[
https://issues.apache.org/jira/browse/HBASE-5778?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jean-Daniel Cryans updated HBASE-5778:
--------------------------------------
Attachment: HBASE-5778-0.94.patch
Attaching a first pass on making replication and HLog compression best buddies
(here against 0.94).
Most of the changes are leaks since I need the context all over the place.
The meatier part is keeping track of the context in ReplicationSource.
Basically we get a new one the first time we read the HLog then we just keep
passing it back. I made sure to set the context to null when sending
{{WALEdits}} to the sink.
The second part was managing the missing dict entries when recovering a log
with a last known position. I could think of a few solutions:
- Reset the last known position back to 0 and resend all the edits. Basically
this ignores the problem.
- Add a "fast forward" method in the code to just read the file up to the last
known position.
- Introduce new checks in order to read the log from 0 (using the normal code
path) but then skip all the entries until we get to the last known position.
I implemented the last one. It adds a lot of new things to track which I don't
like but it should be "correct".
I also added a new test which is just enabling WAL compression on
TestReplication's master cluster. Everything passes.
> Turn on WAL compression by default
> ----------------------------------
>
> Key: HBASE-5778
> URL: https://issues.apache.org/jira/browse/HBASE-5778
> Project: HBase
> Issue Type: Improvement
> Reporter: Jean-Daniel Cryans
> Assignee: Lars Hofhansl
> Priority: Blocker
> Fix For: 0.96.0
>
> Attachments: 5778.addendum, 5778-addendum.txt, HBASE-5778-0.94.patch,
> HBASE-5778.patch
>
>
> I ran some tests to verify if WAL compression should be turned on by default.
> For a use case where it's not very useful (values two order of magnitude
> bigger than the keys), the insert time wasn't different and the CPU usage 15%
> higher (150% CPU usage VS 130% when not compressing the WAL).
> When values are smaller than the keys, I saw a 38% improvement for the insert
> run time and CPU usage was 33% higher (600% CPU usage VS 450%). I'm not sure
> WAL compression accounts for all the additional CPU usage, it might just be
> that we're able to insert faster and we spend more time in the MemStore per
> second (because our MemStores are bad when they contain tens of thousands of
> values).
> Those are two extremes, but it shows that for the price of some CPU we can
> save a lot. My machines have 2 quads with HT, so I still had a lot of idle
> CPUs.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
