Fix #825: add matplotlib to fix missing module error under docker image

[jimthompson5802]

Fix #825: ImportError: No module named matplotlib.pyplot in examples/sklearn_elasticnet_diabetes/train_diabetes.py in mlflow docker image

Overview of change:

• Added Python package matplotlib to test-requirements.txt
• Added support fo C compiler to Docker image to support install of one of matplotlib dependencies: subprocess32. pip install of this package is via source code and requires code compilation. As a result add to Dockerfile apt-get install -y build-essential command.

[jimthompson5802]

Failure in Travis CI (Job 2921) is on Python 3 keras related tests. AFAICT, these changes do not affect anything re: Keras. Two PR submitted after mine passed their CI checks. Possible one-time fluke. Trying to restart CI check by closing PR and re-opening.

[jimthompson5802]

reopening to try and restart CI tests based on this posting

[jimthompson5802]

Travis CI failed (Job 2926) again but this time on the R tests. The Python 3 Keras test passed w/o issue. Trying one more time to restart CI tests.

[jimthompson5802]

Restarting CI tests.

[dbczumar]

@jimthompson5802 Thank you for the contribution. After discussing it with the team, we feel that it may make sense to introduce an MLproject specification with a conda.yaml specification that enumerates the required dependencies, including matplotlib (just like the example you tweaked in #829).

Ideally, the Dockerfile included in the repository root is intended to include the minimal set of dependencies for developing MLflow and running existing tests, so I'd like to avoid extending the set of test requirements (which are also installed for every Travis build) as much as possible.

Would you be willing to add an MLproject with a Conda environment for this example? Users could then execute mlflow run ... inside the Docker container to achieve the desired result.

[jimthompson5802]

@dbczumar Yes I can convert to MLProject structure.

One process question: Should I do this work on this current pull request or should I close this pull request and submit a new pull request when I've the MLProject in place?

[dbczumar]

@jimthompson5802 Thanks so much! I think it's fine to perform this work in the same PR, no reason to submit a new one.

[jimthompson5802]

@dbczumar I've completed conversion to MLProject structure. Ready for your review.

This is the log for a test run:
sklearn_diabetes_mlproject_testrun.txt

[codecov-io]

Codecov Report

Merging #827 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #827   +/-   ##
=======================================
  Coverage   65.75%   65.75%           
=======================================
  Files          22       22           
  Lines         800      800           
=======================================
  Hits          526      526           
  Misses        274      274

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 820eff1...ce163fc. Read the comment docs.

[jimthompson5802]

prep to restart travis ci build

[jimthompson5802]

restart travis ci build

[dbczumar]

LGTM!

[dbczumar]

@jimthompson5802 I tried the MLproject on OSX and Ubuntu. It ran properly on Ubuntu, but I encountered a matplotlib issue on OSX. It looks like there are extra dependencies for the conda environment that must be installed on OSX and are unavailable on Linux systems. They are documented here: https://matplotlib.org/faq/osx_framework.html#conda.

I've updated the PR to divide the example into two subdirectories, one for each platform. Please let me know if this makes sense!

[jimthompson5802]

@dbczumar No problem. Glad you guys have the resources to find these platform specific issues. Thank you.