Hi Adrien, We find that on Lucene 911: 1. A MUST of two ConstantScoreQuery: (+ConstantScore($meta/fieldNames:searchableField) +ConstantScore($meta/fieldNames:_bus_key)) invokes the score 3.6M times. The query latency is quite high. And the the Max Conjunction scorer shows up in the flamegraph <https://htmlpreview.github.io/?https://github.com/wurui90/scratch/blob/main/flamegraphs/exists-with-limit-200-lucene911-must.html> .
while 2. A FILTER of two ConstantScoreQuery: #ConstantScore($meta/fieldNames:searchableField) #ConstantScore($meta/fieldNames:_bus_key) invokes the score 1001 times.In my mental model, the 1's query result is identical to 2 and 1 can be optimized to 2. I wonder why doesn't Lucene internal does this optimization?" On Lucene97, both queries invoke the score 1001 times. Thanks! On Fri, Sep 20, 2024 at 11:53 AM Rui Wu <rui...@mongodb.com> wrote: > Hi Adrien, > > You are right, the Max Conjunction scorer shows up in the flamegraph for > 12 MUST clauses: > https://htmlpreview.github.io/?https://github.com/wurui90/scratch/blob/main/flamegraphs/exists-with-limit-200-lucene911-must.html > > > Thanks! > > On Fri, Sep 20, 2024 at 2:27 AM Adrien Grand <jpou...@gmail.com> wrote: > >> This suggests that BlockMaxConjunctionBulkScorer has a similar issue, >> I'll look into it too. >> >> On Thu, Sep 19, 2024 at 2:48 AM Rui Wu <rui...@mongodb.com> wrote: >> >>> Hi Adrien, >>> >>> Thanks for your help and putting up a fix! >>> >>> Another experiment I did without your PR: if the 12 SHOULD clauses are >>> changed to 12 MUST clauses, the problem is the same: it collects 3.6M docs >>> on Lucene911 but 1001 docs on Lucene97. Does this data point align with how >>> MaxScoreBulkScorer works? >>> >>> Thanks! >>> >>> On Wed, Sep 18, 2024 at 1:51 AM Adrien Grand <jpou...@gmail.com> wrote: >>> >>>> Thank you, this last comment was helpful and helped me understand the >>>> problem. I opened a PR at https://github.com/apache/lucene/pull/13800. >>>> >>>> On Tue, Sep 17, 2024 at 7:45 PM Rui Wu <rui...@mongodb.com> wrote: >>>> >>>>> Another information is that, in Lucene97, this query (12 SHOULD >>>>> clauses) collected 1001 results; while in Lucene911, this query (12 SHOULD >>>>> clauses) collected all docs (3.6M collect count). >>>>> >>>>> In Lucene911, if the query has only one SHOULD clause, it collects >>>>> 1001 results. If the query has multiple clauses, it collects 3.6M results. >>>>> >>>>> On Tue, Sep 17, 2024 at 9:09 AM Rui Wu <rui...@mongodb.com> wrote: >>>>> >>>>>> This query latency increased from 14.65 to 20.90ms. >>>>>> >>>>>> We use the `TopScoreDocCollector.createSharedManager(/*batchSize*/ >>>>>> 101, /*searchAfterFieldDoc*/ null, /*hitsThreshold*/ 1000); ` >>>>>> >>>>>> Thanks a lot! >>>>>> >>>>>> On Tue, Sep 17, 2024 at 6:45 AM Adrien Grand <jpou...@gmail.com> >>>>>> wrote: >>>>>> >>>>>>> Can you tell us how long this query used to take, and how long it >>>>>>> takes now? >>>>>>> Also are you using IndexSearcher's default total hit count threshold >>>>>>> of 1,000, or are you passing a custom value to >>>>>>> TopScoreDocCollectorManager? >>>>>>> >>>>>>> On Tue, Sep 17, 2024 at 10:14 AM Rui Wu <rui...@mongodb.com> wrote: >>>>>>> >>>>>>>> Hi Adrien, >>>>>>>> >>>>>>>> Thanks for looking into this! Here are more screenshots of the >>>>>>>> flamegraph. The original flamegraph HTMLs have stack traces from our >>>>>>>> app so >>>>>>>> I don't share it here. >>>>>>>> [image: Screenshot 2024-09-17 at 1.13.07 AM.png][image: Screenshot >>>>>>>> 2024-09-17 at 1.12.01 AM.png] >>>>>>>> >>>>>>>> On Tue, Sep 17, 2024 at 1:00 AM Adrien Grand <jpou...@gmail.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> Hello Rui, >>>>>>>>> >>>>>>>>> We actually released a change that should make MaxScoreBulkScorer >>>>>>>>> faster on dense disjunctions in 9.8: >>>>>>>>> https://github.com/apache/lucene/pull/12444. Your benchmark case >>>>>>>>> is quite specific though as all clauses match all docs and produce >>>>>>>>> constant >>>>>>>>> scores, so I would expect the scorer to quickly realize that it can >>>>>>>>> skip >>>>>>>>> all documents once it's scored the first k docs. This makes me wonder >>>>>>>>> if it >>>>>>>>> bottleneck on skipping blocks of documents rather than on scoring >>>>>>>>> them. >>>>>>>>> Would you be able to share your whole flame graph, it looks like it >>>>>>>>> may be >>>>>>>>> truncated a the top? >>>>>>>>> >>>>>>>>> On Mon, Sep 16, 2024 at 10:01 PM Rui Wu <rui...@mongodb.com> >>>>>>>>> wrote: >>>>>>>>> >>>>>>>>>> Correction: The index has 3.6 million documents. >>>>>>>>>> >>>>>>>>>> On Mon, Sep 16, 2024 at 1:00 PM Rui Wu <rui...@mongodb.com> >>>>>>>>>> wrote: >>>>>>>>>> >>>>>>>>>>> Dear experts, >>>>>>>>>>> >>>>>>>>>>> In our Mongodb Atlas Search performance regression test between >>>>>>>>>>> Lucene 9.7 and Lucene 9.11, we detect a 43% latency regression in >>>>>>>>>>> this >>>>>>>>>>> query shape: >>>>>>>>>>> 12 SHOULD clause, and each clause matches all of the documents. >>>>>>>>>>> Each should clause is wrapped in ConstantScoreQuery. >>>>>>>>>>> >>>>>>>>>>> The index has 3.6 documents, and every document is identical: >>>>>>>>>>> Every document is {"path": ["1", "2", "3" ... "12"]} >>>>>>>>>>> The query shape is a BooleanQuery of SHOULD "1", SHOULD "2", ... >>>>>>>>>>> SHOULD "12". >>>>>>>>>>> >>>>>>>>>>> Our flamegraphs show that most of the time in search() is spent >>>>>>>>>>> on the MaxScoreBulkScorer class: >>>>>>>>>>> [image: image.png] >>>>>>>>>>> >>>>>>>>>>> We wonder if this extreme test case is expected to be slow on >>>>>>>>>>> MaxScoreBulkScorer? >>>>>>>>>>> >>>>>>>>>>> Thanks a lot! >>>>>>>>>>> >>>>>>>>>>> Rui Wu >>>>>>>>>>> Lead Engineer, MongoDB >>>>>>>>>>> >>>>>>>>>> >>>>>>>>> >>>>>>>>> -- >>>>>>>>> Adrien >>>>>>>>> >>>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Adrien >>>>>>> >>>>>> >>>> >>>> -- >>>> Adrien >>>> >>> >> >> -- >> Adrien >> >
