We have somewhat of a similar solution to what Rob mentioned. Except, we
store our docker images in quay and use the ECSOperator to run those images
on our ECS clusters. The setup works fairly smoothly.  For some of our jobs
that require much larger machines, we use the BashOperator to execute an
AWS Scale Up/Down before/after the jobs to add/remove machines from our
cluster.

-Shoumitra

On Fri, Feb 9, 2018 at 6:19 PM, Rob Goretsky <robert.goret...@gmail.com>
wrote:

> My team has solved for this with Docker.   When a developer works on a
> single project, they freeze their Python library versions via
> pip freeze > requirements.txt
> for that project, And then we build one Docker image per project, using
> something very similar to the official 'onbuild' version of the Python
> Docker image from here https://hub.docker.com/_/python/..
> We have Jenkins automatically build and push an updated image per project
> to ECR whenever code is pushed to GitHub's master branch for that project.
>
> This means we have currently 80 different Docker images (one per project)
> stored in ECR, but each one is completely isolated from each other in terms
> of their dependencies.   This means we never have to worry about the impact
> of upgrading a python library version for anything but the current project
> we're working on..  This has opened up some nice opportunities to start
> playing more with Python 3.x while keeping all of our older stuff running
> smoothly on Python 2.7.
>
> Airflow then simply calls a version of the DockerOperator each time to run
> the script/program within the project..   Working great for us!
>
> -rob
>
>
> On Mon, Feb 5, 2018 at 3:11 PM, Dennis O'Brien <den...@dennisobrien.net>
> wrote:
>
> > Hi Andrew,
> >
> > I think the issue is that each worker has a single airflow entry point
> > (what does `which airflow` point to) which has an associated environment
> > and list of packages installed, whether those are managed via conda,
> > virtualenv, or the available python environment.  So the executor would
> > need to know which environment you want to run.  I don't know how this
> > would be possible with the LocalExecutor or SequentialExecutor since both
> > are tied to the original python environment.  (Someone correct me if I am
> > wrong here.  I'm definitely not an expert on the Airflow internals.)
> >
> > The BashOperator will allow you to run any process you want, including
> any
> > Python environment, but there is some plumbing overhead required if you
> > want access to the context, etc.  The CeleryExecutor (and any of the
> > executors that support distributed workers) plus a queue gets around the
> > issue of the worker environment tied to the scheduler environment.
> >
> > That said, I don't want to discourage you from trying things out.  I am
> > sure there are some mysteries of Python that might make this possible.
> For
> > example, this project from Armin Ronacher that allows modules to use
> > different versions of available libraries.  (Warning: I wouldn't use this
> > in production.  I think it was more proof of concept.)
> > https://github.com/mitsuhiko/multiversion
> >
> > cheers,
> > Dennis
> >
> >
> >
> > On Mon, Feb 5, 2018 at 5:06 AM Andrew Maguire <andrewm4...@gmail.com>
> > wrote:
> >
> > > I am curious about similar issue. I'm wondering if we could use
> > > https://github.com/pypa/pipenv - so each dag is in a folder say and
> that
> > > folder has pipfile.lock that i think could then sort of bundle the
> > required
> > > environment into the dag code folder itself.
> > >
> > > I've not used this yet or anything but seems interesting...
> > >
> > > On Mon, Feb 5, 2018 at 7:17 AM Dennis O'Brien <den...@dennisobrien.net
> >
> > > wrote:
> > >
> > > > Thanks for the input!  I'll take a look at using queues for this.
> > > >
> > > > thanks,
> > > > Dennis
> > > >
> > > > On Tue, Jan 30, 2018 at 4:17 PM Hbw <br...@heisenbergwoodworking.com
> >
> > > > wrote:
> > > >
> > > > > Run them on different workers by using queues?
> > > > > That way different workers can have different 3rd party libs while
> > > > sharing
> > > > > the same af core.
> > > > >
> > > > > B
> > > > >
> > > > > Sent from a device with less than stellar autocorrect
> > > > >
> > > > > > On Jan 30, 2018, at 9:13 AM, Dennis O'Brien <
> > den...@dennisobrien.net
> > > >
> > > > > wrote:
> > > > > >
> > > > > > Hi All,
> > > > > >
> > > > > > I have a number of jobs that use scikit-learn for scoring
> players.
> > > > > > Occasionally I need to upgrade scikit-learn to take advantage of
> > some
> > > > new
> > > > > > features.  We have a single conda environment that specifies all
> > the
> > > > > > dependencies for Airflow as well as for all of our DAGs.  So
> > > currently
> > > > > > upgrading scikit-learn means upgrading it for all DAGs that use
> it,
> > > and
> > > > > > retraining all models for that version.  It becomes a very
> involved
> > > > task
> > > > > > and I'm hoping to find a better way.
> > > > > >
> > > > > > One option is to use BashOperator (or something that wraps
> > > > BashOperator)
> > > > > > and have bash use a specific conda environment with that version
> of
> > > > > > scikit-learn.  While simple, I don't like the idea of limiting
> task
> > > > input
> > > > > > to the command line.  Still, an option.
> > > > > >
> > > > > > Another option is the DockerOperator.  But when I asked around
> at a
> > > > > > previous Airflow Meetup, I couldn't find anyone actually using
> it.
> > > It
> > > > > also
> > > > > > adds some complexity to the build and deploy process in that now
> I
> > > have
> > > > > to
> > > > > > maintain docker images for all my environments.  Still, not
> ruling
> > it
> > > > > out.
> > > > > >
> > > > > > And the last option I can think of is just heterogeneous workers.
> > We
> > > > are
> > > > > > migrating our Airflow infrastructure to AWS ECS (from EC2) and
> plan
> > > on
> > > > > > having support for separate worker clusters, so this could
> include
> > > > > workers
> > > > > > with different conda environments.  I assume as long as a few key
> > > > > packages
> > > > > > are identical between scheduler and worker instances (airflow,
> > redis,
> > > > > > celery?) the rest can be whatever.
> > > > > >
> > > > > > Has anyone faced this problem and have some advice?  Am I missing
> > any
> > > > > > simpler options?  Any thoughts much appreciated.
> > > > > >
> > > > > > thanks,
> > > > > > Dennis
> > > > >
> > > >
> > >
> >
>

Reply via email to