[ 
https://issues.apache.org/jira/browse/LUCENE-8753?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16809124#comment-16809124
 ] 

Mike Sokolov commented on LUCENE-8753:
--------------------------------------

The behavior I'm referring to isn't a problem with the benchmark _per se_, it's 
an intrinsic feature of trying to optimize blocks of things that work better 
when you have a bunch of similar things right next to each other sharing a 
common prefix. The PKLookup test is very sensitive to such optimizations (or 
failures to optimize) because it indexes a consecutive block of terms: a00, 
a01, a02, etc. and at the same time, this nice consecutive property can be 
disturbed by index fragmentation, especially when only each term is indexed for 
only a single document. 

> New PostingFormat - UniformSplit
> --------------------------------
>
>                 Key: LUCENE-8753
>                 URL: https://issues.apache.org/jira/browse/LUCENE-8753
>             Project: Lucene - Core
>          Issue Type: Improvement
>          Components: core/codecs
>    Affects Versions: 8.0
>            Reporter: Bruno Roustant
>            Priority: Major
>         Attachments: Uniform Split Technique.pdf, luceneutil.benchmark.txt
>
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> This is a proposal to add a new PostingsFormat called "UniformSplit" with 4 
> objectives:
>  - Clear design and simple code.
>  - Easily extensible, for both the logic and the index format.
>  - Light memory usage with a very compact FST.
>  - Focus on efficient TermQuery, PhraseQuery and PrefixQuery performance.
> (the pdf attached explains visually the technique in more details)
>  The principle is to split the list of terms into blocks and use a FST to 
> access the block, but not as a prefix trie, rather with a seek-floor pattern. 
> For the selection of the blocks, there is a target average block size (number 
> of terms), with an allowed delta variation (10%) to compare the terms and 
> select the one with the minimal distinguishing prefix.
>  There are also several optimizations inside the block to make it more 
> compact and speed up the loading/scanning.
> The performance obtained is interesting with the luceneutil benchmark, 
> comparing UniformSplit with BlockTree. Find it in the first comment and also 
> attached for better formatting.
> Although the precise percentages vary between runs, three main points:
>  - TermQuery and PhraseQuery are improved.
>  - PrefixQuery and WildcardQuery are ok.
>  - Fuzzy queries are clearly less performant, because BlockTree is so 
> optimized for them.
> Compared to BlockTree, FST size is reduced by 15%, and segment writing time 
> is reduced by 20%. So this PostingsFormat scales to lots of docs, as 
> BlockTree.
> This initial version passes all Lucene tests. Use “ant test 
> -Dtests.codec=UniformSplitTesting” to test with this PostingsFormat.
> Subjectively, we think we have fulfilled our goal of code simplicity. And we 
> have already exercised this PostingsFormat extensibility to create a 
> different flavor for our own use-case.
> Contributors: Juan Camilo Rodriguez Duran, Bruno Roustant, David Smiley



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org

Reply via email to