Обсуждение: Pg/CyberCluster test results
I've been testing Cybercluster (which is a modified PgCluster) ... I have two back-end databases, one load balancer, andone replicator. I've been testing failover and rebuilding a degraded cluster, and I'm finidng that it is REALLY easy forthe two back-ends to get out of sync with each other. This is very disturbing. I was wondering if anyone has experiencewith solving this problem.
I hoped that it would be easier to get the nodes back in sync but it seems that all Postgres Multi-Master solutions are not reliable at the moment. I've planed to test CyberCluster this weekend but I already suspected that this rsync solutions have some shortcomings. Sniff... It seems that we have to wait for PGCluster-II which isn't a "shared nothing" solution. Instead all files are on a shared medium like SAN or iSCSI and all instances uses this medium (similar to Oracle). Robert CG wrote: > I've been testing Cybercluster (which is a modified PgCluster) ... I have two back-end databases, one load balancer, andone replicator. I've been testing failover and rebuilding a degraded cluster, and I'm finidng that it is REALLY easy forthe two back-ends to get out of sync with each other. This is very disturbing. I was wondering if anyone has experiencewith solving this problem. > > > > >
Maybe some people haven't read this site here: http://www.postgresqldocs.org/wiki/Replication,_Clustering,_and_Connection_Pooling It gives a nice overview. Another interesting project which isn't production ready yet is Postgres-R (http://www.postgres-r.org/) Robert RW wrote: > I hoped that it would be easier to get the nodes back in sync > but it seems that all Postgres Multi-Master solutions are not > reliable at the moment. I've planed to test CyberCluster > this weekend but I already suspected that this rsync solutions > have some shortcomings. Sniff... > > It seems that we have to wait for PGCluster-II which isn't a > "shared nothing" solution. Instead all files are on a shared > medium like SAN or iSCSI and all instances uses this medium > (similar to Oracle). > > Robert > > CG wrote: >> I've been testing Cybercluster (which is a modified PgCluster) ... I >> have two back-end databases, one load balancer, and one replicator. >> I've been testing failover and rebuilding a degraded cluster, and I'm >> finidng that it is REALLY easy for the two back-ends to get out of >> sync with each other. This is very disturbing. I was wondering if >> anyone has experience with solving this problem. >> >> >> >> > >
On Fri, Aug 22, 2008 at 03:52:39PM +0200, RW wrote: > It seems that we have to wait for PGCluster-II which isn't a > "shared nothing" solution. Instead all files are on a shared > medium like SAN or iSCSI and all instances uses this medium > (similar to Oracle). That's not shared-nothing, it's shared-storage. A -- Andrew Sullivan ajs@commandprompt.com +1 503 667 4564 x104 http://www.commandprompt.com/
On Fri, Aug 22, 2008 at 06:33:50AM -0700, CG wrote: > cluster, and I'm finidng that it is REALLY easy for the two > back-ends to get out of sync with each other. When I investigated that product, I came to the conclusion that it's in the family of replication by query-dispatch. Everything in that family has this problem, and it's a fundamental limitation of the approach. A -- Andrew Sullivan ajs@commandprompt.com +1 503 667 4564 x104 http://www.commandprompt.com/