[Swift-devel] Re: replication vs site score

[Mihael Hategan]

On Mon, 2009-04-06 at 16:11 +0000, Ben Clifford wrote:
> On Mon, 6 Apr 2009, Mihael Hategan wrote:
> 
> > I'm still not following. From what I understand, you are suggesting
> > what's already there. So either that is true and you think the current
> > scheme is not what it is, or I don't understand how your suggestion is
> > different than the current scheme. 
> 
> Its not the case, as I understand it, that replica jobs will always be run 
> after primary jobs - they will be run in the order they arrive in the job 
> queue. Jobs that Swift puts in the queue after that replication decision 
> has been made (for example, jobs that were waiting for dependent data) 
> will run after the replicas submitted before that dependent data become 
> available.
> 
>  a=p(x)
>  b=p(y)
>  c=q(a);
> 
> a, b run. eventually swift gets bored and resubmits b to the local job 
> queue. then a completes, and so c gets queued in the local job queue.
> 
> replica_of_b gets submitted to a site before c does.

I see what you're saying now.

I think scheduler priorities are not a bad idea.