On Tue, Dec 14, 2010 at 5:28 PM, Igor Ranitovic <[email protected]> wrote: > Hi Stack, > > We have been running a small cluster (name node + 5 rs) on 0.20.3 for a long > time now. We are currently at 1100 regions per RS. As far as I can tell, I > have not seen any problems or changes in behavior due this. >
Thats a bunch Igor. If one column family only per region, and allowing an average of 2 or so store files per family, thats: 5 * 1100 * 2 files open on a pretty small hdfs cluster (HBase opens files on deploy and keeps them open while the cluster is up). If less regions, the number of storefiles will be less overall if they are allowed grow bigger (but files are indexed so lookups should be relatively the same though some more memory consumed since index is kept in memory). > What kind of problems can I expect with 1K+ regions per RS? What is a > consequence of upping region size from 256M to let's 512M. > Go to 1G I'd say Igor. You'd up the file size and probably the flush size to by 4x. You should upgrade to 0.20.6 too boss. St.Ack
