I'm a little reluctant to drop the conda_linux configuration because we previously ran into an issue that was specific to how continuum compiled scipy for linux. I do use conda environments on linux because it's an easy way for me to manage custom python environments without root access.

Is a conda_mac configuration more representative of end user environments than a bare_mac configuration? I don't think we need both.

can we spread the load over several CI's from different providers, some travis, some azure, and some appveyor or circle? what are their free offerings? Does azure offer free concurrency for oss? I'll try to answer these if useful

I've thought it would be easier to maintain the CI if we reduce to just 1 provider. They all have their quirks, but lately Azure has been my favorite. In this PR, Azure completes 12 configurations across 3 platforms in roughly the time it takes Travis to complete 4 configurations on 1 platform.

I'm a little reluctant to drop the conda_linux configuration because we previously ran into an issue that was specific to how continuum compiled scipy for linux. I do use conda environments on linux because it's an easy way for me to manage custom python environments without root access.

I'm not familiar with the severity of that issue you had before but if it was a "one-off" occurance I would advocate for a quicker CI experience than covering than for trying to cover all avenues of usability all the time. Having to wait an extended amount of time for the CI through every iteration of a PR could scare off new contributors and a drag on the frequent contributors. We can always spin up a wider scoped CI on our own branch if an issue comes up with another build that's in heavy use.

You've earned plenty of wisdom in this area so I'd defer to you from a maintainers perspective. The rest of the team may have a strong opinion one way or another as well.

Is a conda_mac configuration more representative of end user environments than a bare_mac configuration? I don't think we need both.

This may be a good question for a survey. I personally use Homebrew on Mac which shields the user from using the system Python and doesn't require sudo. I feel like conda is another layer of abstraction that's not necessary for *nix systems but worthwhile on Windows.

I don't have a strong preference between bare_mac, conda_mac, or both.

I'm not that concerned about total CI speed so long as results begin to be returned quickly. The vast majority of errors are caught within a single platform's bare (~1 min) or a single platform's conda (~4 min) configurations, and then the total build is immediately marked failed. That's faster than any of the Travis results (~6 minutes). LGTM is always the slowest (typically 30-60 minutes).

I don't have a strong preference between bare_mac, conda_mac, or both.

Let's just keep it bare_mac for now. If we hear that everyone is using conda on mac and/or there's a significant difference in bug occurrence then we can add it in.

I'm not that concerned about total CI speed so long as results begin to be returned quickly.
I wasn't aware of the "fail fast" mode. From that perspective, adding tests shouldn't be as big of a deal especially with LGTM making us wait significantly longer as well.

Thanks for the feedback!