Fix 330: dealing with multiple python versions and pip

[tonyyang-svail]

Fix #330

[zou000]

FYI: I found using a python version manager, such as conda, pyenv, is more convenient. For example, we are using this approach in tf-plus: https://github.com/tensorflow/io/blob/master/dev/Dockerfile#L19

[tonyyang-svail]

@zou000 Thanks for the suggestion. Indeed, adding source activate tfio-dev to ~/.bashrc neatly isolate the Linux built-in Python2 and our Python3. This approach avoids the error-prone

# Make python3 as default Python interpreter
RUN ln -s /usr/bin/python3 /usr/bin/python
RUN chmod +x /usr/bin/python

[tonyyang-svail]

@zou000 enabling python env in .bashrc has certain drawbacks. For example, if you do docker run image_name bash test.sh i.e. non-interactive bash, .bashrc is skipped, so the python env is not activated. One workaround is to define the source activate in entrypoint.

[zou000]

I can think of 2 other ways:

1. Invoke bash interactively e.g. docker run image_name bash -i test.sh
2. The idea of a pyenv is very simple, it just needs to find the python binary under the env root dir, e.g. in your case, you just need to modify PATH:

ENV PATH=/miniconda/envs/sqlflow-dev/bin:$PATH

I like the PATH approach better. You can verify the result by:

docker run image_name python -c 'import sys; print(sys.path)'

[tonyyang-svail]

CI failed at

=== RUN   TestExecutorTrainAndPredictDNN
FAIL	github.com/sql-machine-learning/sqlflow/sql	0.265s

Update: TestExecutorTrainAndPredictDNN failed because go test was not able to find tensorflow Python package. Because we enable the Python virtual environment in .bashrc, but .bashrc is skipped on non-interactive bashes such as docker run image_names bash test.sh.

[hungry1526]

If we are duplicating some code in .Dev, can we group the duplicated ones and add some notes in case when they need to be synchronized?

[weiguoz]

LGTM