[
https://issues.apache.org/jira/browse/LUCENE-2798?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13018365#comment-13018365
]
Robert Muir commented on LUCENE-2798:
-------------------------------------
{quote}
I wrote a test including just [32] and [28,777] as indexed strings, and the
same mismatch occurs for random locales, regardless of collator decomposition,
and for all collator strengths except PRIMARY.
{quote}
Without looking too hard (are these hex values?) in your debugging it would be
useful to print the sort key as well. Are the sort keys the same?
But FYI the bugs i found in collation, somehow corrupted the internal state of
RuleBasedCollator, so the exact strings you are looking at might simply be a
symptom.
> Randomize indexed collation key testing
> ---------------------------------------
>
> Key: LUCENE-2798
> URL: https://issues.apache.org/jira/browse/LUCENE-2798
> Project: Lucene - Java
> Issue Type: Test
> Components: Analysis
> Affects Versions: 3.1, 4.0
> Reporter: Steven Rowe
> Assignee: Steven Rowe
> Priority: Minor
> Fix For: 4.0
>
> Attachments: LUCENE-2798.patch
>
>
> Robert Muir noted on #lucene IRC channel today that Lucene's indexed
> collation key testing is currently fragile (for example, they had to be
> revisited when Robert upgraded the ICU dependency in LUCENE-2797 because of
> Unicode 6.0 collation changes) and coverage is trivial (only 5 locales
> tested, and no collator options are exercised). This affects both the JDK
> implementation in {{modules/analysis/common/}} and the ICU implementation
> under {{modules/icu/}}.
> The key thing to test is that the order of the indexed terms is the same as
> that provided by the Collator itself. Instead of the current set of static
> tests, this could be achieved via indexing randomly generated terms'
> collation keys (and collator options) and then comparing the index terms'
> order to the order provided by the Collator over the original terms.
> Since different terms may produce the same collation key, however, the order
> of indexed terms is inherently unstable. When performing runtime collation,
> the Collator addresses the sort stability issue by adding a secondary sort
> over the normalized original terms. In order to directly compare Collator's
> sort with Lucene's collation key sort, a secondary sort will need to be
> applied to Lucene's indexed terms as well. Robert has suggested indexing the
> original terms in addition to their collation keys, then using a Sort over
> the original terms as the secondary sort.
> Another complication: Lucene 3.X uses Java's UTF-16 term comparison, and
> trunk uses UTF-8 order, so the implemented secondary sort will need to
> respect that.
> From #lucene:
> {quote}
> rmuir__: so i think we have to on 3.x, sort the 'expected list' with
> Collator.compare, if thats equal, then as a tiebreak use String.compareTo
> rmuir__: and in the index sort on the collated field, followed by the
> original term
> rmuir__: in 4.x we do the same thing, but dont use String.compareTo as the
> tiebreak for the expected list
> rmuir__: instead compare codepoints (iterating character.codepointAt, or
> comparing .getBytes("UTF-8"))
> {quote}
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
