Add CI job for oldest supported dependencies

[jrbourbeau]

This PR adds a CircleCI job to run our tests against the oldest supported version of our dependencies. Closes #456.

[stsievert]

Should the minimum required version of dask by 1.0.0 instead of 0.18.2? #455 (comment) and #458 (comment)

[jrbourbeau]

To start, I just wanted to match the minimum versions specified in setup.py (i.e. dask 0.18.2). But bumping the minimum supported version of dask to 1.0.0 would certainly be okay with me.

[jrbourbeau]

Several of the test failures here (ref https://circleci.com/gh/dask/dask-ml/3484?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link) are related to newer features in our dependencies (e.g. dd.DataFrame.to_dask_array was included in the dask 0.19.0 release, the dtype argument for pandas.get_dummies was included in the pandas 0.23.0 release, etc.)

What minimum version numbers for our dependencies do we want to support? As @stsievert pointed out, there has been mention of dask 1.0.0 as a lower bound for dask. @TomAugspurger any thoughts on a minimum version for pandas? setup.py currently has pandas 0.21.0 as the minimum (which was released in October 2017).

[TomAugspurger]

Dask 1.0 sounds good. I'm fine with a relatively new version of pandas. 0.23.4 probably makes sense.

…

On Fri, Feb 15, 2019 at 9:44 PM James Bourbeau ***@***.***> wrote: Several of the test failures here (ref https://circleci.com/gh/dask/dask-ml/3484?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link) are related to newer features in our dependencies (e.g. dd.DataFrame.to_dask_array was included in the dask 0.19.0 release, the dtype argument for pandas.get_dummies was included in the pandas 0.23.0 release, etc.) What minimum version numbers for our dependencies do we want to support? As @stsievert <https://github.com/stsievert> pointed out <#461 (comment)>, there has been mention of dask 1.0.0 as a lower bound for dask. @TomAugspurger <https://github.com/TomAugspurger> any thoughts on a minimum version for pandas? setup.py currently has pandas 0.21.0 as the minimum (which was released in October 2017). — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#461 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIu6gjmZ2s-unyPltjDKaUZvkVTITks5vN36kgaJpZM4a-Wp4> .

[TomAugspurger]

Hmm apparently the flake8 pulled from conda-forge is checking things in type comments? I'd recommend just removing those for now.

[TomAugspurger]

Unrelated failures should be fixed, if you could merge master and push again.

[jrbourbeau]

I think this should be ready for review now (cc @TomAugspurger). Summary of changes made in this PR:

• Add CircleCI build to test against earliest supported versions of our dependencies
• Update minimum versions in setup.py
• Update a few tests that were failing with older dependency versions
• Unpin NumPy version in ci/environment-2.7.yml (I suspect this could have been done in TST: Try numba RC #363)
• Remove a couple of Python 2 style type annotations that newer versions of flake8 didn't like

[TomAugspurger]

Thanks.