[
https://issues.apache.org/jira/browse/HBASE-4150?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13079499#comment-13079499
]
Gary Helmling commented on HBASE-4150:
--------------------------------------
Throwing a RuntimeException where it was not thrown before in a way that could
be disruptive to clients is IMO an even bigger change. Especially in this
case, this is relatively likely to occur if careful attention is not paid to
aligning pool size with thread allocations.
I also don't think it's reasonable to commit a change like this when there is
ongoing debate about the correct approach. I'm willing to be convinced of
alternatives, but I'm not convinced this is currently the correct approach.
-1 to this change.
> Potentially too many connections may be opened if ThreadLocalPool or
> RoundRobinPool is used
> -------------------------------------------------------------------------------------------
>
> Key: HBASE-4150
> URL: https://issues.apache.org/jira/browse/HBASE-4150
> Project: HBase
> Issue Type: Bug
> Reporter: Ted Yu
> Assignee: Doug Meil
> Fix For: 0.92.0
>
> Attachments: 4150-1.txt, 4150.txt
>
>
> See 'Problem with hbase.client.ipc.pool.type=threadlocal in trunk' discussion
> started by Lars George.
> From Lars Hofhansl:
> Looking at HBaseClient.getConnection(...) I see this:
> {code}
> synchronized (connections) {
> connection = connections.get(remoteId);
> if (connection == null) {
> connection = new Connection(remoteId);
> connections.put(remoteId, connection);
> }
> }
> {code}
> At the same time PoolMap.ThreadLocalPool.put is defined like this:
> {code}
> public R put(R resource) {
> R previousResource = get();
> if (previousResource == null) {
> ...
> if (poolSize.intValue() >= maxSize) {
> return null;
> }
> ...
> }
> {code}
> So... If the ThreadLocalPool reaches its capacity it always returns null and
> hence all new threads will create a
> new connection every time getConnection is called!
> I have also verified with a test program that works fine as long as the
> number of client threads (which include
> the threads in HTable's threadpool of course) is < poolsize. Once that is no
> longer the case the number of
> connections "explodes" and the program dies with OOMEs (mostly because each
> Connection is associated with
> yet another thread).
> It's not clear what should happen, though. Maybe (1) the ThreadLocalPool
> should not have a limit, or maybe
> (2) allocations past the pool size should throw an exception (i.e. there's a
> hard limit), or maybe (3) in that case
> a single connection is returned for all threads while the pool it over its
> limit or (4) we start round robin with the other
> connection in the other thread locals.
> For #1 means that the number of client threads needs to be more carefully
> managed by the client app.
> In this case it would also be somewhat pointless that Connection have their
> own threads, we just pass stuff
> between threads.
> #2 would work, but puts more logic in the client.
> #3 would lead to hard to debug performance issues.
> And #4 is messy :)
> From Ted Yu:
> For HBaseClient, at least the javadoc doesn't match:
> {code}
> * @param config configuration
> * @return either a {@link PoolType#Reusable} or {@link
> PoolType#ThreadLocal}
> */
> private static PoolType getPoolType(Configuration config) {
> return PoolType.valueOf(config.get(HConstants.HBASE_CLIENT_IPC_POOL_TYPE),
> PoolType.RoundRobin, PoolType.ThreadLocal);
> {code}
> I think for RoundRobinPool, we shouldn't allow maxSize to be
> Integer#MAX_VALUE. Otherwise connection explosion described by Lars may incur.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
