On Tue, Dec 14, 2010 at 1:46 AM, Anze <[email protected]> wrote:
>
> First of all, thank you all for the answers. I appreciate it!
>
> To recap:
> - 0.20.4 is known to be "fragile"

Yes.  It had a bug that would cause deadlock.


> - upgrade to 0.89 (cdh3b3) would improve stability
> - GC should be monitored and system tuned if necessary (not sure how to do
> that - yet :)
> - memory should be at least 2GB, better 4GB+ (we can't go that far)

Yes, it'd help having more memory though that said 0.90 seems fine w/
1G heaps (caveat, the more memory you have the more cache you have and
the faster your reads will be).

> - more nodes would help with stability issues
>
> @Jonathan: yes, we are using 2 nodes that run both Hadoop (namenode, sec.
> namenode, datanodes, jobtracker, tasktrackers) and Hbase. The reason is that
> performance-wise we don't need more than that yet, but we have plans to make
> operation much larger in future. So while this is in production, it is really
> a test-case for much larger system.

Two nodes is a particularly 'bad' number.  It probably runs slower
than 1 node.  Go to 3, 4 or even 5?  Life will be rosier.


> However, Hadoop runs reliably, even under pressure (I do understand it is much
> more mature project though).

Hadoop has ten minute timeouts and retries each task up to 4 times as
opposed to HBase which has much shorter timeouts, etc.

I would expect HBase to be written with the
> mantra "any machine may fail at any time" in mind - and with error recovery in
> that spirit. In out experience, with 0.20.4 this just isn't the case (data
> loss of a few hours' worth of Put()-s is very common when it crashes).

Sorry if you picked up the wrong impression.  In 0.20.x HBase, you
will lose data.  The Hadoop it runs on does not have a working sync.


> But I
> really hope we can make it work reliably, we have put a lot of work in
> building a system around it... We'll see how it goes with 0.89 (fingers
> crossed :).
>

Please come back to the list if you have issues getting it all going.

St.Ack

Reply via email to