/cc @terrytangyuan @dmitrievanthony to take a look.

Basic story: provide coverage for supported platforms.

When we provide a list of supported platforms, we are talking about several things:

1. The supported platforms when user install from python tensorflow-io PyPI.org
2. The supported platforms when user compile tensorflow-io from source.
3. The supported platforms when user install from R (CRAN)

In detail:

1. When user install tensorflow-io from PyPI.org, they could install with python 2.7, 3.4, 3.5, 3.6. Our position is that tensorflow-io support Linux platform that could run tensorflow. As such we have to build tensorflow-io exactly the same way as tensorflow for python 2.7/3.4/3.5/3.6. That is where customer-op tests comes from.
2. When user install tensorflow-io from PyPI.org, they could install it on a platform different from custom-op docker image. Mostly likely user will install tensorflow-io/tensorflow with system default python on Ubuntu (16.04 and 18.04). So we have to test the binaries of *.whl files generated from customer-op could run on Ubuntu (16.04 and 18.04) with system-default python.
3. When developer want to build tensorflow-io from source, they mostly likely will use a default Ubuntu 16.04/18.04 to compile (with system -default GCC version) so tests also have to be done on those platforms.
4. For R support, users will install tensorflow-io from PyPI.org so the binary of .whl are actually build from custom-op. So we need to test the scenario of building whl from custom-op docker image, but runs on Ubuntu + R 3.5. (@terrytangyuan may have additional input on R)

All tests passes now. I think this PR is ready for review.

@yongtang I changed the travis config for R to this recently:

This includes the following:

• language: r gives more R friendly setup and provides some built-in functionalities besides devtools::test() that are otherwise harder to set up ourselves.
• echo "options(repos = c(CRAN='http://cran.rstudio.com'))" >> ~/.Rprofile configures R to use a specific CRAN mirror
• --user travis is needed to get around some permission issues when language: r is specified
• The cache section significantly speeds up the build

Could you incorporate the above mentioned items in the PR?

@terrytangyuan The PR made changes to CI so now R is tested inside docker containers with Ubuntu 16.04 and 18.04. This is actually different from Travis CI's Ubuntu 14.04 VM and R language spec.

The reason to test inside Ubuntu 16.04 and 18.04 containers is that users are more likely to use Ubuntu 16.04/18.04 than Travis CI's Ubuntu 14.04. (The Ubuntu 14.04 will be out of support in April). Also, Travis CI still hasn't been able to support Ubuntu 18.04 today.

With tests running inside containers some of the optimizations may not be applications. I managed to have the following two in pace in containers:

• echo "options(repos = c(CRAN='http://cran.rstudio.com'))" >> ~/.Rprofile configures R to use a specific CRAN mirror
• The directory cache of $HOME/.cache/pip (by mapping inside and outside of containers) which could speeds up the build

Please take a look and see if it helps.

@yongtang Thanks. That makes sense and the changes look fine to me. I'll probably add the additional work that language: r does on Travis later besides devtools::test(). The R tests actually failed with:

Message: AttributeError: 'NoneType' object has no attribute 'video_dataset'
Detailed traceback: 
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/data/ops/dataset_ops.py", line 195, in make_one_shot_iterator
    _make_dataset.add_to_graph(ops.get_default_graph())
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/function.py", line 479, in add_to_graph
    self._create_definition_if_needed()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/function.py", line 335, in _create_definition_if_needed
    self._create_definition_if_needed_impl()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/function.py", line 344, in _create_definition_if_needed_impl
    self._capture_by_value, self._caller_device)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/function.py", line 864, in func_graph_from_py_func
    outputs = func(*func_graph.inputs)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/data/ops/dataset_ops.py", line 192, in _make_dataset
    return dataset._as_variant_tensor()  # pylint: disable=protected-access
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/data/ops/dataset_ops.py", line 2131, in _as_variant_tensor
    self._input_dataset._as_variant_tensor(),  # pylint: disable=protected-access
  File "/usr/local/lib/python2.7/dist-packages/tensorflow_io/video/python/ops/video_dataset_ops.py", line 71, in _as_variant_tensor
    return video_ops.video_dataset(self._filenames)

Looks like the exit code of R -e 'devtools::test()' wasn't captured properly.

Thanks @terrytangyuan. I updated the PR and the failed R tests on ffmpeg have been fixed. Also changed the check code to R -e 'stopifnot(all(data.frame(devtools::test())$failed == 0L))' so that the error exit code could be captured. Please take a look.