hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Marcos Ortiz <mlor...@uci.cu>
Subject Re: importing a large table
Date Sat, 31 Mar 2012 17:19:53 GMT
Well, doing some calculations, you have 18 TB of data, divided in 9200 
regions, you have approximately 2.4 GB by regions. Is this correct?

Well, my first advice is that you have to unable the automatic split 
mechanism in HBase. It better to do this manually, but you will have an
insane number on regions in short time.

The second is to enable compression (Gzip, LZO, Snappy) in all your 
HBase cluster. This brings to you less data to work, and less network

Omer, one of the Software Engineer at the LA Hadoop User Group gave a 
excellent talk about HBase called: "HBase Do's and Don'ts". I recommend 
that you should see this talk.

See the post first in the Cloudera's blog:

- Video

On 3/31/2012 5:33 AM, Rita wrote:
> I have close to 9200 regions. Is there an example I can follow? or are
> there tools to do this already?
> On Fri, Mar 30, 2012 at 10:11 AM, Marcos Ortiz <mlortiz@uci.cu
> <mailto:mlortiz@uci.cu>> wrote:
>     On 03/30/2012 04:54 AM, Rita wrote:
>>     Thanks for the responses. I am using 0.90.4-cdh3. i exported the table
>>     using hbase exporter. Yes, the previous table still exists but on a
>>     different cluster.My region servers are large, close to 12GB in size.
>     Which is the total number of your regions?
>>     I want to understand regarding Hfiles. We export the table as a series of
>>     Hfiles and then import them in?
>     Yes, The simplest way to do this is using the TableOutputFormat, but
>     if you use instead the HFileOutputFormat, the process will be more
>     efficient, because using this feature (bulk loads) will use less CPU
>     and network. With a MapReduce job, you prepare your data using the
>     HFileOutputFormat (Hadoop's TotalOrderPartitioner class in used to
>     partition the map output
>     into disjoint ranges of the key space, corresponding to the key
>     ranges of the regions in the table).
>>     What is the difference between that in the
>>     regular MR export job?
>     The main difference with regular MR jobs is the output, instead to
>     use the classic ouput formats like TextOutputFormat,
>     MultipleOutputFormat, SequenceFileOutputFormat, etc, you will use
>     the HFileOutputFormat, that is the native data file type for HBase
>     (HFile).
>>       I idea sounds good because it sounds simple on the
>>     surface :-)
>>     On Fri, Mar 30, 2012 at 12:08 AM, Stack<stack@duboce.net>  <mailto:stack@duboce.net>
>>>     On Thu, Mar 29, 2012 at 7:57 PM, Rita<rmorgan466@gmail.com>  <mailto:rmorgan466@gmail.com>
>>>>     Hello,
>>>>     I am importing a 40+ billion row table which I exported several months
>>>     ago.
>>>>     The data size is close to 18TB on hdfs (3x replication).
>>>     Does the table from back then still exist?  Or do you remember what
>>>     the key spread was like?  Could you precreate the old table?
>>>>     My problem is when I try to import it with mapreduce it takes a few days
>>>     --
>>>>     which is ok -- however when the job fails to whatever reason, I have
>>>>     restart everything. Is it possible to import the table in chunks like,
>>>>     import 1/3, 2/3, and then finally 3/3  of the table?
>>>     Yeah.  Funny how the plug gets pulled on the rack when the three day
>>>     job is at the end 95% done.
>>>>     Btw, the jobs creates close to 150k mapper jobs, thats a problem waiting
>>>     to
>>>>     happen :-)
>>>     Are you running 0.92?  If not, you should and go for bigger regions.   10G?
>>>     St.Ack
>     --
>     Marcos Luis Ortíz Valmaseda (@marcosluis2186)
>       Data Engineer at UCI
>       http://marcosluis2186.posterous.com
>     <http://www.uci.cu/>
> --
> --- Get your facts first, then you can distort them as you please.--

Marcos Luis Ortíz Valmaseda (@marcosluis2186)
  Data Engineer at UCI



View raw message