hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From James Birchfield <jbirchfi...@stumbleupon.com>
Subject Re: HBase Table Row Count Optimization - A Solicitation For Help
Date Sat, 21 Sep 2013 01:21:02 GMT
Thanks Lars.  I like your time calculations much better than mine.

So this is where my inexperience is probably going to come glaring through.  And maybe the
root of all this.  I am not running the MapReduce job on a node in the cluster.  It is running
on a development server that connects remotely to the cluster.  Further more, I am not executing
the MpReduce job from the command line using the CLI as seen in many of the examples.  I am
executing them in process of a stand-alone Java process I have written.  It is simple in nature,
it simply creates an HBaseAdmin connection, list the tables and looks up the column families,
code the admin connection, then loops over the table list, and runs the following code:

public class RowCounterRunner {

	public static long countRows(String tableName) throws Exception {

		Job job = RowCounter.createSubmittableJob(
				ConfigManager.getConfiguration(), new String[]{tableName});
		boolean waitForCompletion = job.waitForCompletion(true);
		Counters counters = job.getCounters();
		Counter findCounter = counters.findCounter(hbaseadminconnection.Counters.ROWS);
		long value2 = findCounter.getValue();
		return value2;

	}
}

I sort of stumbled on to this implementation as a fairly easy way to automate the process.
 So based on your comments, and the fact that I see this in my log:

2013-09-20 23:41:05,556 INFO  [LocalJobRunner Map Task Executor #0] LocalJobRunner       
         : map

makes me think I am not taking advantage of the cluster effectively, if at all.  I do not
mind at all running the MapReduce job using the hbase/hadoop CLI, I can script that as well.
 I just thought this would work decently enough.

It does seem like it will be possible to use the Agregation coprocessor as suggested a little
earlier in this thread.  It may speed things up as well.  But either way, I need to understand
if I am losing significant performance running in the manner I am.  Which at this point sounds
like I probably am.

Birch
On Sep 20, 2013, at 6:09 PM, lars hofhansl <larsh@apache.org> wrote:

> From your numbers below you have about 26k regions, thus each region is about 545tb/26k
= 20gb. Good.
> 
> How many mappers are you running?
> And just to rule out the obvious, the M/R is running on the cluster and not locally,
right? (it will default to a local runner when it cannot use the M/R cluster).
> 
> Some back of the envelope calculations tell me that assuming 1ge network cards, the best
you can expect for 110 machines to map through this data is about 10h. (so way faster than
what you see).
> (545tb/(110*1/8gb/s) ~ 40ks ~11h)
> 
> 
> We should really add a rowcounting coprocessor to HBase and allow using it via M/R.
> 
> -- Lars
> 
> 
> 
> ________________________________
> From: James Birchfield <jbirchfield@stumbleupon.com>
> To: user@hbase.apache.org 
> Sent: Friday, September 20, 2013 5:09 PM
> Subject: Re: HBase Table Row Count Optimization - A Solicitation For Help
> 
> 
> I did not implement accurate timing, but the current table being counted has been running
for about 10 hours, and the log is estimating the map portion at 10%
> 
> 2013-09-20 23:40:24,099 INFO  [main] Job                            :  map 10% reduce
0%
> 
> So a loooong time.  Like I mentioned, we have billions, if not trillions of rows potentially.
> 
> Thanks for the feedback on the approaches I mentioned.  I was not sure if they would
have any effect overall.
> 
> I will look further into coprocessors.
> 
> Thanks!
> Birch
> On Sep 20, 2013, at 4:58 PM, Vladimir Rodionov <vrodionov@carrieriq.com> wrote:
> 
>> How long does it take for RowCounter Job for largest table to finish on your cluster?
>> 
>> Just curious.
>> 
>> On your options:
>> 
>> 1. Not worth it probably - you may overload your cluster
>> 2. Not sure this one differs from 1. Looks the same to me but more complex.
>> 3. The same as 1 and 2
>> 
>> Counting rows in efficient way can be done if you sacrifice some accuracy :
>> 
>> http://highscalability.com/blog/2012/4/5/big-data-counting-how-to-count-a-billion-distinct-objects-us.html
>> 
>> Yeah, you will need coprocessors for that.
>> 
>> Best regards,
>> Vladimir Rodionov
>> Principal Platform Engineer
>> Carrier IQ, www.carrieriq.com
>> e-mail: vrodionov@carrieriq.com
>> 
>> ________________________________________
>> From: James Birchfield [jbirchfield@stumbleupon.com]
>> Sent: Friday, September 20, 2013 3:50 PM
>> To: user@hbase.apache.org
>> Subject: Re: HBase Table Row Count Optimization - A Solicitation For Help
>> 
>> Hadoop 2.0.0-cdh4.3.1
>> 
>> HBase 0.94.6-cdh4.3.1
>> 
>> 110 servers, 0 dead, 238.2364 average load
>> 
>> Some other info, not sure if it helps or not.
>> 
>> Configured Capacity: 1295277834158080 (1.15 PB)
>> Present Capacity: 1224692609430678 (1.09 PB)
>> DFS Remaining: 624376503857152 (567.87 TB)
>> DFS Used: 600316105573526 (545.98 TB)
>> DFS Used%: 49.02%
>> Under replicated blocks: 0
>> Blocks with corrupt replicas: 1
>> Missing blocks: 0
>> 
>> It is hitting a production cluster, but I am not really sure how to calculate the
load placed on the cluster.
>> On Sep 20, 2013, at 3:19 PM, Ted Yu <yuzhihong@gmail.com> wrote:
>> 
>>> How many nodes do you have in your cluster ?
>>> 
>>> When counting rows, what other load would be placed on the cluster ?
>>> 
>>> What is the HBase version you're currently using / planning to use ?
>>> 
>>> Thanks
>>> 
>>> 
>>> On Fri, Sep 20, 2013 at 2:47 PM, James Birchfield <
>>> jbirchfield@stumbleupon.com> wrote:
>>> 
>>>>        After reading the documentation and scouring the mailing list
>>>> archives, I understand there is no real support for fast row counting in
>>>> HBase unless you build some sort of tracking logic into your code.  In our
>>>> case, we do not have such logic, and have massive amounts of data already
>>>> persisted.  I am running into the issue of very long execution of the
>>>> RowCounter MapReduce job against very large tables (multi-billion for many
>>>> is our estimate).  I understand why this issue exists and am slowly
>>>> accepting it, but I am hoping I can solicit some possible ideas to help
>>>> speed things up a little.
>>>> 
>>>>        My current task is to provide total row counts on about 600
>>>> tables, some extremely large, some not so much.  Currently, I have a
>>>> process that executes the MapRduce job in process like so:
>>>> 
>>>>                        Job job = RowCounter.createSubmittableJob(
>>>>                                        ConfigManager.getConfiguration(),
>>>> new String[]{tableName});
>>>>                        boolean waitForCompletion =
>>>> job.waitForCompletion(true);
>>>>                        Counters counters = job.getCounters();
>>>>                        Counter rowCounter =
>>>> counters.findCounter(hbaseadminconnection.Counters.ROWS);
>>>>                        return rowCounter.getValue();
>>>> 
>>>>        At the moment, each MapReduce job is executed in serial order, so
>>>> counting one table at a time.  For the current implementation of this whole
>>>> process, as it stands right now, my rough timing calculations indicate that
>>>> fully counting all the rows of these 600 tables will take anywhere between
>>>> 11 to 22 days.  This is not what I consider a desirable timeframe.
>>>> 
>>>>        I have considered three alternative approaches to speed things up.
>>>> 
>>>>        First, since the application is not heavily CPU bound, I could use
>>>> a ThreadPool and execute multiple MapReduce jobs at the same time looking
>>>> at different tables.  I have never done this, so I am unsure if this would
>>>> cause any unanticipated side effects.
>>>> 
>>>>        Second, I could distribute the processes.  I could find as many
>>>> machines that can successfully talk to the desired cluster properly, give
>>>> them a subset of tables to work on, and then combine the results post
>>>> process.
>>>> 
>>>>        Third, I could combine both the above approaches and run a
>>>> distributed set of multithreaded process to execute the MapReduce jobs in
>>>> parallel.
>>>> 
>>>>        Although it seems to have been asked and answered many times, I
>>>> will ask once again.  Without the need to change our current configurations
>>>> or restart the clusters, is there a faster approach to obtain row counts?
>>>> FYI, my cache size for the Scan is set to 1000.  I have experimented with
>>>> different numbers, but nothing made a noticeable difference.  Any advice
or
>>>> feedback would be greatly appreciated!
>>>> 
>>>> Thanks,
>>>> Birch
>> 
>> 
>> Confidentiality Notice:  The information contained in this message, including any
attachments hereto, may be confidential and is intended to be read only by the individual
or entity to whom this message is addressed. If the reader of this message is not the intended
recipient or an agent or designee of the intended recipient, please note that any review,
use, disclosure or distribution of this message or its attachments, in any form, is strictly
prohibited.  If you have received this message in error, please immediately notify the sender
and/or Notifications@carrieriq.com and delete or destroy any copy of this message and its
attachments.


Mime
View raw message