Matplotlib attempts to import PyQt4 when PyQt5 is not available

[cpcloud]

Bug report

Bug summary

• matplotlib is falling back to PyQt4 when importing PyQt5 fails.
• PyQt5 can fail to import for reasons other than not being installed. For example, if dlopening any of the transitive shared object dependencies of PyQt5 fails (happy to dive into the details of how and why this happens, if needed), ImportError is raised.

Code for reproduction

Unfortunately, this case is specific to the installation that I'm working with which is doing some unconventional things. I'm able to reproduce the bug, but not outside the environment I'm working in, and I can't share any code.

Actual outcome

matplotlib attempts to import PyQt5, which fails because QtGui cannot be loaded by dlopen. This causes matplotlib to think that PyQt5 isn't available, because failing dlopen calls raise an ImportError and subsequently matplotlib will attempt to import PyQt4.

Expected outcome

If I say my backend is Qt5Agg, that's what I mean. I'd like to know if the backend I've requested is not available, so I can take action to remedy the situation.

Matplotlib should examine the error message of the raised ImportError and reraise if it's not the standard ImportError: No module named 'PyQt5' I-haven't-installed-something-properly error message. It's very unfortunate that dlopen failures during the course of import are ImportErrors, but that ship has sailed.

Matplotlib version

• matplotlib: 2.0.0
• python: 3.5.3
• platform: linux
• I installed matplotlib through the default anaconda channel

I'm happy to put a patch up for this.

[anntzer]

I would support having qt5agg mean exactly that but 1. you need to be check the QT_API environment variable as well and 2. there may be backcompatibility issues.

[cpcloud]

The existing code is fine, except for this fallback code. I think it could go through a deprecation cycle as well to address major back compat issues

[anntzer]

Please do submit a patch if you can.

[tacaswell]

The source of this is: #5153 which is to play nice with people directly using qt_compat and the qt* backends without setting the rcparams.

I suspect the best fix here is to add (yet another) flag to this so we only fall back if we are in the list. Put in a PR because that was faster than explaining how I thought it should be fixed.