lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Uwe Schindler (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (LUCENE-4546) SorterTemplate.quicksort incorrect
Date Wed, 07 Nov 2012 21:34:12 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-4546?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13492720#comment-13492720
] 

Uwe Schindler commented on LUCENE-4546:
---------------------------------------

Hi Stefan,
it is some time ago when I worked on SorterTemplate, so I don't have all the facts in mind.
There are different implementations of QuickSort available, also those working without any
pivot (like yours). But as far as I remember, the performance tests showed, that the additional
swaps and compares added some slowdown (depending on the order of input data), so the explicit
pivot methods helped. The SorterTemplate quicksort implementation is also the one that was
used in Lucene from the beginning, so I did not want to change the algorithm in a minor release.
We could add some new performance tests with your implementation and compare the speed, but
I think, e.g. CollectionUtil, which uses Collections.swap() would get much slower by this.

I agree, the class is very nice for sorting of non-array data, but it is currently marked
as @lucene.internal, so the usability for non-lucene code was never thought of, performance
was the only driving force :-) But I checked the javadocs, it is clearly documented that setPivot(i)
has to store the value of slot i for later comparison with comparePivot(j).
                
> SorterTemplate.quicksort incorrect
> ----------------------------------
>
>                 Key: LUCENE-4546
>                 URL: https://issues.apache.org/jira/browse/LUCENE-4546
>             Project: Lucene - Core
>          Issue Type: Bug
>          Components: core/other
>    Affects Versions: 3.6.1, 4.0, 4.1
>            Reporter: Stefan Pohl
>            Assignee: Uwe Schindler
>              Labels: patch
>             Fix For: 3.6.1, 4.0, 4.1
>
>         Attachments: SorterTemplate.java.patch, TestSorterTemplate.java, TestSorterTemplate.java
>
>
> On trying to use the very useful o.a.l.utils.SorterTemplate, I stumbled upon inconsistent
sorting behaviour, of course, only a randomized test caught this;)
> Because SorterTemplate.quicksort is used in several places in the code (directly BytesRefList,
ArrayUtil, BytesRefHash, CollectionUtil and transitively index and search), I'm a bit puzzled
that this either hasn't been caught by another higher-level test or that neither my test nor
my understanding of an insufficiency in the code is valid;)
> If the former holds and given that the same code is released in 3.6 and 4.0, this might
even be a more critical issue requiring a higher priority than 'major'.
> So, can a second pair of eyes please have a timely look at the attached test and patch?
> Basically the current quicksort implementation seems to assume that luckily always the
median is chosen as pivot element by grabbing the mid element, not handling the case where
the initially chosen pivot ends up not in the middle. Hope this and the test helps to understand
the issue.
> Reproducible, currently failing test and a patch attached.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

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


Mime
View raw message