hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Busbey <bus...@apache.org>
Subject Re: [DISCUSS] Re: Replication resiliency
Date Thu, 26 Jan 2017 20:54:43 GMT
I've noticed a few other places where we can lose a worker thread and
the RegionServer happily continues. One notable example is the worker
threads that handle syncs for the WAL. I'm generally a
fail-fast-and-loud advocate, so I like aborting when things look
wonky. I've also had to deal with a lot of pain around silent and thus
hard to see replication failures, so strong signals that the
replication system is in a bad way sound good to me atm.

Do we have worker threads that we can't safely continue without
indefinitely? Can we solve the general problem of "unhandled exception
in threads cause a RS Abort"?

As mentioned on the jira, I do worry a bit about cluster stability and
cascading failures, given the ability to have user-provided endpoints
in the replication process. Ultimately, I don't see that as different
than all the other places coprocessors can put the cluster at risk.

On Thu, Jan 26, 2017 at 2:48 PM, Sean Busbey <busbey@apache.org> wrote:
> (edited subject to ensure folks filtering for DISCUSS see this)
> On Thu, Jan 26, 2017 at 1:58 PM, Gary Helmling <ghelmling@gmail.com> wrote:
>> Over in HBASE-17381 there has been some discussion around whether an
>> unhandled exception in a ReplicationSourceWorkerThread should trigger a
>> regionserver abort.
>> The current behavior in the case of an unexpected exception in
>> ReplicationSourceWorkerThread.run() is to log a message and simply let the
>> thread die, allowing replication for this source to back up.
>> I've seen this happen in an OOME scenario, which seems like a clear case
>> where we would be better off aborting the regionserver.
>> However, in the case of any other unexpected exceptions out of the run()
>> method, how do we want to handle this?
>> I'm of the general opinion that where we would be better off aborting on
>> all unexpected exceptions, as it means that:
>> a) we have some missing error handling
>> b) failing fast raises visibility and makes it easier to add any error
>> handling that should be there
>> c) silently stopping up replication creates problems that are difficult for
>> our users to identify operationally and hard to troubleshoot.
>> However, the current behavior has been there for quite a while, and maybe
>> there are other situations or concerns I'm not seeing which would justify
>> having regionserver stability over replication stability.
>> What are folks thoughts on this?  Should the regionserver abort on all
>> unexpected exceptions in the run method or should we more narrowly focus
>> this on OOME's?

View raw message