[
https://issues.apache.org/jira/browse/KUDU-3011?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16984137#comment-16984137
]
Andrew Wong commented on KUDU-3011:
-----------------------------------
The difficulty with making tservers non-scannable during maintenance is that
during a rolling restart of an entire cluster, tservers will enter maintenance
at different points in time. That means that Impala might be in the middle of
scanning one replica on a tserver when all of the sudden the tserver enters
maintenance. At that point, what should Kudu do?
Another approach to consider would be to have Impala scan nodes transparently
retry a Kudu scan when it fails. If there's a way to get Impala to throw away
any partial scan of the tablet so far, that'd be one way around this. It might
be a little wasteful if we restart scanning a partition when we've already made
it most of the way through, but that seems preferable to restarting the entire
Impala query.
> Support for smooth maintenance window
> -------------------------------------
>
> Key: KUDU-3011
> URL: https://issues.apache.org/jira/browse/KUDU-3011
> Project: Kudu
> Issue Type: New Feature
> Reporter: LiFu He
> Priority: Major
>
> A scan corresponding to a tablet failure causes the entire SQL to fail on the
> common query engines, such as Impala. Though we have the fault-tolerant
> feature by "SetFaultTolerant()", Impala doesn't use it right now since that
> will make lower throughput. Thus, lots of SQL that are running will fail when
> we shutdown/reboot/upgrade the tserver. That can be scary.
> Maybe we can do some improvement in this area, for example, the tablets are
> not allowed to be scanned after the tserver is in maintenance mode
> (KUDU-2069). And for the LEADER_ONLY mode scanning, the leader role needs to
> be shifted from the maintenance tserver. Then we can shutdown the tserver
> smoothly after all the existing SQL are completed.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
