[ https://issues.apache.org/jira/browse/LUCENE-1575?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12695876#action_12695876 ]
Shai Erera commented on LUCENE-1575: ------------------------------------ bq. Shai which OS/JRE did you run on? Windows XP, 1.5 (32 bit) I chose the sort by doc since I was very curious as to the major drop in performance your queries show. When you sort by doc, the queue converges exactly after 10 documents (assuming you ask for 10 results). After that, every document that's collected fails at the comparator level (first line in the "if (queueFull)" code) and the collector literally does nothing. So all that's left are the calls to scorer.score(). I chose to sort by docs reversed so that every call will actually do something. I measured scorer.score() calls in OCSC and they were insignificant (~14 ms total on a single query run). bq. Also, I think it's better to take min rather than avg across all the runs. I don't know. Some runs showed 112 ms and some 256 ms ... Taking the minimum just measures "when the OS cache is fully optimized for your run" case, that's why I prefer to take avg. Anyway, I don't think it matters since we're taking the same for both runs. I reviewed the code in TopFieldCollector and it's almost identical to before the patch, besides I differences I noted above, which don't seem like to have an effect. That's why I wanted to have a synthetic test, which we know exactly what it'll do. BTW, this test shows far smaller drop in performance comparing to your previous sort-by-doc runs, and it does collect ~x20 times more documents ... > Refactoring Lucene collectors (HitCollector and extensions) > ----------------------------------------------------------- > > Key: LUCENE-1575 > URL: https://issues.apache.org/jira/browse/LUCENE-1575 > Project: Lucene - Java > Issue Type: Improvement > Components: Search > Reporter: Shai Erera > Assignee: Michael McCandless > Fix For: 2.9 > > Attachments: LUCENE-1575.1.patch, LUCENE-1575.2.patch, > LUCENE-1575.3.patch, LUCENE-1575.4.patch, LUCENE-1575.5.patch, > LUCENE-1575.6.patch, LUCENE-1575.7.patch, LUCENE-1575.patch, > LUCENE-1575.patch, LUCENE-1575.patch, sortBench5.py, sortCollate5.py > > > This issue is a result of a recent discussion we've had on the mailing list. > You can read the thread > [here|http://www.nabble.com/Is-TopDocCollector%27s-collect()-implementation-correct--td22557419.html]. > We have agreed to do the following refactoring: > * Rename MultiReaderHitCollector to Collector, with the purpose that it will > be the base class for all Collector implementations. > * Deprecate HitCollector in favor of the new Collector. > * Introduce new methods in IndexSearcher that accept Collector, and deprecate > those that accept HitCollector. > ** Create a final class HitCollectorWrapper, and use it in the deprecated > methods in IndexSearcher, wrapping the given HitCollector. > ** HitCollectorWrapper will be marked deprecated, so we can remove it in 3.0, > when we remove HitCollector. > ** It will remove any instanceof checks that currently exist in IndexSearcher > code. > * Create a new (abstract) TopDocsCollector, which will: > ** Leave collect and setNextReader unimplemented. > ** Introduce protected members PriorityQueue and totalHits. > ** Introduce a single protected constructor which accepts a PriorityQueue. > ** Implement topDocs() and getTotalHits() using the PQ and totalHits members. > These can be used as-are by extending classes, as well as be overridden. > ** Introduce a new topDocs(start, howMany) method which will be used a > convenience method when implementing a search application which allows paging > through search results. It will also attempt to improve the memory > allocation, by allocating a ScoreDoc[] of the requested size only. > * Change TopScoreDocCollector to extend TopDocsCollector, use the topDocs() > and getTotalHits() implementations as they are from TopDocsCollector. The > class will also be made final. > * Change TopFieldCollector to extend TopDocsCollector, and make the class > final. Also implement topDocs(start, howMany). > * Change TopFieldDocCollector (deprecated) to extend TopDocsCollector, > instead of TopScoreDocCollector. Implement topDocs(start, howMany) > * Review other places where HitCollector is used, such as in Scorer, > deprecate those places and use Collector instead. > Additionally, the following proposal was made w.r.t. decoupling score from > collect(): > * Change collect to accecpt only a doc Id (unbased). > * Introduce a setScorer(Scorer) method. > * If during collect the implementation needs the score, it can call > scorer.score(). > If we do this, then we need to review all places in the code where > collect(doc, score) is called, and assert whether Scorer can be passed. Also > this raises few questions: > * What if during collect() Scorer is null? (i.e., not set) - is it even > possible? > * I noticed that many (if not all) of the collect() implementations discard > the document if its score is not greater than 0. Doesn't it mean that score > is needed in collect() always? > Open issues: > * The name for Collector > * TopDocsCollector was mentioned on the thread as TopResultsCollector, but > that was when we thought to call Colletor ResultsColletor. Since we decided > (so far) on Collector, I think TopDocsCollector makes sense, because of its > TopDocs output. > * Decoupling score from collect(). > I will post a patch a bit later, as this is expected to be a very large > patch. I will split it into 2: (1) code patch (2) test cases (moving to use > Collector instead of HitCollector, as well as testing the new topDocs(start, > howMany) method. > There might be even a 3rd patch which handles the setScorer thing in > Collector (maybe even a different issue?) -- 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: java-dev-unsubscr...@lucene.apache.org For additional commands, e-mail: java-dev-h...@lucene.apache.org