[
https://issues.apache.org/jira/browse/CRUNCH-604?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15247283#comment-15247283
]
Josh Wills commented on CRUNCH-604:
-----------------------------------
[~blendmaster] thanks for this, looks awfully sensible. Unfortunately, the
patch to TupleWritable didn't apply cleanly for me (some stuff about
EMPTY_WRITABLES in the def that I don't have in master)-- are we missing some
files, or is this a patch off of a forked version I need to adapt the patch
against?
> Avoid expensive Writables.reloadWritableComparableCodes where possible
> ----------------------------------------------------------------------
>
> Key: CRUNCH-604
> URL: https://issues.apache.org/jira/browse/CRUNCH-604
> Project: Crunch
> Issue Type: Improvement
> Components: Core
> Affects Versions: 0.13.0
> Reporter: Steven Ruppert
> Assignee: Josh Wills
> Attachments:
> 0001-TupleWritable-only-reload-codes-once-on-setConf.patch,
> 0001-Writables-cache-reloadWritableComparables-when-it-ha.patch
>
>
> Every time `setConf` is called on TupleWritable,
> `Writables.reloadWritableComparableCodes(conf)` is called. Unfortunately,
> `SequenceFile$Reader.readValue` calls `setConf` every single time. This burns
> a regrettable amount of CPU time.
> Attached is a patch that prevents a given TupleWritable instance from
> reloading the code more than once, as well as a patch to cache
> (hashCode-wise) reading from the actual hadoop config, which has to run
> regexes and stuff. I can construe situations where this would break (somehow,
> you modify the configuration in between reading to two values), but nothing
> actually sane comes to mind.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
