[
https://issues.apache.org/jira/browse/HBASE-2987?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12908400#action_12908400
]
Jonathan Gray commented on HBASE-2987:
--------------------------------------
bq. You have not tried to use LZMA compression before, or something like it.
I'm not saying that kind of thing would be common.
Agreed (isn't that an argument to not change default?).
Let me try to be more clear on my point. In all tests I've run previously (w/
and w/o LZO compression) it is not the flush that bites you, it's the
compaction. If you are tripping over the flush build-up, you are worse off
than other tests because normally that is fine and it is the compactions that
end up blocking stuff.
Disabling compression for flushes may prevent the flushing from stacking but
then you will get hit by the even longer-running compactions.
Thinking through this again, perhaps what I'm really concluding is that it
might make sense to disable compression on flushes *and* minor compactions and
to enable it on major compactions (in an archival example like this using LZMA
or the like). It could even make sense to use something like LZO on
flushes/minors and LZMA on majors.
> Avoid compressing flush files
> -----------------------------
>
> Key: HBASE-2987
> URL: https://issues.apache.org/jira/browse/HBASE-2987
> Project: HBase
> Issue Type: Improvement
> Reporter: Andrew Purtell
> Assignee: Andrew Purtell
> Priority: Minor
> Attachments: HBASE-2987.patch
>
>
> I've extended Hadoop compression to use the LZMA algorithm and HFile to
> provide an option for selecting it. With typical input, the LZMA algorithm
> produces 30% smaller output than GZIP at max compression (which is currently
> the best available option for HFiles) and 15% smaller output than BZIP2. I'm
> aware of the "disk is cheap" mantra but for a multi-peta-scale archival
> application, where we still want random read and random update capabilities,
> 30% less disk is a substantial cost savings. LZMA compression speed is ~1
> MB/second on a 2 GHz CPU, decompression speed is ~20 MB/second. This is 4x
> slower than BZIP2 to compress but at least 2x faster to decompress for 15%
> better results. For an archival application these properties would be
> acceptable if not for the very significant problem of flushing. Obviously the
> low throughput of the LZMA compressor means it is unsuitable for foreground
> processing. In HBase terms, it can be used for compaction but not for flush
> files.
> Attached patch, against 0.20 branch, turns off compression for flushes. This
> could be implemented as a config option, but I wonder if with the possible
> exception of LZO should we be compressing flushes at all? Any significant
> reduction in flush throughput can stall writers during periods of high write
> activity. Maybe globally disabling compression on flush flies is a good
> thing?
> I have tested this and confirmed the result is the desired behavior: 'file'
> shows flush files as uncompressed data, compacted files as compressed.
> Compaction merges files with different compression properties. LZMA provides
> rather extreme space savings over the other available options without slowing
> down writers if the regionservers are configured with enough write buffering
> to ride over the significantly lengthened compaction times.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
