[Sheepdog] How to look up objects in any case

MORITA Kazutaka morita.kazutaka at lab.ntt.co.jp
Fri Dec 18 03:24:02 CET 2009

On Thu, Dec 17, 2009 at 7:37 PM, Chris Webb <chris at arachsys.com> wrote:
> MORITA Kazutaka <morita.kazutaka at lab.ntt.co.jp> writes:
>> On Thu, Dec 17, 2009 at 1:08 AM, Chris Webb <chris at arachsys.com> wrote:
>> >
>> > Are you planning on the sheepdog daemon proxying the request to another
>> > node, or returning a 'please ask the following node' response? The former
>> > gives maximum simplicity in the client; the latter might improve network
>> > bandwidth consumption at the expensive of an extra piece of client
>> > complexity.
>> I'm planning on the former. I think its network bandwidth consumption is not
>> more than the latter's one if sheepdog daemon forwards the request
>> intelligently.
> Well, the other point here is that if you're reading data from another node,
> the reason you're doing it is that you should have it locally but don't...
> so we can write it to disk as well as forwarding it on the client, helping
> the rebalancing of the cluster.

Yes, all the data objects should be placed based on the current node
map, so you are quite right.

Sheepdog moves objects in the background to place them in the right
node, so write contention between the background access and the client
access would be a problem. We should carefully implement these, I


MORITA, Kazutaka

More information about the sheepdog mailing list