Greg, Thanks a lot for the education! Sincerely, Yuan On Tue, Jan 20, 2015 at 2:37 PM, Gregory Farnum <greg@xxxxxxxxxxx> wrote: > You don't need to list them anywhere for this to work. They set up the > necessary communication on their own by making use of watch-notify. > > On Mon, Jan 19, 2015 at 6:55 PM ZHOU Yuan <dunk007@xxxxxxxxx> wrote: >> >> Thanks Greg, that's a awesome feature I missed. I find some >> explanation on the watch-notify thing: >> http://www.slideshare.net/Inktank_Ceph/sweil-librados. >> >> Just want to confirm, it looks like I need to list all the RGW >> instances in ceph.conf, and then these RGW instances will >> automatically do the cache invalidation if necessary? >> >> >> Sincerely, Yuan >> >> >> On Mon, Jan 19, 2015 at 10:58 PM, Gregory Farnum <greg@xxxxxxxxxxx> wrote: >> > On Sun, Jan 18, 2015 at 6:40 PM, ZHOU Yuan <dunk007@xxxxxxxxx> wrote: >> >> Hi list, >> >> >> >> I'm trying to understand the RGW cache consistency model. My Ceph >> >> cluster has multiple RGW instances with HAProxy as the load balancer. >> >> HAProxy would choose one RGW instance to serve the request(with >> >> round-robin). >> >> The question is if RGW cache was enabled, which is the default >> >> behavior, there seem to be some cache inconsistency issue. e.g., >> >> object0 was cached in RGW-0 and RGW-1 at the same time. Sometime later >> >> it was updated from RGW-0. In this case if the next read was issued to >> >> RGW-1, the outdated cache would be served out then since RGW-1 wasn't >> >> aware of the updates. Thus the data would be inconsistent. Is this >> >> behavior expected or is there anything I missed? >> > >> > The RGW instances make use of the watch-notify primitive to keep their >> > caches consistent. It shouldn't be a problem. >> > -Greg _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |