You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@cassandra.apache.org by Wayne Schroeder <ws...@pinsightmedia.com> on 2014/11/26 21:16:48 UTC
Never running repair: No need vs consequences in our usage pattern
I have a 30+ node cluster that is under heavy read and write load. Based on the fact that we never delete data, and all data is inserted with TTLs and is somewhat temporal if not upserted, and we are fine with the consistency of one and read repair chance, we elected to never repair. The reasoning behind this is that the data is so temporal and would simply vanish through normal compaction. We also adhere to the policy of trying to do full row writes so we do not have to do reassembly during reads. Are there any consequences we should be aware of with this strategy? We don’t even run repair when adding nodes to the cluster—we just wait for the data to invalidate itself via TTL and be compacted away.
Based on everything I’ve read, running repair only really helps us on consistency (which we don’t care about because data is updated so often that being one update behind is fine) and deleted data re-appearing (and we never delete, we just always use TTLs). Perhaps there is some other reason to run repair that we are not aware of?
Wayne
Re: Never running repair: No need vs consequences in our usage pattern
Posted by Robert Coli <rc...@eventbrite.com>.
On Wed, Nov 26, 2014 at 12:16 PM, Wayne Schroeder <
wschroeder@pinsightmedia.com> wrote:
> I have a 30+ node cluster that is under heavy read and write load. Based
> on the fact that we never delete data, and all data is inserted with TTLs
> and is somewhat temporal if not upserted, and we are fine with the
> consistency of one and read repair chance, we elected to never repair. The
> reasoning behind this is that the data is so temporal and would simply
> vanish through normal compaction. We also adhere to the policy of trying
> to do full row writes so we do not have to do reassembly during reads. Are
> there any consequences we should be aware of with this strategy? We don’t
> even run repair when adding nodes to the cluster—we just wait for the data
> to invalidate itself via TTL and be compacted away.
>
You ultimately don't care about consistency OR durability, which is the
other thing repair helps with. You are correct that you shouldn't bother
running repair in this case.
=Rob