Hi Udo,I didn't read the whole thread but can you define the type of workload you're looking at? Do you have jobs that require reading the whole data stored in your database? For example one big column family that needs to be read entirely by a job? Because the amount of time required to read a whole disk (SATA II) of 1TB is roughly 2-2.5 hours. Now add RAID to this and you can modify this amount of time but your bottleneck will pretty much always be your disks. On our cluster, we currently have more than 1TB per node and it holds but we find that our sweet spot should be around 400-500GB per node.Regards,
SCOn Mon, Jul 4, 2011 at 3:01 PM, ZFabrik Subscriber <firstname.lastname@example.org> wrote:
Let's assume you have 50 nodes and their work-load grows simultaneously. You discover that the nodes are about to reach their limits (btw. what is the actual limit of a Cassandra node? 100GB? 500GB? 1TB?)You decide to add another 50 nodes. Do you do this within one step? Or one after the other? Or in several rounds, always every RF-rd node?Or you add 20 nodes and move the token ranges. Again, in one step? 20 steps? 4 steps 5 nodes each?This could take a while (in terms of days, if not weeks) in larger clusters!Does anybody has experience with real life scale-outs?RegardsUdo
Am 04.07.2011 um 16:21 schrieb Paul Loy:Well, by issuing a nodetool move when a node is under high load, you basically make that node unresponsive. That's fine, but a nodetool move on one node also means that that node's replica data needs to move around the ring and possibly some replica data from the next (or previous) node in the ring. So how does this affect other nodes wrt RF and quorum? Will quorum fail until the replicas have moved also?
On Mon, Jul 4, 2011 at 3:08 PM, Dan Hendry <email@example.com> wrote:
That's basically how I understand it.
However, I think it gets better with larger clusters as the proportion of the ring you move around at any time is much lower.
On Mon, Jul 4, 2011 at 10:54 AM, Subscriber <firstname.lastname@example.org> wrote:
I read a lot of Cassandra's high scalability feature: allowing seamless addition of nodes, no downtime etc.
But I wonder how one will do this in practice in an operational system.
In the system we're going to implement we're expecting a huge number of writes with uniformly distributed keys
(the keys are given and cannot be generated). That means using RandomPartitioner will (more or less) result in
the same work-load per node as any other OrderPreservePartitioner - right?
But how do you scale a (more or less) balanced Cassandra cluster? I think that in the end
you always have to double the number of nodes (adding just a handful of nodes disburdens only the split regions, the
work-load of untouched regions will grow with unchanged speed).
This seems to be ok for small clusters. But what do you do with when you have several 100s of nodes in your cluster?
It seems to me that a balanced cluster is a bless for performance but a curse for scalability...
What are the alternatives? One could re-distribute the token ranges, but this would cause
downtimes (AFAIK); not an option!
Is there anything that I didn't understand or do I miss something else? Is the only left strategy to make sure that
the cluster grows unbalanced so one can add nodes to the hotspots? However in this case you have to make sure
that this strategy is lasting. Could be too optimistic...
No virus found in this incoming message.
Checked by AVG - www.avg.com
Version: 9.0.901 / Virus Database: 271.1.1/3743 - Release Date: 07/04/11 02:35:00