hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stack <st...@duboce.net>
Subject Re: all regions unregistered over time.
Date Wed, 22 Sep 2010 18:06:03 GMT
Are you lzo'ing Jack?  If not, you probably should.

On Wed, Sep 22, 2010 at 3:17 AM, Jack Levin <magnito@gmail.com> wrote:
> So our cell sizes will be 350kb on average with 5-10 terabytes per server, I just want
to keep the count of Regions under 1000, per server
> -Jack
> On Sep 22, 2010, at 2:44 AM, Ryan Rawson <ryanobjc@gmail.com> wrote:
>> Region size is one of those tricky things, there are a few factors to consider:
>> - regions are the basic element of availability and distribution.
>> - HBase scales by having regions across many servers.  Thus if you
>> have 2 regions for 16GB data, on a 20 node machine you are a net loss
>> there.
>> - High region count has been known to make things slow, this is
>> getting better, but it is probably better to have 700 regions than
>> 3000 for the same amount of data.
>> - Low region count prevents parallel scalability as per point #2.
>> This really cant be stressed enough, since a common problem is loading
>> 200MB data into HBase then wondering why your awesome 10 node cluster
>> is mostly idle.
>> - There is not much memory footprint difference between 1 region and
>> 10 in terms of indexes, etc, held by the regionserver.
>> Generally speaking I stick to the default, go smaller for hot tables,
>> or manually split them, and go with a 1GB region size on our largest
>> 900 GB table.
>> -ryan
>> On Wed, Sep 22, 2010 at 12:01 AM, Jack Levin <magnito@gmail.com> wrote:
>>> Yes, I am thinking to put 10 to 15 million files on each regionserver
>>> (well, not literally, but be controlled by regionserver).   So thats
>>> close to 4 TB worth of regions, which is about 4GB per region should
>>> we target 1000 regions per server.  Note, not all files are 'hot', and
>>> I only expect to keep about 1% super hot, and 5% relatively hot, the
>>> rest are cold.  So in terms of keeping hbase blocks in RAM, that
>>> should be adequate, for the rest we can afford a trip into hdfs.
>>> If servers are running 8 GB of ram, and are shared for regionservers
>>> and datanodes, how much heap should I allocate to each?  6GB for RS
>>> and 1GB  for DN?
>>> Also, on the question whether 8 core x 16G Ram helps a Master server
>>> to bring up the cluster faster, the answer is definitely - yes.   It
>>> took only 90 seconds to load 5000 regions across 13 servers, where
>>> same task for Dual Core 8G Ram, took nearly 10 minutes.
>>> -Jack
>>> On Tue, Sep 21, 2010 at 11:38 PM, Stack <stack@duboce.net> wrote:
>>>> On Tue, Sep 21, 2010 at 11:11 PM, Jack Levin <magnito@gmail.com> wrote:
>>>>> Its definitely binary, and I can even load it in my browser but
>>>>> setting appropriate headers.  So I guess for PUT and GET via Accept:
>>>>> application/octet-stream there is no base64 encoding at all.
>>>> OK.  Good.  If it were base64'd, you'd see it.
>>>>> Btw, out of curiosity I have region max file size set to 1GB now, but
>>>>> what if I set it to say 10G or 50G?  Is their significant overhead in
>>>>> address seeking via HDFS?
>>>> You could do that.  We don't have much experience running regions of
>>>> that size.  You should for sure pre-split your table on creation if
>>>> you go this route (See HBaseAdmin API [1]).  This method is not
>>>> available in shell so you'd have to script it or write a little java
>>>> to do it).
>>>> St.Ack
>>>> 1. http://hbase.apache.org/docs/r0.89.20100726/apidocs/org/apache/hadoop/hbase/client/HBaseAdmin.html#createTable(org.apache.hadoop.hbase.HTableDescriptor,
>>>> byte[][])

View raw message