lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Lance Norskog (JIRA)" <j...@apache.org>
Subject [jira] Commented: (LUCENE-2373) Change StandardTermsDictWriter to work with streaming and append-only filesystems
Date Thu, 06 May 2010 22:37:48 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-2373?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12864957#action_12864957
] 

Lance Norskog commented on LUCENE-2373:
---------------------------------------

.bq HDFS uses 64 or 128 _Mega_Byte blocks. 
Yet another reason to manage memory carefully. 

It should be possible to hit this watermark by using the NoMergePolicy and a RamBuffer size
of 64M or 128M:. Hitting the RAMBuffer size causes a segment to flush to a file with little
breakage (unused disk space), and it will never be merged again, cutting HDFS overheads. This
should give a predictable and consistent segment writing overhead, right?


> Change StandardTermsDictWriter to work with streaming and append-only filesystems
> ---------------------------------------------------------------------------------
>
>                 Key: LUCENE-2373
>                 URL: https://issues.apache.org/jira/browse/LUCENE-2373
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Index
>            Reporter: Andrzej Bialecki 
>             Fix For: 4.0
>
>
> Since early 2.x times Lucene used a skip/seek/write trick to patch the length of the
terms dict into a place near the start of the output data file. This however made it impossible
to use Lucene with append-only filesystems such as HDFS.
> In the post-flex trunk the following code in StandardTermsDictWriter initiates this:
> {code}
>     // Count indexed fields up front
>     CodecUtil.writeHeader(out, CODEC_NAME, VERSION_CURRENT); 
>     out.writeLong(0);                             // leave space for end index pointer
> {code}
> and completes this in close():
> {code}
>       out.seek(CodecUtil.headerLength(CODEC_NAME));
>       out.writeLong(dirStart);
> {code}
> I propose to change this layout so that this pointer is stored simply at the end of the
file. It's always 8 bytes long, and we known the final length of the file from Directory,
so it's a single additional seek(length - 8) to read it, which is not much considering the
benefits.

-- 
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