hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "jiraposter@reviews.apache.org (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-4192) Optimize HLog for Throughput Using Delayed RPCs
Date Fri, 12 Aug 2011 23:45:29 GMT

    [ https://issues.apache.org/jira/browse/HBASE-4192?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13084477#comment-13084477

jiraposter@reviews.apache.org commented on HBASE-4192:

This is an automatically generated e-mail. To reply, visit:

(Updated 2011-08-12 23:44:23.648920)

Review request for hbase.


Address most of Jon's feedback:

* change 'batchEntries' to 'useDelayedRpc'. Hopefully this describes the new feature better;
* revert the constructor usage in tests (where we were adding an additional null parameter);
* add a bunch of javadoc in HLog;
* document new parameters in hbase-default.xml;
* use wait/notify instead of yield when closing writer;
* unify the two differing append functions and change users of previous one;
* fix coprocessors; note that when using delayed RPCs the postWAL function is called in hlogFlush,
when the entries are added to the writer;

Still TODO:
* parameterize TestLogRolling;
* benchmark results.



1.  Add hbase.region.wal.batchentries configuration parameter.  If this is
enabled, batch entries to the HLog in a queue.
2.  Use delayed RPCs for sync requests when aggresive batching is enabled.
This frees up RPC handler threads for the duration of the sync.
3.  Pass the RPC server instance all the way to down to HLog.  This is needed
to find out the current remote call, mark it as delayed, and finally complete
it when the sync is done.
4.  Use the region read-write consistency control to avoid exposing to
RegionScanners edits which have not yet been synced.
5.  Change a few tests which directly create HRegions or HLogs.  The
rpcServers passed in are null, HLog falls back to classic RPCs when it has no
knowledge of the RPC server.
6.  Add TestBatchEntriesLogRolling, which is identical to TestLogRolling,
except that it uses aggressive batching.  I'm not sure how to add tests that
verify the same functionality but don't duplicate code, suggestion are

The new parameter is disabled by default.

This addresses bug HBASE-4192.

Diffs (updated)

  src/main/java/org/apache/hadoop/hbase/ipc/HBaseServer.java 7117bce 
  src/main/java/org/apache/hadoop/hbase/master/HMaster.java a00b93d 
  src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java 83ff7b2 
  src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java 7a917da 
  src/main/java/org/apache/hadoop/hbase/regionserver/ReadWriteConsistencyControl.java 8ec53d3

  src/main/java/org/apache/hadoop/hbase/regionserver/wal/HLog.java 887f736 
  src/main/resources/hbase-default.xml 66548ca 
  src/test/java/org/apache/hadoop/hbase/regionserver/wal/TestWALActionsListener.java dc43eb2


Diff: https://reviews.apache.org/r/1463/diff


All unit tests run with aggressive batching turned on and off.



> Optimize HLog for Throughput Using Delayed RPCs
> -----------------------------------------------
>                 Key: HBASE-4192
>                 URL: https://issues.apache.org/jira/browse/HBASE-4192
>             Project: HBase
>          Issue Type: New Feature
>          Components: wal
>    Affects Versions: 0.92.0
>            Reporter: Vlad Dogaru
>            Priority: Minor
> Introduce a new HLog configuration parameter (batchEntries) for more aggressive batching
of appends.  If this is enabled, HLog appends are not written to the HLog writer immediately,
but batched and written either periodically or when a sync is requested.  Because sync times
become larger, they use delayed RPCs to free up RPC handler threads.

This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira


View raw message