Stack, I don't think that is my case. I am doing random reads across the namespace and the way the table is designed, they should be distributed across region servers. As I understand, rows are sorted by the key and we should design the table such that we fetch data across regions and I have tried to achieve the same. If there is something else you want me to read, please point me to it. I have read the Hbase Architecture doc and also the one Lars George has posted
I have one 2G file and other smaller ones on the cluster, but currently I am fetching data from this 2G lookup only. The number of regions is as follows: Server1: regions=41, 2G heap , also the hbase master, regionserver, namenode, tasktracker, jobtracker, datanode Server2: regions=36, 4G heap , datanode, tasktracker and regionserver Server3: regions=37 - this server gets 0 requests or 0 hitRatio, 4G heap , datanode, tasktracker and regionserver Total:114 That link mentioned that some servers have 0 hitRatio and says that is acceptable (?) , but that's for inserts- I am not sure if same applies to reads. http://search-hadoop.com/m/ESeeZ1B082l How do I confirm where the .META is hosted. Currently, I look the master log and check the machine it is hitting for .META table. My main concern is that before the upgrade to 0.20.6, .5M rows took 520 seconds (which you though was slow) on this 3-node cluster and now, after the upgrade and whatever other changes hbase/hdfs went through, it takes nearly an hour to do the same (with the same data and same rows being fetched). There is something really wrong with HDFS/Hbase here. I need help with diagnosing this. Let me know if you need any logs from me for this. I did send some logs last time. Did you get a chance to look at those? Thanks. -----Original Message----- From: saint....@gmail.com [mailto:saint....@gmail.com] On Behalf Of Stack Sent: Monday, September 06, 2010 12:04 PM To: user@hbase.apache.org Subject: Re: regionserver skew On Fri, Sep 3, 2010 at 6:22 PM, Sharma, Avani <agsha...@ebay.com> wrote: > I read on the mailing list that the region server that has .META table > handles more requests. That sounds okay, but in my case the 3rd regionserver > has 0 requests! And I feel that's what slowing down the read performance. > Also the hit ratio at the other regionserver is 87% or so. Only the one that > hosts .META has 95+% hit ratio. > Are your reads distributed across the whole namespace or are they only fetching some subset? If a subset, it can be the case that the subset is totally hosted by a single regionserver and while your test is running, its only pulling form this single server. Is that your case? (You do understand how rows are distributed on an hbase cluster?) Also, how many regions do you have? You said you have 2G of data total at one stage. That likely does not make for many regions. If so, it could also be the case that all the server that is not fielding requests may not be actually carrying data, or little data. Is this your case? St.Ack