11-14-2017 11:26 AM
We often need to restart a node to do some quick maintenance, such as reconfiguring a disk or changing an OS setting which requires a machine restart.
Also, we use Impala not only for interactive user queries but also for many of our ETL job queries, and these slightly longer queries of course die if a single node processing fragments becomes unavailable, killing the corresponding job, exposing us to data corruption. Therefore we are always forced to pause all our jobs.
I know we can gracefully decommission a node, but it can take hours to move all the dfs data out and then back in, so not worth it when trying to do a quick restart.
So is there a way (via CM, shell or API) to tell an impalad to simply stop taking new fragments in preparation for a restart? (we can also easily remove from haproxy so it doesn't take new queries as coordinator).
11-14-2017 12:17 PM
Impala currently does not support graceful node decommissioning. We're tracking work on this feature in IMPALA-1760, but we currently are not targeting it for a particular release. Unfortunately that only leaves the option of killing the daemon.