[
https://issues.apache.org/jira/browse/HIVE-20854?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16675879#comment-16675879
]
Gopal V commented on HIVE-20854:
--------------------------------
Sure, will have a clearer reading of that soon - I had a prod outage lasting
from 6:05 AM to 6:25 AM which "magically" resolved itself, which I was
debugging.
The 20 minute timing ended up being too long for a prod outage and this was
specifically a customer who was running ~800ms average queries at 60+
queries/sec, so a lot of queries went missing.
bq. So I think we should still keep the session.timeout in order of minutes.
That I agree to, but I think 2 minutes is a sane default instead of 20, since
SLAs have gone from hours to seconds since hive1.
> Sensible Defaults: Hive's Zookeeper heartbeat interval is 20 minutes, change
> to 2
> ---------------------------------------------------------------------------------
>
> Key: HIVE-20854
> URL: https://issues.apache.org/jira/browse/HIVE-20854
> Project: Hive
> Issue Type: Bug
> Reporter: Gopal V
> Assignee: Gopal V
> Priority: Major
> Attachments: HIVE-20854.1.patch
>
>
> {code}
> HIVE_ZOOKEEPER_SESSION_TIMEOUT("hive.zookeeper.session.timeout",
> "1200000ms",
> new TimeValidator(TimeUnit.MILLISECONDS),
> "ZooKeeper client's session timeout (in milliseconds). The client is
> disconnected, and as a result, all locks released, \n" +
> "if a heartbeat is not sent in the timeout."),
> {code}
> That's 1,200,000ms which is too long for all practical purposes - a 20 minute
> outage in case a node has a failure is too long.
> That is too long for the JDBC load-balancing, LLAP failure tolerance and the
> lock manager expiry.
> Change to 2 minutes, as a sensible default
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
