[
https://issues.apache.org/jira/browse/LUCENE-1536?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12670455#action_12670455
]
Michael McCandless commented on LUCENE-1536:
--------------------------------------------
Some ideas / further things to explore:
* Deletions, top-level filters, and BooleanQuery that "factors" to a
toplevel AND really should be handled by the same code.
* Even an AND'd filter on a sub-clause of a BooleanQuery can be
pushed down to the TermDocs under that tree.
* That common code should send a top-level filter down to the lowest
level, used by random access API, if the filter supports random
access (not all do) and it's not super sparse.
* I think one thing slowing down trunk is the lack of a
Scorer.skipToButNotNext API. We now ask the filter for its
next(), which gives us a filterDocID. Then we call
Scorer.skipTo(filterDocID). If the scorer does not match that
filterDocID, it internally does next(), which for an expensive
scorer is alot of likely wasted work: it advances to a docID that
the filter may not accept. If we had a "skipToButNotNext" API we
could avoid that wasted work. I'm curious what gains this change
alone would provide.
* I'm thinking (but haven't tested this) if the filter is relatively
sparse compared to the other iterators, it'd be better to convert
it to a sparse repr (eg SortedVIntList) and drive the search by
iteration through the filter, after fixing the above skipTo issue.
Maybe a "low iterator" access.
* We may need a "filter optimizer" utility class somewhere, somehow.
For filters you do not plan to re-use, you would not bother with
this. But for filters that will be re-used, you should 1) convert
them to sparse or non-sparse repr depending on their density, 2)
maybe invert them and make sparse if they are close to 100%
density, 3) maybe factor in deletions to the filter so there is
only a single top-level filter to apply.
* I'm not yet sure how to make this change cleanly to the APIs...
> if a filter can support random access API, we should use it
> -----------------------------------------------------------
>
> Key: LUCENE-1536
> URL: https://issues.apache.org/jira/browse/LUCENE-1536
> Project: Lucene - Java
> Issue Type: Improvement
> Components: Search
> Affects Versions: 2.4
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Priority: Minor
> Attachments: LUCENE-1536.patch
>
>
> I ran some performance tests, comparing applying a filter via
> random-access API instead of current trunk's iterator API.
> This was inspired by LUCENE-1476, where we realized deletions should
> really be implemented just like a filter, but then in testing found
> that switching deletions to iterator was a very sizable performance
> hit.
> Some notes on the test:
> * Index is first 2M docs of Wikipedia. Test machine is Mac OS X
> 10.5.6, quad core Intel CPU, 6 GB RAM, java 1.6.0_07-b06-153.
> * I test across multiple queries. 1-X means an OR query, eg 1-4
> means 1 OR 2 OR 3 OR 4, whereas +1-4 is an AND query, ie 1 AND 2
> AND 3 AND 4. "u s" means "united states" (phrase search).
> * I test with multiple filter densities (0, 1, 2, 5, 10, 25, 75, 90,
> 95, 98, 99, 99.99999 (filter is non-null but all bits are set),
> 100 (filter=null, control)).
> * Method high means I use random-access filter API in
> IndexSearcher's main loop. Method low means I use random-access
> filter API down in SegmentTermDocs (just like deleted docs
> today).
> * Baseline (QPS) is current trunk, where filter is applied as iterator up
> "high" (ie in IndexSearcher's search loop).
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
