hadoop-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael Segel <michael_se...@hotmail.com>
Subject Re: New Production Cluster Criticisms/Advice
Date Wed, 15 Aug 2012 02:42:00 GMT
Real clusters are a tad harder than the pseudo cluster.

You may want to consider EMR where you can choose between Amazon's Hadoop release (Its Apache),
MapR M3 or MapR M5


On Aug 14, 2012, at 8:36 PM, Eli Finkelshteyn <iefinkel@gmail.com> wrote:

> Hey Mohammad,
> Thanks for the reply. I've been using Hadoop and Pig for a while, and I've setup a pseudo-cluster
before. I've just never setup anything production-scale yet and wanted advice on that.
> 
> Cheers, 
> 
> On Tue, Aug 14, 2012 at 6:20 PM, Mohammad Tariq <dontariq@gmail.com> wrote:
> Hello Eli,
> 
>     If this is your first time with Hadoop then I would suggest to configure a cluster
locally just to get yourself familiar with Hadoop(a pseudo setup would do).
> 
> For your analytical stuff you can have a look at Pig, another member of the Hadoop ecosystem.
It's a dataflow language that makes analytics really easy.
> 
> As a data store Hbase would definitely be a good move.
> 
> For data aggregation, you can also have a look at Flume and Chukwa, apart from Scribe.

> 
> On Wednesday, August 15, 2012, Eli Finkelshteyn <iefinkel@gmail.com> wrote:
> > Hey Folks,
> > I'm going to be setting up my first new production cluster soon, and was hoping
to get some advice and criticism on my current plan of action. Here's my current plan:
> > Background/Requirements:
> > I'm setting this up for a start-up that's not gathering very big data yet, but will
be in the next few months (I hope, anyway). I'd like to use the cluster for a few things,
at least at first:
> > 1. logging stuff it doesn't make sense to write to a normal database (as well as
duplicates of what I am throwing in my database so I can use that stuff from HDFS later on).
Basically, just logging a ton of information I might want for analytics/model training later.
> > 2. analytics processing.
> > 3. model training (for machine learning). I'll primarily do this through Mahout.
> > 4. will probably want hbase on there as well for real time reading of some data.
I'm not married to this, and haven't played around much with hbase yet, but wanted to leave
the possibility open.
> > The Plan:
> > I'm thinking I'll set this up in Amazon. We have most of the rest of our hardware
there, and I really like the option to be able to spin up a bunch of extra workers at will
to have them train some ML model for me and then kill them off. For now, just to get things
off the ground, I'm going to setup a small 4 machine cluster (1 NameNode, 1 SecondaryNameNode/JobTracker,
2 DataNode/TaskTrackers). I'll start playing around with that setup, and will add more to
it as needed. Since everything will be puppetized, adding more machines shouldn't be too bad
(I think). I've been using Cloudera so far, and I haven't seen any good reason to switch,
so I'll use CDH4. For logging, I'll just use scribe and wind up storing stuff as lzos (a good
tutorial on the best way to do this would be awesome).
> > Thoughts?
> > Eli
> 
> -- 
> Regards,
>     Mohammad Tariq
> 
> 


Mime
View raw message