lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Earwin Burrfoot (JIRA)" <j...@apache.org>
Subject [jira] Commented: (LUCENE-2506) A Stateful Filter That Works Across Index Segments
Date Fri, 26 Nov 2010 03:10:36 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-2506?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12935927#action_12935927
] 

Earwin Burrfoot commented on LUCENE-2506:
-----------------------------------------

bq. We are not deleting documents, and have not witnessed merges changing doc IDs in any other
situation.
That assumption's gonna break very soon. Very very soon, when IndexWriter learns how to merge
non-sequential segments.

You can do it same way Zoie authors did. They have a special per-segment cache that maps external
ID <-> docID.
The cache is pre-loaded when opening the segment from a field. It costs some memory, but nothing
frightening.
They found it optimal to add the same Term("ID", "ID") to all the docs and then store external
ID in that term's payload -> no TermsIndex pressure + gets read fast.
Then you can convert between IDs at a cost of simple array lookup.

Now, this IS reality. Bulletproof, fast lookups by external ID, and no broken cheats.

P.S. Not sure it was exactly Zoie, but definetly some of LinkedIn's opensourced Lucene stuff.

> A Stateful Filter That Works Across Index Segments
> --------------------------------------------------
>
>                 Key: LUCENE-2506
>                 URL: https://issues.apache.org/jira/browse/LUCENE-2506
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Index
>    Affects Versions: 3.0.2
>            Reporter: Karthick Sankarachary
>         Attachments: LUCENE-2506.patch
>
>
> By design, Lucene's Filter abstraction is applied once for every segment in the index
during searching. In particular, the reader provided to its #getDocIdSet method does not represent
the whole underlying index. In other words, if the index has more than one segment the given
reader only represents a single segment.  As a result, that definition of the filter suffers
the limitation of not having the ability to permit/prohibit documents in the search results
based on the terms that reside in segments that precede the current one.
> To address this limitation, we introduce here a StatefulFilter which specifically builds
on the Filter class so as to make it capable of remembering terms in segments spanning the
whole underlying index. To reiterate, the need for making filters stateful stems from the
fact that some, although not most, filters care about the terms that they may have come across
in prior segments. It does so by keeping track of the past terms from prior segments in a
cache that is maintained in a StatefulTermsEnum instance on a per-thread basis. 
> Additionally, to address the case where a filter might want to accept the last matching
term, we keep track of the TermsEnum#docFreq of the terms in the segments filtered thus far.
By comparing the sum of such TermsEnum#docFreq with that of the top-level reader, we can tell
if the current segment is the last segment in which the current term appears. Ideally, for
this to work correctly, we require the user to explicitly set the top-level reader on the
StatefulFilter. Knowing what the top-level reader is also helps the StatefulFilter to clean
up after itself once the search has concluded.
> Note that we leave it up to each concrete sub-class of the stateful filter to decide
what to remember in its state and what not to. In other words, it can choose to remember as
much or as little from prior segments as it deems necessary. In keeping with the TermsEnum
interface, which the StatefulTermsEnum class extends, the filter must decide which terms to
accept or not, based on the holistic state of the search.  

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


Mime
View raw message