Folks -- how are folks handling the "productionalization" of their Pig
submit nodes?

For our PROD environment, I originally thought we'd just have a few VMs
from which Pig jobs would be submitted onto our cluster.  But on our 8GB
VMs, I found that we were often hitting heap OOM errors on a relatively
small set of approx. 50 analytics jobs.  As a short-term solution, we ended
up scaling these VMs horizontally, which seemed a bit messy to me, since we
have to manage which jobs are executed where.

Is this heap footprint (300-400 MB/per Pig process) consistent with your
environment?

Norbert

Norbert,

You mean 8GB memory on client side to launch Pig right? That seems
like a lot for simply spawning jobs. We use Azkaban to schedule jobs
and there are 10s of jobs spawned at once. Pig by itself should not be
so memory intensive.

Thanks,
Prashant

On Mar 21, 2012, at 6:50 AM, Norbert Burger <[EMAIL PROTECTED]> wrote:

> Folks -- how are folks handling the "productionalization" of their Pig
> submit nodes?
>
> For our PROD environment, I originally thought we'd just have a few VMs
> from which Pig jobs would be submitted onto our cluster.  But on our 8GB
> VMs, I found that we were often hitting heap OOM errors on a relatively
> small set of approx. 50 analytics jobs.  As a short-term solution, we ended
> up scaling these VMs horizontally, which seemed a bit messy to me, since we
> have to manage which jobs are executed where.
>
> Is this heap footprint (300-400 MB/per Pig process) consistent with your
> environment?
>
> Norbert

Hi Prashant -- yes, 8 GB total RAM, but we're seeing 300-400 MB heap
consumption per Pig invocation client-side.

We're also migrating soon to Azkaban, but it doesn't seem like it'd resolve
this issue, since from what I understand it simply wraps Grunt.

Norbert

On Wed, Mar 21, 2012 at 10:18 AM, Prashant Kommireddi
<[EMAIL PROTECTED]>wrote:

> Norbert,
>
> You mean 8GB memory on client side to launch Pig right? That seems
> like a lot for simply spawning jobs. We use Azkaban to schedule jobs
> and there are 10s of jobs spawned at once. Pig by itself should not be
> so memory intensive.
>
> Thanks,
> Prashant
>
> On Mar 21, 2012, at 6:50 AM, Norbert Burger <[EMAIL PROTECTED]>
> wrote:
>
> > Folks -- how are folks handling the "productionalization" of their Pig
> > submit nodes?
> >
> > For our PROD environment, I originally thought we'd just have a few VMs
> > from which Pig jobs would be submitted onto our cluster.  But on our 8GB
> > VMs, I found that we were often hitting heap OOM errors on a relatively
> > small set of approx. 50 analytics jobs.  As a short-term solution, we
> ended
> > up scaling these VMs horizontally, which seemed a bit messy to me, since
> we
> > have to manage which jobs are executed where.
> >
> > Is this heap footprint (300-400 MB/per Pig process) consistent with your
> > environment?
> >
> > Norbert
>