A specialized Trie for Block Tree Index #14333

gf2121 · 2025-03-06T08:02:34Z

Context

Write MSB VLong for better outputs sharing in block tree index #12631 introduced a MSBVLong format to encode the first fp of FST output. It is the first time we benefit from the output sharing in blocktree. The change reduces ~13% tip size, in turn caused a performance regression when accumulating output bytes Nightly benchmark regression for term dict queries #12659. Then Disable suffix sharing for block tree index #12722 introduce a complex and tricky OutputAccumulator to get the performance back a bit, while still slower than no output prefix sharing.
Disable suffix sharing for block tree index #12722 we disabled suffix sharing as we find that very few suffix get shared in block tree.

Proposal

Before the PRs mentioned above, the fst in block tree is almost like a trie - no output prefix sharing and few suffix sharing. When output prefix get shared, we saved space, but also lost performance. IMO for tip, performance is more important than storage size, which is usually a very small part of the whole index, and loaded off-heap.

This PR tries to introduce a simple trie that specialized for blocktree index:

It can have assumptions that keys are bytes so numbers are better encoded.
The output bytes array can be read from IndexInput without copy.
It also improves performance by replacing vint / vlong with a over-read and mask.
...

PR is still a draft, but the number looks promising.

Storage

	Baseline	Candidate	diff
_32_Lucene101_0.tip	4425601	3827402	-13.52%
_65_Lucene101_0.tip	4458107	3884702	-12.86%
_98_Lucene101_0.tip	4791217	4185525	-12.64%
_cb_Lucene101_0.tip	4832497	4183043	-13.44%
_fe_Lucene101_0.tip	4807799	4173039	-13.20%
_fp_Lucene101_0.tip	720343	587765	-18.40%
_g0_Lucene101_0.tip	721438	585146	-18.89%
_gb_Lucene101_0.tip	694205	566073	-18.46%
_gm_Lucene101_0.tip	688145	561999	-18.33%
_gx_Lucene101_0.tip	819804	657046	-19.85%
_gy_Lucene101_0.tip	142276	100526	-29.34%
_gz_Lucene101_0.tip	125578	87568	-30.27%
_h0_Lucene101_0.tip	109982	77425	-29.60%
_h1_Lucene101_0.tip	113266	80308	-29.10%
_h2_Lucene101_0.tip	104672	72846	-30.41%
sum	27554930	23630413	-14.24%

Search

TaskQPS baseline      StdDevQPS my_modified_version      StdDev                Pct diff p-value
                          IntSet      595.28      (3.3%)      584.37      (3.9%)   -1.8% (  -8% -    5%) 0.110
                      OrHighRare      108.03      (8.7%)      106.53      (9.2%)   -1.4% ( -17% -   18%) 0.624
                  FilteredPhrase       16.86      (2.9%)       16.69      (3.0%)   -1.0% (  -6% -    5%) 0.288
                IntervalsOrdered        7.89      (3.1%)        7.81      (4.9%)   -1.0% (  -8% -    7%) 0.448
                        SpanNear       12.64      (2.1%)       12.56      (2.2%)   -0.6% (  -4% -    3%) 0.359
                    FilteredTerm      128.98      (3.2%)      128.26      (2.9%)   -0.6% (  -6% -    5%) 0.565
                 CountAndHighMed       84.00      (2.3%)       83.55      (1.6%)   -0.5% (  -4% -    3%) 0.389
                AndMedOrHighHigh       54.23      (4.6%)       53.99      (4.2%)   -0.5% (  -8% -    8%) 0.745
             CountFilteredPhrase       12.76      (2.3%)       12.70      (2.4%)   -0.4% (  -4% -    4%) 0.556
                  CountOrHighMed      172.95      (3.8%)      172.23      (3.6%)   -0.4% (  -7% -    7%) 0.724
                CountAndHighHigh       84.36      (2.0%)       84.03      (1.6%)   -0.4% (  -3% -    3%) 0.496
                     TermGroup1M       23.31      (1.9%)       23.24      (2.1%)   -0.3% (  -4% -    3%) 0.635
                          Phrase       50.78      (3.1%)       50.65      (2.9%)   -0.3% (  -6% -    5%) 0.781
                     CountPhrase       18.16      (1.8%)       18.12      (2.0%)   -0.2% (  -4% -    3%) 0.698
                    TermGroup10K       17.28      (1.4%)       17.24      (1.9%)   -0.2% (  -3% -    3%) 0.665
                  FilteredOrMany        7.97      (1.9%)        7.95      (1.8%)   -0.2% (  -3% -    3%) 0.730
                   TermTitleSort      138.40      (3.0%)      138.24      (3.6%)   -0.1% (  -6% -    6%) 0.910
     FilteredAnd2Terms2StopWords      436.82      (3.3%)      436.31      (3.1%)   -0.1% (  -6% -    6%) 0.906
             CountFilteredOrMany       13.23      (2.1%)       13.22      (2.2%)   -0.1% (  -4% -    4%) 0.889
             CombinedAndHighHigh        9.59      (2.7%)        9.59      (2.2%)   -0.1% (  -4% -    5%) 0.930
             FilteredOrStopWords       25.57      (3.9%)       25.56      (3.3%)   -0.0% (  -6% -    7%) 0.966
                 AndHighOrMedMed       15.54      (1.4%)       15.53      (1.4%)   -0.0% (  -2% -    2%) 0.939
         CountFilteredOrHighHigh       37.22      (1.2%)       37.21      (0.9%)   -0.0% (  -2% -    2%) 0.923
                    TermGroup100       16.94      (2.4%)       16.95      (2.3%)    0.0% (  -4% -    4%) 0.989
                     CountOrMany       15.03      (3.4%)       15.03      (3.4%)    0.0% (  -6% -    7%) 0.988
                 CountOrHighHigh       89.26      (1.9%)       89.28      (1.7%)    0.0% (  -3% -    3%) 0.960
             CountFilteredIntNRQ       23.19      (1.6%)       23.20      (1.7%)    0.0% (  -3% -    3%) 0.944
                  FilteredIntNRQ       35.57      (1.8%)       35.60      (2.3%)    0.1% (  -3% -    4%) 0.887
                          IntNRQ       35.72      (1.9%)       35.76      (2.1%)    0.1% (  -3% -    4%) 0.860
                    TermBGroup1M       23.72      (2.4%)       23.75      (2.2%)    0.1% (  -4% -    4%) 0.848
              CombinedAndHighMed       78.87      (3.2%)       79.00      (3.3%)    0.2% (  -6% -    6%) 0.876
               FilteredOrHighMed      147.44      (4.0%)      147.69      (2.8%)    0.2% (  -6% -    7%) 0.876
          CountFilteredOrHighMed       46.61      (1.1%)       46.69      (1.0%)    0.2% (  -1% -    2%) 0.609
              FilteredOrHighHigh       30.50      (3.9%)       30.58      (3.1%)    0.3% (  -6% -    7%) 0.816
                    AndStopWords        6.02      (6.8%)        6.04      (4.4%)    0.3% ( -10% -   12%) 0.881
                       OrHighMed       55.30      (4.4%)       55.45      (5.8%)    0.3% (  -9% -   10%) 0.868
                          OrMany        5.73      (6.5%)        5.75      (7.0%)    0.3% ( -12% -   14%) 0.898
                      AndHighMed      116.16      (3.6%)      116.50      (3.4%)    0.3% (  -6% -    7%) 0.789
                FilteredOr3Terms      106.20      (3.8%)      106.59      (2.4%)    0.4% (  -5% -    6%) 0.713
                    SloppyPhrase       22.88      (3.1%)       22.98      (3.1%)    0.4% (  -5% -    6%) 0.666
                  TermBGroup1M1P       13.89      (2.4%)       13.96      (1.9%)    0.4% (  -3% -    4%) 0.518
            FilteredAndStopWords       26.26      (4.3%)       26.38      (4.0%)    0.5% (  -7% -    9%) 0.719
               CombinedOrHighMed       15.73      (3.4%)       15.81      (1.5%)    0.5% (  -4% -    5%) 0.564
                 DismaxOrHighMed      100.85      (3.4%)      101.37      (3.6%)    0.5% (  -6% -    7%) 0.642
              CombinedOrHighHigh       12.56      (3.3%)       12.64      (1.0%)    0.6% (  -3% -    5%) 0.403
      FilteredOr2Terms2StopWords      192.94      (2.7%)      194.24      (1.9%)    0.7% (  -3% -    5%) 0.362
                      DismaxTerm      719.66      (5.8%)      724.60      (6.9%)    0.7% ( -11% -   14%) 0.733
                       And3Terms      493.05      (5.2%)      496.54      (5.4%)    0.7% (  -9% -   12%) 0.676
              Or2Terms2StopWords      102.67      (4.8%)      103.43      (5.4%)    0.7% (  -9% -   11%) 0.643
                    CombinedTerm       23.30      (3.5%)       23.48      (1.7%)    0.8% (  -4% -    6%) 0.365
             FilteredAndHighHigh       25.26      (4.6%)       25.46      (4.7%)    0.8% (  -8% -   10%) 0.594
                     OrStopWords       37.40      (4.8%)       37.73      (6.2%)    0.9% (  -9% -   12%) 0.614
                DismaxOrHighHigh       21.65      (3.7%)       21.86      (5.5%)    1.0% (  -7% -   10%) 0.521
               FilteredAnd3Terms       99.12      (5.5%)      100.08      (5.3%)    1.0% (  -9% -   12%) 0.573
                      OrHighHigh       24.82      (4.6%)       25.07      (5.8%)    1.0% (  -9% -   12%) 0.546
                     AndHighHigh       83.00      (4.4%)       83.91      (3.1%)    1.1% (  -6% -    9%) 0.369
                            Term      706.24      (4.0%)      713.95      (4.9%)    1.1% (  -7% -   10%) 0.440
                        Or3Terms      253.05      (5.5%)      256.03      (5.2%)    1.2% (  -9% -   12%) 0.487
                      TermDTSort      243.15      (3.3%)      246.59      (5.2%)    1.4% (  -6% -   10%) 0.305
                   TermMonthSort     2863.45      (7.1%)     2908.45      (7.4%)    1.6% ( -12% -   17%) 0.493
                 FilteredPrefix3      144.48      (2.9%)      146.86      (2.3%)    1.6% (  -3% -    7%) 0.046
             And2Terms2StopWords      247.08      (3.1%)      251.49      (3.1%)    1.8% (  -4% -    8%) 0.066
              FilteredAndHighMed      146.72      (7.4%)      149.37      (7.0%)    1.8% ( -11% -   17%) 0.426
                         Prefix3       68.50      (3.3%)       69.91      (1.6%)    2.1% (  -2% -    7%) 0.012
               TermDayOfYearSort      443.75      (5.2%)      453.71      (4.2%)    2.2% (  -6% -   12%) 0.134
                        Wildcard      100.44      (4.4%)      104.04      (3.3%)    3.6% (  -3% -   11%) 0.004
                          Fuzzy2       47.50      (3.4%)       49.31      (3.6%)    3.8% (  -3% -   11%) 0.001
                         Respell       64.88      (2.9%)       68.90      (3.7%)    6.2% (   0% -   13%) 0.000
                          Fuzzy1      121.48      (2.1%)      129.95      (3.1%)    7.0% (   1% -   12%) 0.000
                       CountTerm     7936.16     (10.6%)     8914.66      (8.8%)   12.3% (  -6% -   35%) 0.000
                        PKLookup      269.87      (3.6%)      361.18      (4.9%)   33.8% (  24% -   43%) 0.000

mikemccand

This looks great! Thank you for persisting @gf2121!

mikemccand · 2025-04-14T14:21:59Z

lucene/core/src/java/org/apache/lucene/codecs/Codec.java


    @SuppressWarnings("NonFinalStaticField")
-    static Codec defaultCodec = LOADER.lookup("Lucene101");
+    static Codec defaultCodec = LOADER.lookup("Lucene103");


Yay! This means default Codec is now using trie!

I think this is causing the OOMs. See #14487

gf2121 · 2025-04-14T15:27:43Z

Thank you @mikemccand and @jpountz for the patient review and all these great suggestions!

I raised mikemccand/luceneutil#369 to switch codec for luceneutil.

gf2121 · 2025-04-15T14:04:54Z

Nightly confirmed the speedup.

https://benchmarks.mikemccandless.com/2025.04.14.18.05.20.html
https://benchmarks.mikemccandless.com/PKLookup.html

Annotation PR: mikemccand/luceneutil#370

mikemccand · 2025-04-15T15:50:56Z

Nightly confirmed the speedup.

That's ~1.2 M primary key lookups per second, yay!

PK lookups are used all the time, not just the "key/value store" use case... e.g. updating or deleting a document by its id.

msokolov · 2025-04-15T19:31:17Z

I saw a message raising concern about some OOM test failures possibly related to this? Am I confused? Is excessive memory usage something we should be concerned about?

Hmm maybe it was a test issue only #14487

stefanvodita · 2025-04-23T14:56:07Z

If we think the issue last week was with the tests, should we go ahead and back-port this change?

gf2121 · 2025-04-23T15:11:43Z

Hi @stefanvodita

We plan to allow CI chew on this change for a couple of weeks and backport it if everything goes well. See #14333 (comment).

When backporting, i'll try to catch all changes made to Lucene103Codec, including #14447. So for now, feel free to just merge it into main :)

stefanvodita · 2025-04-23T15:14:44Z

You anticipated my concern with #14447 😄
Thank you for handling this!

* A specialized Trie for Block Tree Index (#14333) * Fix OOM of TestTrie (#14488) * Compute the doc range more efficiently when flushing doc block (#14447) * Fix TestForTooMuchCloning (#14547) * Fix tests: too many calls to IndexInput.clone during merging (#14595) * Simplify rootCodeFp (#14596) --------- Co-authored-by: panguixin <[email protected]>

Coqueue · 2025-05-20T08:48:55Z

Thanks for the fantastic change!

Want to share that we adopted and backported this codec to Lucene 912, and ran it against an Amazon Search internal benchmark, from which we observed an increase of 2.7% in Searcher throughput : D

Looking forward to seeing the actual improvements this brings in production!

mikemccand · 2025-05-21T12:11:30Z

Thank you @Coqueue.

ran it against an Amazon Search internal benchmark, from which we observed an increase of 2.7% in Searcher throughput : D

Small correction: Amazon Product Search internal benchmarks.

Also, we did this backport just to evaluate performance impact to our service from this change since we so heavily use Lucene's terms dictionary, but for actually running this in production we will upgrade to 10.x.

Hmm this change doesn't have the milestone set -- I'll set it.

jpountz · 2025-05-21T12:12:51Z

This is great info, thanks for sharing!

gf2121 added 17 commits March 4, 2025 16:30

trie test pass and writer test pass

8a54c70

iter

2f1da21

most tests passed

e5507bb

concurrency issue

a261ea7

concurrency issue

2ee3af9

all tests passed

aa12a84

rename classes and impl stats

583de5d

No need msb

5d30fd4

tidy

1d24e48

remove vlong

5dbb77a

no fst

53dfcb8

try to reduce virtual call

21e1798

iter

35760bf

shift one

13f10ae

delta codec

cb4c420

iter

7570df1

specialized single child

af2781a

github-project-automation bot added this to OpenSearch Lucene & Core Performance Tracking Mar 6, 2025

github-project-automation bot moved this to Open in OpenSearch Lucene & Core Performance Tracking Mar 6, 2025

github-actions bot added module:core/codecs module:test-framework labels Mar 6, 2025

gf2121 marked this pull request as draft March 6, 2025 08:20

gf2121 added 3 commits March 6, 2025 17:13

license

63b0cdd

iter

72fa0ba

iter

8387fc7

gf2121 closed this Mar 7, 2025

github-project-automation bot moved this from Open to Closed in OpenSearch Lucene & Core Performance Tracking Mar 7, 2025

gf2121 reopened this Mar 7, 2025

github-project-automation bot moved this from Closed to Open in OpenSearch Lucene & Core Performance Tracking Mar 7, 2025

iter

82b3172

gf2121 mentioned this pull request Apr 10, 2025

Fix TestInt7HnswBackwardsCompatibility #14460

Merged

gf2121 and others added 2 commits April 10, 2025 19:41

Merge branch 'main' into trie

b8191a4

fix 10.2.0.zip

c1b0044

gf2121 mentioned this pull request Apr 10, 2025

Add TestInt7HnswBackwardsCompatibility index for 10.2.0 #14461

Merged

gf2121 requested review from jpountz and mikemccand April 10, 2025 12:50

fix 103 -> 90

e2384a6

mikemccand approved these changes Apr 14, 2025

View reviewed changes

gf2121 merged commit 878e23f into apache:main Apr 14, 2025
7 checks passed

github-project-automation bot moved this from Open to Merged in OpenSearch Lucene & Core Performance Tracking Apr 14, 2025

gf2121 mentioned this pull request Apr 23, 2025

TestForTooMuchCloning: too many calls to IndexInput.clone during TermRangeQuery: 7 #14546

Closed

jpountz pushed a commit to jpountz/lucene that referenced this pull request Apr 24, 2025

A specialized Trie for Block Tree Index (apache#14333)

452ff11

gf2121 mentioned this pull request Apr 25, 2025

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

Merged

gf2121 added a commit to gf2121/lucene that referenced this pull request Apr 29, 2025

A specialized Trie for Block Tree Index (apache#14333)

5b57c0d

gf2121 added a commit to gf2121/lucene that referenced this pull request Apr 29, 2025

A specialized Trie for Block Tree Index (apache#14333)

b5f1948

gf2121 mentioned this pull request Apr 29, 2025

[Backport] A specialized Trie for Block Tree Index #14563

Merged

gf2121 added a commit to gf2121/lucene that referenced this pull request May 7, 2025

A specialized Trie for Block Tree Index (apache#14333)

df06c78

mikemccand added this to the 10.3.0 milestone May 21, 2025

romseygeek mentioned this pull request Oct 30, 2025

Sometimes high memory usage of Lucene103BlockTreeTermsWriter during merging elastic/elasticsearch#137359

Open

A specialized Trie for Block Tree Index #14333

A specialized Trie for Block Tree Index #14333

Uh oh!

Conversation

gf2121 commented Mar 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mikemccand left a comment

Choose a reason for hiding this comment

Uh oh!

mikemccand Apr 14, 2025

Choose a reason for hiding this comment

Uh oh!

dweiss Apr 14, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

gf2121 commented Apr 14, 2025

Uh oh!

gf2121 commented Apr 15, 2025

Uh oh!

mikemccand commented Apr 15, 2025

Uh oh!

msokolov commented Apr 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

stefanvodita commented Apr 23, 2025

Uh oh!

gf2121 commented Apr 23, 2025

Uh oh!

stefanvodita commented Apr 23, 2025

Uh oh!

Coqueue commented May 20, 2025

Uh oh!

mikemccand commented May 21, 2025

Uh oh!

jpountz commented May 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

gf2121 commented Mar 6, 2025 •

edited

Loading

msokolov commented Apr 15, 2025 •

edited

Loading