Thanks Bryan and Jeremy, good info. When you say "replication", you are referring to HDFS block replication, not HBase cluster/table replication, correct?
Thanks, Otis -- Search Analytics - http://sematext.com/search-analytics/index.html HBASE Performance Monitoring - http://sematext.com/spm/index.html On Mon, May 6, 2013 at 5:44 PM, Bryan Beaudreault <[email protected]> wrote: > We run multiple HBase clusters of various sizes, each is spread over 4 AZs. > Yes, there is a definite write latency added due to replication, and yes > there is a transfer cost that is not negligible but I can't comment on the > specifics of. > > Despite that though we use HBase as the primary datastore for multiple > online APIs and the sink for many hadoop jobs, and are able to keep total > response times plenty low for our apps. Of course this is very specific to > our read/write patterns, etc, so YMMV. > > > > > On Mon, May 6, 2013 at 4:59 PM, Jeremy Carroll <[email protected]> wrote: > >> We tried it. It did not work well. According to AWS, cross AZ latency >> should be ~1ms. In practice we have seen large variances. Also if you are >> using larger instances with placement groups, they do not work cross AZ as >> well. >> >> With this in mind we went the multi-cluster route, and kept all nodes for a >> cluster in 1 AZ. >> >> >> On Mon, May 6, 2013 at 10:37 AM, Otis Gospodnetic < >> [email protected]> wrote: >> >> > Hi, >> > >> > Do people spread HBase clusters over multiple EC2 Availability Zones, >> > or is that a big no-no? Cross-AZ communication is supposedly going >> > over fast fiber optic cables. >> > >> > Thanks, >> > Otis >> > -- >> > Search Analytics - http://sematext.com/search-analytics/index.html >> > HBASE Performance Monitoring - http://sematext.com/spm/index.html >> > >>
