On Mon, Aug 30, 2010 at 11:03, Apollon Oikonomopoulos <[email protected]> wrote: > Hi, > > On 11:23 Mon 23 Aug , Guido Trotter wrote: >> As iustin has noted before we need the iallocator framework to support >> this as well. This can probably be done by considering all nodes which >> can map the external storage to be "possible secondaries" for the case >> of live migration/failover (which currently doesn't call iallocator >> but will need to) and to restrict initial allocation to the "possible" >> nodes for that storage. (we can add a "iallocator changes" part to >> this design to cover this). > > This could be determined in a dynamic fashion by a dry-run "attach" > then, which would check if the node can map the storage. This check in > turn would be called on all nodes just before migration?
But that would make migration extremely expensive, both in terms of locks (it would need to lock all nodes), and in terms of performance. I'd rather we have this declared statically (with some method yet-to-be-determined). regards, iustin
