@Nick: In my initial production release, I dont think we should have compaction pain since we wont be doing heavy real-time writes. Now, i m thinking of using 1-2 spindle out of 10 spindle for temp dir of Hadoop and HBase. My cluster will only have 2-4 Map slots per node. Thanks for you valuable input.
PS: corrected the typo of "Roll" to "Role" On Mon, Dec 17, 2012 at 5:46 PM, Nick Dimiduk <[email protected]> wrote: > On Mon, Dec 17, 2012 at 5:20 PM, anil gupta <[email protected]> wrote: > > > @Nick: I am using HBase 0.92.1, CompactionTool.java is part of HBase 0.96 > > as per https://issues.apache.org/jira/browse/HBASE-7253. > > > > Fair enough; I grepped against trunk. > > I have 10 disks on my slave node that will primarily be used for serving > > HBase queries(very less MR ). So, i was trying to distribute my disk I/O > > load evenly among the disk. Will it be fine if i just dedicate 1 disk for > > hadoop.tmp.dir or 1 disk is also a overkill for hbase.tmp.dir. > > > > The dedicated IO could help to alleviate compaction pain -- the question > is, will you experience compaction pain? Does your workload include > frequent mutations (Puts, Deletes)? If the answer is 'no' (as your > description above implies), you'll likely not benefit very much from the > dedicated platter; better use of the spindle will likely be for the > DataNode. You can probably co-locate tmp.dir with a low-intensity resource. > Then again, if you have 10 drives, why not? > -- Thanks & Regards, Anil Gupta
