Sam, I think using N=2 and RAID0 might be playing it a little fast and loose. You may only have 4 machines but since each has 24 disks you're looking at 4*24=96 disks. Given that one failed disk takes out an entire machine I wonder the probability of 2 disks failing on two different machines which would cause you to lose data. I think if you're set on N=2 going with RAID6 (or even RAID10, stripes of mirrors) would be a better idea as it would take 4 disk failures where half occur on one machine and half on another. You could slice this up a number of different ways but my point is that the more disks you have the greater the probability of failure.
I wouldn't run 24 nodes per server because you could then have data all on one server, not to mention your node count will be much higher which requires more overhead (e.g. ring size, TCP connections, file handles). -Ryan On Fri, Jan 13, 2012 at 4:39 AM, Sam R <[email protected]> wrote: > (not actually raccoons but I wanted a fourth R) > > We have (not yet, but we will soon) 4 servers with 24x1TB raw data drives > (OS on ssd). These servers will be providing static object storage > (specifically, photos). Performance isn't important since all the data will > be stored by a CDN sitting in front of the servers, were predicting low > read levels <1 reads per second, and reads don't need to be fast since > they'll be background jobs from our CDN (CDNs will pre-cache all data and > will only require origin reads when the object becomes invalid, if user > accesses a server without an object they will be routed to a different > server or an error page, basically users will never be waiting on our > origin servers). I've been evaluating lots of solutions and RIAK (with > commercial support) seemed the best option, but I'm not fully decided yet. > > We want N=2 redundancy on objects (CDN loses about 5% of objects per month > so we have a 3rd copy of 95% of objects), with the object on two servers. > We want servers to survive a disk failure but we don't care about copies of > objects since objects will be on another server. The object should never > have all of its copies on the same server. > > If we use RAID0, a disk failure would take out an entire server. With > RAID5/6 we reduce our avaliable capacity and have object level redundancy > on each single server (which we don't care about). The only other option I > can see is running 24 RIAK nodes per server keeping each disk separate and > putting a single nodes data on each disk, then somehow getting RIAK to make > sure it never stores the same object on the same server but I'm not sure if > that's possible. > > Any other suggestions? > > -Sam > > _______________________________________________ > riak-users mailing list > [email protected] > http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com > >
_______________________________________________ riak-users mailing list [email protected] http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
