Restarting KSQL with active transient queries can leave orphaned internal topics

[big-andy-coates]

Describe the bug
It may be the case that restarting KSQL, while there are running push/transient query that uses internal changelog or repartition topics, results in the internal topics not being deleted

This has been removed on the community slack:

https://confluentcommunity.slack.com/archives/C6UJNMY67/p1574980967041200

If this is the case, it's probably because we're not close()ing transient queries on shutdown, which would clean up these internal topics. This would fix half the problem.

Of course, we may still have transient topics left around if a node crashes or is killed while there are transient queries. At the moment, a specific node would not be able to clean these up on restart as they could belong to another node. We'd need to encode a node identify into the topic name to allow a restarted node to clean up orphaned topics.

[progocz]

We could also add support for transient topics to kafka in order to properly solve the issue.

[big-andy-coates]

Did some testing. Internal topics are generally cleaned up, but occasionally not. Looks to be some kind of race condition.

Regardless, on hard-kill / crash they will not be, so need some way of cleaning up. e.g. store set
of transient queries in topic/table. Use this to clean up on restart?

[apurvam]

I'm dropping this off the milestone, as I don't think we are going to work on it: it's just been carried over blindly.

[rodesai]

I think we should consider prioritizing this in 0.14 or 0.15 as we've seen numerous leaks on our soak cluster (see #6360). I'll mark as needs-triage again so we can discuss.

[apurvam]

I think we should do this for 0.15.