[
https://issues.apache.org/jira/browse/HIVE-12790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15086232#comment-15086232
]
Naveen Gangam commented on HIVE-12790:
--------------------------------------
[~thejas] These logs are based on Hive 0.13 and I just realized the Hive object
and SessionState are no longer associated in the latest release. Regardless,
given that a new connection to the HMS is established every time right after
the log message "Launching Job" is printed, it points to the the following code
in Driver.launchTask()
{code}
console.printInfo("Launching Job " + cxt.getCurJobNo() + " out of " +
jobs);
}
tsk.initialize(conf, plan, cxt);
TaskResult tskRes = new TaskResult();
TaskRunner tskRun = new TaskRunner(tsk, tskRes);
{code}
The last line above launches a new thread and the first message from that new
thread is
{code}
2015-12-03 04:21:08,360 INFO [Thread-405782()]: hive.metastore
(HiveMetaStoreClient.java:open(311)) - Trying to connect to metastore with URI
thrift://<anonymizedURL>:9083
{code}
So it appeared that the following codepath was creating new HMS connections
TaskRunner.run()-->SessionState.start()-->Hive.get().getMSC()
Hive.get() should return a shared instance but for some reason, is it possible
that it is closing the current and returning a new instance?
> Metastore connection leaks in HiveServer2
> -----------------------------------------
>
> Key: HIVE-12790
> URL: https://issues.apache.org/jira/browse/HIVE-12790
> Project: Hive
> Issue Type: Bug
> Components: HiveServer2
> Affects Versions: 1.1.0
> Reporter: Naveen Gangam
> Assignee: Naveen Gangam
> Attachments: HIVE-12790.patch, snippedLog.txt
>
>
> HiveServer2 keeps opening new connections to HMS each time it launches a
> task. These connections do not appear to be closed when the task completes
> thus causing a HMS connection leak. "lsof" for the HS2 process shows
> connections to port 9083.
> {code}
> 2015-12-03 04:20:56,352 INFO [HiveServer2-Background-Pool: Thread-424756()]:
> ql.Driver (SessionState.java:printInfo(558)) - Launching Job 11 out of 41
> 2015-12-03 04:20:56,354 INFO [Thread-405728()]: hive.metastore
> (HiveMetaStoreClient.java:open(311)) - Trying to connect to metastore with
> URI thrift://<anonymizedURL>:9083
> 2015-12-03 04:20:56,360 INFO [Thread-405728()]: hive.metastore
> (HiveMetaStoreClient.java:open(351)) - Opened a connection to metastore,
> current connections: 14824
> 2015-12-03 04:20:56,360 INFO [Thread-405728()]: hive.metastore
> (HiveMetaStoreClient.java:open(400)) - Connected to metastore.
> ....
> 2015-12-03 04:21:06,355 INFO [HiveServer2-Background-Pool: Thread-424756()]:
> ql.Driver (SessionState.java:printInfo(558)) - Launching Job 12 out of 41
> 2015-12-03 04:21:06,357 INFO [Thread-405756()]: hive.metastore
> (HiveMetaStoreClient.java:open(311)) - Trying to connect to metastore with
> URI thrift://<anonymizedURL>:9083
> 2015-12-03 04:21:06,362 INFO [Thread-405756()]: hive.metastore
> (HiveMetaStoreClient.java:open(351)) - Opened a connection to metastore,
> current connections: 14825
> 2015-12-03 04:21:06,362 INFO [Thread-405756()]: hive.metastore
> (HiveMetaStoreClient.java:open(400)) - Connected to metastore.
> ...
> 2015-12-03 04:21:08,357 INFO [HiveServer2-Background-Pool: Thread-424756()]:
> ql.Driver (SessionState.java:printInfo(558)) - Launching Job 13 out of 41
> 2015-12-03 04:21:08,360 INFO [Thread-405782()]: hive.metastore
> (HiveMetaStoreClient.java:open(311)) - Trying to connect to metastore with
> URI thrift://<anonymizedURL>:9083
> 2015-12-03 04:21:08,364 INFO [Thread-405782()]: hive.metastore
> (HiveMetaStoreClient.java:open(351)) - Opened a connection to metastore,
> current connections: 14826
> 2015-12-03 04:21:08,365 INFO [Thread-405782()]: hive.metastore
> (HiveMetaStoreClient.java:open(400)) - Connected to metastore.
> ...
> {code}
> The TaskRunner thread starts a new SessionState each time, which creates a
> new connection to the HMS (via Hive.get(conf).getMSC()) that is never closed.
> Even SessionState.close(), currently not being called by the TaskRunner
> thread, does not close this connection.
> Attaching a anonymized log snippet where the number of HMS connections
> reaches north of 25000+ connections.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
