I need some direction on how we want to fix the travis_ci:

here are some notes about the build chain:

in the travis build we set several options for forcing package versions in our conda environment (cython, scipy, numpy). However, these all get overridden by pip during the python setup.py develop call, which installs new versions over top of what we have specified in order to satisfy our requirements.txt . That is why we are installing scikit-learn==0.23.1 and failing to install on the python3.5 version.

I'm guessing we just want to test that hdbscan runs across the different python3.x versions with whatever versions of the dependencies that conda thinks is best.

Also; Due to the bug in numpy we bumped our numpy>=1.16.0 #272 , which requires python3.6+ . This invalidates the second test in the travis matrix.

Suggestions:

• allow conda to figure out dependency package versions instead of passing them in manually through travis conda create --file hdbscan/requirements.txt
• move the second test case in travis to python3.6 OR revisit requirements.txt to make minimum versions lower that can actually run on python3.5