.5%) 4.1% ( -14% - 26%) 0.192
LowSpanNear 40.56 (4.1%) 42.49
(6.4%) 4.8% ( -5% - 15%) 0.005
Respell 56.68 (7.6%) 59.68
(5.5%) 5.3% ( -7% - 19%) 0.012
AndHighLow 1062.10 (6.2%) 1118.34
(5.9%) 5.3% ( -6% - 18%) 0.005
HighIntervalsOrdered 9.10 (4.5%) 9.64
(6.5%) 5.9% ( -4% - 17%) 0.001
LowIntervalsOrdered 15.02 (5.1%) 15.96
(5.2%) 6.2% ( -3% - 17%) 0.000
BrowseDateSSDVFacets 2.29 (6.5%) 2.44
(19.1%) 6.4% ( -18% - 34%) 0.158
OrHighLow 931.34 (4.9%) 995.65
(3.7%) 6.9% ( -1% - 16%) 0.000
PKLookup 244.93 (10.0%) 262.86
(11.4%) 7.3% ( -12% - 31%) 0.031
BrowseMonthTaxoFacets 23.30 (42.2%) 28.97
(50.4%) 24.3% ( -47% - 202%) 0.098
BrowseDateTaxoFacets 23.39 (44.9%) 29.54
(52.5%) 26.3% ( -49% - 224%) 0.089
BrowseDayOfYearTaxoFacets 23.38 (44.2%) 30.09
(54.9%) 28.7% ( -48% - 228%) 0.069
BrowseRandomLabelTaxoFacets 22.47 (55.6%) 29.49
(68.4%) 31.2% ( -59% - 349%) 0.113
```
I've attached the tasks file for reference here as well
[wikimedium.10M.nostopwords.tasks.txt](https://github.com/apache/lucene/files/8962486/wikimedium.10M.nostopwords.tasks.txt)
@mikemccand , do you have any suggestion where this discrepancy might be
coming from? I'll continue to run experiments as well to see if I can pinpoint
the issue.
--
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]
