[
https://issues.apache.org/jira/browse/HBASE-6309?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13553495#comment-13553495
]
stack commented on HBASE-6309:
------------------------------
[~devaraj] So, when callback is called, you'd stick the even up in zk queue?
Then in another thread process it? Queuing an executor to do the NN work would
be less than this. A queue would be good in case server died, yes. I wonder
what else we should be queuing around split log trans? Would having a Q
violate 15.10.4 Invariant: http://hbase.apache.org/book.html#developing?
[~zjushch] That is interesting. If so, we should be able to just close this?
What you reckon DD?
> [MTTR] Do NN operations outside of the ZK EventThread in SplitLogManager
> ------------------------------------------------------------------------
>
> Key: HBASE-6309
> URL: https://issues.apache.org/jira/browse/HBASE-6309
> Project: HBase
> Issue Type: Improvement
> Affects Versions: 0.92.1, 0.94.0, 0.96.0
> Reporter: Jean-Daniel Cryans
> Assignee: Devaraj Das
> Priority: Critical
> Fix For: 0.96.0
>
>
> We found this issue during the leap second cataclysm which prompted a
> distributed splitting of all our logs.
> I saw that none of the RS were splitting after some time while the master was
> showing that it wasn't even 30% done. jstack'ing I saw this:
> {noformat}
> "main-EventThread" daemon prio=10 tid=0x00007f6ce46d8800 nid=0x5376 in
> Object.wait() [0x00007f6ce2ecb000]
> java.lang.Thread.State: WAITING (on object monitor)
> at java.lang.Object.wait(Native Method)
> at java.lang.Object.wait(Object.java:485)
> at org.apache.hadoop.ipc.Client.call(Client.java:1093)
> - locked <0x00000005fdd661a0> (a org.apache.hadoop.ipc.Client$Call)
> at org.apache.hadoop.ipc.RPC$Invoker.invoke(RPC.java:226)
> at $Proxy9.rename(Unknown Source)
> at sun.reflect.GeneratedMethodAccessor29.invoke(Unknown Source)
> at
> sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25)
> at java.lang.reflect.Method.invoke(Method.java:597)
> at
> org.apache.hadoop.io.retry.RetryInvocationHandler.invokeMethod(RetryInvocationHandler.java:82)
> at
> org.apache.hadoop.io.retry.RetryInvocationHandler.invoke(RetryInvocationHandler.java:59)
> at $Proxy9.rename(Unknown Source)
> at org.apache.hadoop.hdfs.DFSClient.rename(DFSClient.java:759)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem.rename(DistributedFileSystem.java:253)
> at
> org.apache.hadoop.hbase.regionserver.wal.HLogSplitter.moveRecoveredEditsFromTemp(HLogSplitter.java:553)
> at
> org.apache.hadoop.hbase.regionserver.wal.HLogSplitter.moveRecoveredEditsFromTemp(HLogSplitter.java:519)
> at
> org.apache.hadoop.hbase.master.SplitLogManager$1.finish(SplitLogManager.java:138)
> at
> org.apache.hadoop.hbase.master.SplitLogManager.getDataSetWatchSuccess(SplitLogManager.java:431)
> at
> org.apache.hadoop.hbase.master.SplitLogManager.access$1200(SplitLogManager.java:95)
> at
> org.apache.hadoop.hbase.master.SplitLogManager$GetDataAsyncCallback.processResult(SplitLogManager.java:1011)
> at
> org.apache.zookeeper.ClientCnxn$EventThread.processEvent(ClientCnxn.java:571)
> at
> org.apache.zookeeper.ClientCnxn$EventThread.run(ClientCnxn.java:497)
> {noformat}
> We are effectively bottlenecking on doing NN operations and whatever else is
> happening in GetDataAsyncCallback. It was so bad that on our 100 offline
> cluster it took a few hours for the master to process all the incoming ZK
> events while the actual splitting took a fraction of that time.
> I'm marking this as critical and against 0.96 but depending on how involved
> the fix is we might want to backport.
--
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
