Hey Alex, with resource management I meant the Pod CPU and Memory limits
<http://kubernetes.io/docs/admin/limitrange/> which can be set in k8s
deployments. By default the pods will use all available resources on the
node it's on. In theory this could mean a pod ends up by itself on 1 node
when it consumes a lot of resources.

I agree with your remark about delegation in GCP, I tried keeping workers
small so I could run multiple on a single node, and scale the number of
workers up/down accordingly. The reason this failed for me was because I
also had these resource limits apply on scheduler and rabbitmq pods. This
resulted in SIGTERM errors, most likely due to lack of resources.

Where your ConfigMap is now populated from a settings file, I think an
override from env vars would also be very nice.

I'll definitely give this a try when I have some time!



On Mon, Dec 5, 2016 at 1:20 PM, Alex Van Boxel <[email protected]> wrote:

> Hey Koen, I tried running it with 2 but one is for my use-case good enough.
> I don't exactly know what you "in-this-context" mean by managing resources,
> but on GCloud you want to delegate as much as possible to the services
> (DataProc/DataFlow/BigQuery) and soon also other docker containers in the
> same cluster, so I like to keep the workers as light as possible. But good
> point about the max (I should make it configurable).
>
> My cluster for now is 2x4CPU for my Kubernetes cluster, but it has other
> stuff running aside from Airflow.
>
> On Mon, Dec 5, 2016 at 12:42 PM Koen Mevissen <[email protected]>
> wrote:
>
> > Nice one, thanks! I tried this a while ago, and got stuck on resource
> > issues - probably because I capped the pods max resources to hard.
> >
> > I see in the worker yaml it creates 1 worker replica, are you running it
> > with multiple workers as well? Are you managing any resources on the
> pods,
> > or you let the workers consume whatever's available on the nodes?
> >
> >
> > On Fri, Dec 2, 2016 at 10:14 PM, Chris Riccomini <[email protected]>
> > wrote:
> >
> > > Nice, thanks! :)
> > >
> > > On Fri, Dec 2, 2016 at 4:55 AM, Alex Van Boxel <[email protected]>
> wrote:
> > >
> > > > Hi all,
> > > >
> > > > I think I pulled it off to have kind of "one-execute" install of
> > Airflow
> > > on
> > > > Google Container Engine. Personally I find this my preferred setup
> > > (because
> > > > I can have a production environment and staging environment). It's
> > what I
> > > > use for our production/staging setup. I think someone else could pull
> > it
> > > > off with everything in this repo:
> > > >
> > > > https://github.com/alexvanboxel/airflow-gcp-k8s
> > > >
> > > > It has a README that should be enough to get it setup.
> > > >
> > > > Now I'll work on my sync'er so my dags will refresh as soon as I push
> > to
> > > > git :-)
> > > >
> > >
> >
> >
> >
> > --
> > Kind regards,
> > Met vriendelijke groet,
> >
> > *Koen Mevissen*
> > Principal BI Developer
> >
> >
> > *Travix Nederland B.V.*
> > Piet Heinkade 55
> > 1019 GM Amsterdam
> > The Netherlands
> >
> > T. +31 (0)20 203 3241 <+31%2020%20203%203241>
> > E: [email protected]
> > www.travix.com
> >
> > *Brands: * CheapTickets  |  Vliegwinkel  |  Vayama  |  BudgetAir  |
> >  Flugladen
> >
>



-- 
Kind regards,
Met vriendelijke groet,

*Koen Mevissen*
Principal BI Developer


*Travix Nederland B.V.*
Piet Heinkade 55
1019 GM Amsterdam
The Netherlands

T. +31 (0)20 203 3241
E: [email protected]
www.travix.com

*Brands: * CheapTickets  |  Vliegwinkel  |  Vayama  |  BudgetAir  |
 Flugladen

Reply via email to