Agreed with Jean-Daniel. To recap: your layout is generally sensible except you can 'overlay' the zookeeper nodes on any of the other ones, no need to use dedicated machines for them.
2012/7/6 Jean-Daniel Cryans <[email protected]> > On Fri, Jul 6, 2012 at 3:01 AM, Amlan Roy <[email protected]> wrote: > > 1. I have used Hbase-0.92.0 with Hadoop-1.0.0 for our POC. Are > these > > versions production ready? > > If you look at each website you'll see that HBase 0.92.1 is out (bug > fix release over 0.92.0) and so is Hadoop 1.0.3 (same release > concept). I'd recommend finalizing on those. > > > 2. Initial setup I have thought is: > > a. One Master machine. It will be used as Secondary Namenode also. > > b. One Namenode. It will be used as backup Master also. > > c. Three dedicated Zookeeper servers. > > d. Five machines that will be used as Datanode and Regionserver. > > Lots of waste on the ZK size of things. Considering that your cluster > will be dead if the machine that hosts the NN dies then I would > suggest you put just 1 ZK node there and use the two others for DN+RS. > The only reason at your scale to keep a separate ZK would be if you > need it for other clusters/services. > > Hope this helps, > > J-D >
