It can result in a lot of data on the node you run repair on. Where a lot means
perhaps 2 or more times more data.
My unscientific approach is to repair one CF at a time so you can watch the
disk usage and repair the smaller CF's first. After the repair compact if you
need to.
I think the amount of extra data will be related to how out of sync things
are, so once you get repair working smoothly it will be less of problem.
Cheers
-----------------
Aaron Morton
Freelance Cassandra Developer
@aaronmorton
http://www.thelastpickle.com
On 23/09/2011, at 3:04 AM, Yan Chunlu wrote:
>
> hi Aaron:
>
> could you explain more about the issue about repair make space usage going
> crazy?
>
> I am planning to upgrade my cluster from 0.7.4 to 0.8.6, which is because the
> repair never works on 0.7.4 for me.
> more specifically, CASSANDRA-2280 and CASSANDRA-2156.
>
>
> from your description, I really worried about 0.8.6 might make it worse...
>
> thanks!
>
> On Thu, Sep 22, 2011 at 7:25 AM, aaron morton <aa...@thelastpickle.com> wrote:
> How much data is on the nodes in cluster 1 and how much disk space on cluster
> 2 ? Be aware that Cassandra 0.8 has an issue where repair can go crazy and
> use a lot of space.
>
> If you are not regularly running repair I would also repair before the move.
>
> The repair after the copy is a good idea but should technically not be
> necessary. If you can practice the move watch the repair to see if much is
> transferred (check the logs). There is always a small transfer, but if you
> see data been transferred for several minutes I would investigate.
>
> When you start a repair it will repair will the other nodes it replicates
> data with. So you only need to run it every RF nodes. Start it one one, watch
> the logs to see who it talks to and then start it on the first node it does
> not talk to. And so on.
>
> Add a snapshot before the clean (repair will also snapshot before it runs)
>
> Scrub is not needed unless you are migrating or you have file errors.
>
> If your cluster is online, consider running the clean every RFth node rather
> than all at once (e.g. 1,4, 7, 10 then 2,5,8,11). It will have less impact on
> clients.
>
> Cheers
>
> -----------------
> Aaron Morton
> Freelance Cassandra Developer
> @aaronmorton
> http://www.thelastpickle.com
>
> On 22/09/2011, at 10:27 AM, Philippe wrote:
>
>> Hello,
>> We're currently running on a 3-node RF=3 cluster. Now that we have a better
>> grip on things, we want to replace it with a 12-node RF=3 cluster of
>> "smaller" servers. So I wonder what the best way to move the data to the new
>> cluster would be. I can afford to stop writing to the current cluster for
>> whatever time is necessary. Has anyone written up something on this subject ?
>>
>> My plan is the following (nodes in cluster 1 are node1.1->1.3, nodes in
>> cluster 2 are node2.1->2.12)
>> stop writing to current cluster & drain it
>> get a snapshot on each node
>> Since it's RF=3, each node should have all the data, so assuming I set the
>> tokens correctly I would move the snapshot from node1.1 to node2.1, 2.2, 2.3
>> and 2.4 then node1.2->node2.5,2.6,2.,2.8, etc. This is because the range for
>> node1.1 is now spread across 2.1->2.4
>> Run repair & clean & scrub on each node (more or less in //)
>> What do you think ?
>> Thanks
>
>
