[
https://issues.apache.org/jira/browse/LUCENE-10062?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17404882#comment-17404882
]
Greg Miller commented on LUCENE-10062:
--------------------------------------
The performance improvement, as measured by {{luceneutil}} benchmarks, is
borderline unbelievable by moving to numeric doc values (instead of the custom
binary encoded values). It feels too good to be true, but all tests pass and I
pulled the change into our internal fork and ran all of our tests and
correctness suites, which also all pass.
*I'm seeing almost 400% QPS improvement on the three taxonomy browsing tasks
with this change*.
The following results are using {{wikimediumall}}:
{noformat}
TaskQPS baseline StdDevQPS candidate StdDev
Pct diff p-value
OrHighNotMed 638.52 (5.3%) 602.16 (8.5%)
-5.7% ( -18% - 8%) 0.011
OrHighNotLow 609.27 (4.2%) 588.95 (5.8%)
-3.3% ( -12% - 6%) 0.036
PKLookup 136.76 (3.4%) 133.32 (3.3%)
-2.5% ( -8% - 4%) 0.018
OrHighNotHigh 535.46 (4.6%) 523.63 (5.7%)
-2.2% ( -12% - 8%) 0.181
OrNotHighMed 516.79 (5.6%) 507.71 (6.7%)
-1.8% ( -13% - 11%) 0.367
OrNotHighLow 543.98 (4.6%) 535.62 (6.8%)
-1.5% ( -12% - 10%) 0.403
OrHighLow 222.57 (2.7%) 219.42 (3.8%)
-1.4% ( -7% - 5%) 0.171
Prefix3 52.18 (6.1%) 51.50 (6.0%)
-1.3% ( -12% - 11%) 0.499
Fuzzy1 49.69 (3.3%) 49.12 (4.1%)
-1.1% ( -8% - 6%) 0.340
Wildcard 23.73 (4.1%) 23.53 (4.0%)
-0.8% ( -8% - 7%) 0.512
OrNotHighHigh 471.12 (3.7%) 467.29 (5.6%)
-0.8% ( -9% - 8%) 0.589
HighSloppyPhrase 4.67 (4.8%) 4.63 (5.5%)
-0.8% ( -10% - 10%) 0.635
MedTerm 1510.23 (5.3%) 1498.61 (7.9%)
-0.8% ( -13% - 13%) 0.718
LowIntervalsOrdered 71.09 (3.3%) 70.59 (3.6%)
-0.7% ( -7% - 6%) 0.523
HighPhrase 15.80 (3.1%) 15.72 (3.5%)
-0.5% ( -6% - 6%) 0.607
MedSloppyPhrase 12.99 (2.2%) 12.94 (2.7%)
-0.4% ( -5% - 4%) 0.614
MedPhrase 11.68 (2.7%) 11.63 (2.7%)
-0.4% ( -5% - 5%) 0.646
Respell 42.75 (2.3%) 42.59 (2.7%)
-0.4% ( -5% - 4%) 0.645
LowSloppyPhrase 6.80 (2.3%) 6.77 (2.5%)
-0.3% ( -5% - 4%) 0.682
IntNRQ 32.19 (1.7%) 32.11 (1.8%)
-0.3% ( -3% - 3%) 0.633
LowPhrase 16.49 (2.6%) 16.45 (2.3%)
-0.3% ( -4% - 4%) 0.738
Fuzzy2 12.52 (3.0%) 12.49 (3.9%)
-0.2% ( -6% - 6%) 0.831
LowTerm 1338.97 (5.7%) 1336.19 (7.1%)
-0.2% ( -12% - 13%) 0.919
HighIntervalsOrdered 5.48 (2.3%) 5.47 (2.7%)
-0.2% ( -5% - 4%) 0.827
AndHighLow 295.57 (2.4%) 295.11 (3.2%)
-0.2% ( -5% - 5%) 0.861
LowSpanNear 39.91 (1.4%) 39.86 (1.5%)
-0.1% ( -3% - 2%) 0.775
HighTerm 1014.28 (4.6%) 1013.17 (6.4%)
-0.1% ( -10% - 11%) 0.951
BrowseMonthSSDVFacets 3.23 (5.0%) 3.23 (4.9%)
-0.1% ( -9% - 10%) 0.956
MedSpanNear 10.01 (2.1%) 10.01 (2.2%)
-0.1% ( -4% - 4%) 0.931
AndHighHigh 50.17 (2.5%) 50.17 (2.8%)
-0.0% ( -5% - 5%) 0.997
HighSpanNear 0.90 (1.3%) 0.90 (1.7%)
0.0% ( -2% - 3%) 0.997
MedIntervalsOrdered 18.18 (1.9%) 18.20 (2.2%)
0.1% ( -3% - 4%) 0.853
OrHighHigh 15.91 (1.7%) 15.93 (2.1%)
0.1% ( -3% - 3%) 0.820
HighTermDayOfYearSort 20.48 (8.0%) 20.54 (6.6%)
0.3% ( -13% - 16%) 0.903
OrHighMed 33.57 (1.9%) 33.68 (2.7%)
0.3% ( -4% - 5%) 0.637
BrowseDayOfYearSSDVFacets 2.99 (5.5%) 3.00 (5.0%)
0.4% ( -9% - 11%) 0.809
AndHighMed 44.35 (3.2%) 44.60 (3.1%)
0.6% ( -5% - 7%) 0.574
HighTermMonthSort 41.42 (14.7%) 41.92 (15.8%)
1.2% ( -25% - 37%) 0.805
HighTermTitleBDVSort 34.18 (12.8%) 34.70 (11.7%)
1.5% ( -20% - 29%) 0.699
TermDTSort 45.24 (9.5%) 45.93 (9.7%)
1.5% ( -16% - 22%) 0.616
BrowseDateTaxoFacets 0.72 (3.5%) 3.51 (62.8%)
388.3% ( 311% - 471%) 0.000
BrowseDayOfYearTaxoFacets 0.72 (3.4%) 3.52 (61.3%)
389.9% ( 314% - 470%) 0.000
BrowseMonthTaxoFacets 0.76 (3.5%) 3.95 (84.1%)
419.2% ( 320% - 525%) 0.000
{noformat}
Digging a little deeper, here's what I'm seeing as top CPU time:
baseline:
{noformat}
PERCENT CPU SAMPLES STACK
12.89% 286328
org.apache.lucene.util.packed.DirectMonotonicReader#get()
7.18% 159607
org.apache.lucene.codecs.lucene90.Lucene90DocValuesProducer$15#binaryValue()
6.90% 153297
org.apache.lucene.util.packed.DirectReader$DirectPackedReader12#get()
6.25% 138833
org.apache.lucene.facet.taxonomy.FastTaxonomyFacetCounts#countAll()
{noformat}
candidate:
{noformat}
PERCENT CPU SAMPLES STACK
4.77% 62575
org.apache.lucene.index.SingletonSortedNumericDocValues#nextDoc()
4.30% 56479
org.apache.lucene.codecs.lucene90.Lucene90PostingsReader$EverythingEnum#nextPosition()
4.20% 55120
org.apache.lucene.util.packed.DirectReader$DirectPackedReader12#get()
3.97% 52068
org.apache.lucene.codecs.lucene90.Lucene90DocValuesProducer$18#nextDoc()
3.77% 49425
org.apache.lucene.codecs.lucene90.Lucene90DocValuesProducer$4#longValue()
3.35% 43952
org.apache.lucene.queries.spans.NearSpansOrdered#nextStartPosition()
3.29% 43142
org.apache.lucene.codecs.lucene90.Lucene90PostingsReader$BlockImpactsPostingsEnum#advance()
2.86% 37556
org.apache.lucene.queries.spans.TermSpans#nextStartPosition()
2.83% 37102
org.apache.lucene.codecs.lucene90.Lucene90PostingsReader$EverythingEnum#advance()
2.62% 34434
org.apache.lucene.queries.spans.NearSpansOrdered#stretchToOrder()
2.53% 33236
org.apache.lucene.util.packed.DirectReader$DirectPackedReader4#get()
1.86% 24417
org.apache.lucene.facet.sortedset.SortedSetDocValuesFacetCounts#countOneSegment()
1.85% 24271
org.apache.lucene.queries.spans.SpanScorer#setFreqCurrentDoc()
1.73% 22668
org.apache.lucene.search.similarities.BM25Similarity$BM25Scorer#score()
1.70% 22365
org.apache.lucene.facet.taxonomy.FastTaxonomyFacetCounts#countAll()
{noformat}
So a drop from 6.25% CPU time to 1.7% for {{FastTaxonomyFacetCounts#countAll}}
On top of this, the index actually gets smaller (by ~1.4%).
{noformat}
11504472
wikimediumall.baseline.facets.taxonomy:Date.taxonomy:Month.taxonomy:DayOfYear.sortedset:Month.sortedset:DayOfYear.Lucene90.Lucene90.nd33.3326M
11334516
wikimediumall.candidate.facets.taxonomy:Date.taxonomy:Month.taxonomy:DayOfYear.sortedset:Month.sortedset:DayOfYear.Lucene90.Lucene90.nd33.3326M
{noformat}
And... I haven't even optimized the single-value case yet (which will be easy
to do and may squeeze out a little more performance based on what we saw with
SSDV faceting).
Like I said, almost too good to be true. I've uploaded a PR here and would
appreciate another set of eyes to see if I have something fundamentally wrong:
https://github.com/apache/lucene/pull/264
> Explore using SORTED_NUMERIC doc values to encode taxonomy ordinals for
> faceting
> --------------------------------------------------------------------------------
>
> Key: LUCENE-10062
> URL: https://issues.apache.org/jira/browse/LUCENE-10062
> Project: Lucene - Core
> Issue Type: Improvement
> Components: modules/facet
> Reporter: Greg Miller
> Assignee: Greg Miller
> Priority: Minor
>
> We currently encode taxonomy ordinals using varint style packing in a binary
> doc values field. I suspect there have been a number of improvements to
> SortedNumericDocValues since taxonomy faceting was first introduced, and I
> plan to explore replacing the custom binary format we have today with a
> SORTED_NUMERIC type dv field instead.
> I'll report benchmark results and index size impact here.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
