On 07/16/2013 01:12 PM, Daniel Kinzler wrote: > Also, it proved quite useful to have a separate db server for slow, long > running > queries vs. quick queries from web based tools.
A point to remember is that our replicas do not rely on spinning rust; this does not present the same performance problems nor the same use profile. (And, indeed, doesn't have the same MTBF either). I very much doubt there is a significant gain to be had from segregating classes of queries in our environment (owing to the lack of seek latency during disjoint access to storage) and given our considerably more robust replication strategy, I don't expect reliability to be an issue. (In particular, from the experience setting this up, restoring a failed replica would take us some two days; so unless we have more than yearly catastrophic failures, that leaves us well within two nines). -- Marc _______________________________________________ Labs-l mailing list Labs-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/labs-l
