adopt conda-forge

[djkirkham]

Run Travis tests using the conda-forge channel. Will fail until #2440 is merged.

[djkirkham]

Some tests will continue to fail even if the above is merged, until a thread-safe hdf5 is available on conda-forge (see conda-forge/hdf5-feedstock#48)

[cpelley]

Some tests will continue to fail even if the above is merged, until a thread-safe hdf5 is available on conda-forge

In case this went unnoticed, there is a thread-safe build on conda-forge now.

[djkirkham]

Edit: Moved to a separate PR: #2477

Due to some changes in Numpy 1.12, some aggregator functions in np.ma return a MaskedArray instance with no mask rather than an ndarray. This causes some of our tests to fail which do an instance check for a MaskedArray.

I've modified the checking code to use np.ma.is_masked(), but this causes other tests to fail where the array being checked has always been a MaskedArray instance with no mask (thus, the reference file claims it is masked when it isn't). Commit dfa75a3 demonstrates this - see the Travis test failures. Note that because some tests contain multiple checks, not all of the checks which would fail are listed in the output.

[marqh]

this PR is ready to go, all the tests are passing and conda-forge is providing all of the testing matrix

The developers proximate to me have all met and discussed the strengths and weaknesses of this approach. We have reached a consensus that the benefits outweigh the costs

[cpelley]

Is there a reason why we are testing against python3.5 and not python3.6?

I assume there are missing dependencies for 3.6 at this time on conda-forge? Either way, It would be good detail the reasoning description of this ticket. People are trying to use iris with 3.6, see #2481.

[cpelley]

In addition, the INSTALL instructions still suggest we support python 3.4 not 3.5:
master/INSTALL

[marqh]

In addition, the INSTALL instructions still suggest we support 3.4:
https://github.com/SciTools/iris/blob/master/INSTALL

i have working environments of iris with python 3.4

this is only a minimum, i don't think that the minimum version of Python has changed to 3.5, just the version we are testing with on travis

Is there a reason why we are testing against python3.5 and not python3.6?

we chose to move from 3.4 to 3.5 to ensure dependency resolution and as a minimal change in order to adopt conda forge

we will look to move testing to 3.6, but not as part of this piece of work, that will be planned as a future activity
as you have observed, there may be further wrinkles to iron out

[cpelley]

Thanks @marqh