[
https://issues.apache.org/jira/browse/HBASE-16117?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15370959#comment-15370959
]
Jonathan Hsieh commented on HBASE-16117:
----------------------------------------
For the first concern, it is subtle and a good catch. The original version is
two nested try blocks. The isMasterRunning is already in the first one
(ClusterConnection). If isMasterRunning fails it turns into a ZK exception
which is not what we want. Let me tweak this and see where it leads.
For the second about adding IOE in registry -- the problem was zk failures
would get swallowed and never surface. With the unit test there was no way of
telling if that we have a bad hbase connection (e.g. we run out of zk
connections and then get a connection that doesn't fail fast). In order to
surface this kind of error at initialization time we need to have the exception
come out. An alternative that wouldn't change interface is to convert to a
runtime exception and then inspect to reconvert back which seems gross IMO.
> Fix Connection leak in mapred.TableOutputFormat
> ------------------------------------------------
>
> Key: HBASE-16117
> URL: https://issues.apache.org/jira/browse/HBASE-16117
> Project: HBase
> Issue Type: Bug
> Components: mapreduce
> Affects Versions: 2.0.0, 1.3.0, 1.2.2, 1.1.6
> Reporter: Jonathan Hsieh
> Assignee: Jonathan Hsieh
> Fix For: 2.0.0, 1.3.0, 1.1.6, 1.2.3
>
> Attachments: hbase-16117.branch-1.patch, hbase-16117.patch,
> hbase-16117.v2.branch-1.patch, hbase-16117.v2.patch,
> hbase-16117.v3.branch-1.patch, hbase-16117.v3.patch, hbase-16117.v4.patch
>
>
> Spark seems to instantiate multiple instances of output formats within a
> single process. When mapred.TableOutputFormat (not
> mapreduce.TableOutputFormat) is used, this may cause connection leaks that
> slowly exhaust the cluster's zk connections.
> This patch fixes that.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
