Actually, we're already doing this:

PRE_URL=https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/
pip install -f $PRE_URL --pre matplotlib

Any interest in putting a PR to the install instructions?

Oh, great news! Is that URL stable? Happy to send a PR.

Yes, the URL is stable (I know it doesn't look like it)...

Is there any plan to cull those over time?

No plan - but we could do that ...

Is it possible to have a more legible/canonical redirect somewhere?

It's canonical, but not legible I agree. I don't personally know of any free services that provide https redirects - anyone have any good ideas? It has to be https in order for pip to use the URL without cumbersome extra flags.

We use cloudflare for https on matplotlib.org ( @jkseppan took care of setting that up ). I wonder if we can use that? https://nightly.matplotlib.org ? I am not sure if that is possible given the way that we are using a custom domain with github pages...

That would be very good - if someone could configure that.

I set up a CNAME record at nightly.matplotlib.org pointing to 7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com (@tcaswell may have received a security warning about that) but that doesn't seem to be enough. I couldn't enable Cloudflare's proxying for that address (I suspect they blacklist other CDNs) but the CNAME should work anyway. Perhaps there is some additional step to take on the Rackspace side by whoever has the credentials to do that?

Rackspace instructions indicate that SSL is not going to work with CNAME records. It seems to me that the requirements of a custom domain and SSL support are not compatible with the free-of-charge solution of just adding DNS records at Cloudflare.

This would probably need a very small Heroku app (or similar) to return "302 Redirect" responses at a custom domain, pointing to the rackcdn address. But at least on Heroku the SSL support is only available on paid plans.

The error message at http://nightly.matplotlib.org seems to be an Akamai message, so if anyone has access to the Akamai (or possibly Rackspace) account used to configure the rackcdn.com subdomain, it probably needs to be set up to allow the CNAME redirection. But, again, that will not work with SSL.

Would it be possible to publish the wheel to pypi under a different name such as matplotlib-nightly? That would avoid redirect difficulties and allow users to install the nightly build without having to specify a specific URL.

I do not think that pip can express 'conflicts with' in the requirements which means you could install both matplotlib and matplotlib-nightly which would overlap and would lead to interesting bug reports!

I would much rather work with in the system (here using non-canonical repositories and the '--pre' flag) rather than work around the system (by creating a package with a slightly different name).

I realized that the redirection can be done with Cloudflare's page rules, which we can use in a limited way on the free tier. https://nightly.matplotlib.org/ now redirects to https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/ and pip install --pre -f https://nightly.matplotlib.org matplotlib installs a prerelease version. You can see the redirection target in the output, though:

Collecting matplotlib
  Downloading https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/matplotlib-2.1.1+734.gfb18b4ec1-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (13.2MB)
    100% |████████████████████████████████| 13.2MB 96kB/s
Collecting six>=1.10 (from matplotlib)
  Using cached six-1.11.0-py2.py3-none-any.whl
Collecting pytz (from matplotlib)
  Using cached pytz-2017.3-py2.py3-none-any.whl
Collecting numpy>=1.7.1 (from matplotlib)
  Downloading https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/numpy-1.15.0.dev0+20171214134454_d233e1f-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (4.7MB)
    100% |████████████████████████████████| 4.7MB 262kB/s
Collecting pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 (from matplotlib)
  Using cached pyparsing-2.2.0-py2.py3-none-any.whl
Collecting python-dateutil>=2.0 (from matplotlib)
  Using cached python_dateutil-2.6.1-py2.py3-none-any.whl
Collecting cycler>=0.10 (from matplotlib)
  Using cached cycler-0.10.0-py2.py3-none-any.whl
Installing collected packages: six, pytz, numpy, pyparsing, python-dateutil, cycler, matplotlib
Successfully installed cycler-0.10.0 matplotlib-2.1.1+734.gfb18b4ec1 numpy-1.15.0.dev0+d233e1f pyparsing-2.2.0 python-dateutil-2.6.1 pytz-2017.3 six-1.11.0

There could be some danger in using Cloudflare's services for something that could get used in automated scripts. I think Cloudflare is mainly intended for serving web pages to humans efficiently, and if they detect what looks like a distributed attack they could start serving CAPTCHA challenges or otherwise limiting the usage.

Sweet!

Somewhat fortuitously, I will be having a chat with someone from cloudflare in the near future anyway (we push enough data every month they noticed us). I'll be sure to bring this up.

Where are we on nightly wheels? Anything I can do to help?

We have moved to building the wheels via github actions (see https://github.com/matplotlib/matplotlib/actions/workflows/cibuildwheel.yml). There is a download link, but I am not sure if that is something you can easily point pip at.

Sorry - I am sure I'm being dumb but I couldn't see how to get the wheels from that interface.

But - if we can't point pip at the results - is there a way to upload them them somewhere pip can see? I think the desire is to have something we can test our individual CIs against - certainly I would like to do that.

If you click on the PR tittle you get to https://github.com/matplotlib/matplotlib/actions/runs/695524406 which has a download link at the bottom (sorry, typed this 5 hours ago and failed to hit 'comment' 🤦 ).

Aha - thanks - Firefox deceived me and I couldn't see or easily get to the bottom of the page.

So, the individual jobs generate a .zip file containing the wheels. Could we make a cron job or similar to download the zip file, unpack and upload somewhere like Anaconda.org?

See MacPython/matplotlib-wheels#15

Just a quick note that we would be very interested in nightly wheels of Matplotlib over at sunpy too - currently we test against the Matplotlib master branch.