Snakes in a Buildpack: Deploying Data Science in Cloud Foundry

Ian Huston & James Wen, Pivotal

Presented first at CF Summit Frankfurt 2016

The Cloud Foundry Buildpacks team set out on a journey in February to go where no official buildpack has gone before: Data Science. Working with those most elusive creatures, data scientists, we toiled in the deepest darkest crevices of our stacks and buildpacks to find a way to easily support the PyData ecosystem. We discovered a package manager called Conda and migrated it to the buildpacks’ zoo, where it now lives in glorious captivity.

Join us as we share our story with you! From idea inception to experiment after experiment after experiment after… you get the idea. We’ll share our initial working hypothesis and subsequent surprising discoveries, dive into the innards of buildpacks and stacks and detail what it was like working with the Pivotal Data Science team as our first customer. Finally, we’ll `cf push` a python data science app with no vendored dependencies!

About Ian Huston
Ian Huston is a Senior Data Scientist at Pivotal Labs. Before joining Pivotal, Ian used Python to create and destroy baby universes inside numerical simulations of the inflationary phase of cosmology.

About James Wen
James Wen is the Team Lead (Anchor) of the Cloud Foundry Buildpacks team at Pivotal in New York City. He currently maintains and works on the Cloud Foundry system buildpacks, buildpack tooling, stacks, and the extensive automation behind it all. In his free time, James serves as a core maintainer and contributor to Bundler and also loves to rock climb, whether on plastic or real rock.

Transcript

1. Snakes in a Buildpack
   By Ian Huston and James Wen
   

   View Slide

2. Ian has a problem
   Hard to deploy apps using great Python data science packages
   

   View Slide

3. Why is this important?
   #productsNOTpowerpoints
   

   View Slide

4. Current Solution
   Community Buildpack:
   https://github.com/ihuston/python-conda-buildpack
   Not tested
   No CI pipeline
   No time for maintenance
   

   View Slide

5. Other Attempted Solutions
   Kenneth Reitz Buildpack
   ● https://github.com/kennethreitz/conda-buildpack
   ● Looks unmaintained, Heroku-specific, didn’t
   work with CF
   Continuum’s Conda buildpack
   ● https://github.com/conda/conda-buildpack
   ● No updates for 2 years, Heroku specific,
   would need downstream changes
   Just use Pip!
   ● https://jakevdp.github.io/blog/2016/08/25/co
   nda-myths-and-misconceptions/
   Docker?
   ● Don’t want responsibility of maintaining OS
   layer!
   

   View Slide

6. Containers vs Buildpacks
   runtime layer
   OS image
   application layer
   Container (e.g. Docker)
   system brings
   fixed host OS
   Kernel
   * Devs may bring a custom
   buildpack
   runtime layer*
   OS image
   application layer
   Buildpack
   App container
   System Provides
   Dev Provides
   system brings
   fixed host OS
   Kernel
   

   View Slide

7. Experiments
   - Use miniconda in the CF python buildpack
   - Vendor anaconda
   - Add data science dependencies to the rootfs
   Solution: Port over Ian’s conda buildpack and maintain it as a code
   path in the current python buildpack.
   

   View Slide

8. Improvement/Fixes needed after
   Lock miniconda to a version instead of
   downloading latest
   Miniconda was always installing python 3
   Miniconda needed to save and load from
   natural app cache
   Suppress miniconda progress bar output
   (massive staging logs)
   

   View Slide

9. What can we do now?
   Predict time to delivery for an
   international courier company
   Order the right components
   for an European car
   manufacturer
   Deliver warnings of
   dangerous road conditions
   while you drive
   

   View Slide

10. Hobbyist: OMSCS Machine Learning Course
    - Taking a Machine Learning class for the GTech OMSCS program
    - HW Assignment = Applying 5 supervised machine learning models to 2 large data sets
    

    View Slide

11. Hobbyist: OMSCS Machine Learning Course
    - For the experiments/analyses, using Pivotal Web Services > running on my dinky 4 year old macbook
    - Ran the experiments by wrapping them in minimal Flask apps
    

    View Slide

12. Future of Conda in the Python Buildpack
    Your Feedback?
    - Can’t run this functionality in air-gapped environment
    - Conda cannot vendor your packages (i.e. something
    like `bundle package`)
    - Reduce size of end droplet?
    - Apps using conda often end up large due to slew of
    dependencies
    Vendor miniconda (cached buildpack)
    

    View Slide

13. Where can I get it?
    Official CF Python Buildpack from v1.5.6:
    https://github.com/cloudfoundry/python-buildpack
    

    View Slide

14. Getting your wonderful ideas into Cloud Foundry
    - Open up discussion on cf-dev mailing list - Open up discussion on open source Cloud Foundry slack
    - Do the CF Dojo program and gain the skills to
    work on CF full-time (or just join Pivotal)
    - Create a feature narrative and get feedback
    

    View Slide

15. Any Questions?
    James Wen @rochesterinnyc
    Ian Huston @ianhuston
    https://github.com/cloudfoundry/python-buildpack
    

    View Slide