It's probably non-issue, sincepgi just won't exist there, but there are certain environments, such as Google AppEngine, that restrict the use of ctypes as a security risk so this wouldn't work.

Still, an interesting alternative.

Can you document this as an optional dependency athttps://github.com/matplotlib/matplotlib/blob/master/INSTALL#L212

@mdboom Just won't exist where?

For me my motivation comes from wanting to test Gtk3 on Travis, for me to test MEP22+27 on a GUI backend, but in general, probably good to do some testing... I have tried installingPyGObject from pip, but it doesn't install, not from pip on linux (see MEP27 thread)...

For other people, I don't see why others wouldn't want to actually use it. The fact they have a "how to use with matplotlib" example shows that people do want to use it.

I have no idea about C and CTypes, I excelled with the language during the university physics courses, but whenever I see it in the real world I find it a headache to read, the code always looks so tangled and messy, distributed between various c files and h files, but meh.

@mdboom Just won't exist where?

On restricted Python environments, such as Google AppEngine and other server environments that remove ctypes.

@mdboom Ahh, sorry, I completely misread that. I thought you referred to this PR as a non-issue because of the limitations ofpgi, and so were giving this PR a -1... my brain failed.

Anyway, tomorrow I shall move the tests to this PR, update the INSTALL guide, and think about how to improve this with an rcParam.

@lazka just a heads up to let you know we will build support directly in to MPL for pgi, and also an opportunity to tell you it looks great.

@jenshnielsen failing... my first guess would lie with the virtual machine not finding the library provided bygir1.2-gtk-3.0 which Travis installs as on the line it fails on, it has already importedgi.repository without error, and fails executing:

fromgi.repositoryimportGtk,Gdk,GObject,GLib

How would I validate that hypothesis? Or do you think the problem lies elsewhere...

I would setup a local ubuntu 12.04 VM and test it locally. It's much easier when you can investigate it interactively. Im guessing that it's related to the old packages in Ubuntu 12.04

One of the problems is that pgi doesn't support python 2.6 so you will need to skip the tests there or perhaps test for python version before importing pgi

Im not sure what the problem is on python 2.7 and 3.x. Perhaps missing X or a to old version of gtk3?

Yes, I think it misses X11, seethe docs onGtk.init_check().

I would have hoped we could have "shown" a window without a display device actually plugged in...

Ooh, it looks like I now have the gtk side of it fixed :).

Next fail point, a question I had earlier but didn't ask... why doesbackend_gtk3agg import the cairo backend... I thoughtAgg drew raster objects andCairo drew vector objects... so they shouldn't mix, right?

Cairo is required to blit (copy an image subset) to the window buffer in Gtk3. (Gtk2 had a separate API for that, but Gtk3 has a hard dependency on Cairo, so I guess they decided to not duplicate that functionality). Strictly speaking, it doesn't have to importbackend_cairo -- it was just the most convenient way to get a handle topycairo/cairoffi, since it's a little non-trivial to import. Though I wouldn't be opposed to moving the cairo importing stuff inbackend_cairo to its own separate helper module.

Ahh, okay, thanks, I didn't realise thatCairo existed as a hard dependency of gtk3. Still a bit odd getting a warning about the cairo backend requires when using the Agg, but it doesn't really matter...

Just about done on this. I just have to uncomment theknownfailiureif to make Travis happy for python-2.6.

Before I do that though, should I add an rcParam, I thought about naming itbackend.gi_preference. I could do it as a string, or to future proof it, a list of strings.

Also do I need to add a what's new? Or too small for that?

👍 to whats_new as we are adding support for another way to get at the
gui backend

The idea of a list of strings would be an order of preference?

On Thu, Jul 30, 2015 at 12:00 PM OceanWolfnotifications@github.com wrote:

Just about done on this. I just have to uncomment the knownfailiureif to
make Travis happy for python-2.6.

Before I do that though, should I add an rcParam, I thought about naming
it backend.gi_preference. I could do it as a string, or to future proof
it, a list of strings.

Also do I need to add a what's new? Or too small for that?

—
Reply to this email directly or view it on GitHub
#4810 (comment)
.

Yes, an order of preference, so I would set the default as['gi', 'pgi']. I would then create matching name functions to import and a for loop that can go through calling each function and if it succeeds without error it breaks out the loop, otherwise it tries the next one in the list.

It kind of feels like overkill for now, especially as I doubt any more versions will come along, but then again, it doesn't hurt, right?

Okay, changes implemented, ready for another review.

So what do people think? Ready for merging (after uncommenting the line and squashing the commits)?

I haven't been able to run a simple example withpgi

import matplotlibmatplotlib.use('GTK3Cairo')matplotlib.rcParams['backend.gi_preference'] = ['pgi']import matplotlib.pyplot as pltplt.plot([1,2,3])plt.show()

In python2.7 I get tons of

BUG: PySequence_LengthSystemError: null argument to internal routineBUG: PySequence_LengthSystemError: null argument to internal routine.........

In python3.4 I get

Segmentation fault (core dumped)

Maybe it's my fault...

@fariza Hmm, yes, I just tested on 2.7 and got a segmentation fault...

in cairo_save () from /usr/lib/x86_64-linux-gnu/libcairo.so.2

but like you that could also just occur on my system, I have a very twisted system... the problem only occurs onplt.show() which I didn't run on this system. Just tested cleanly on Travis and it passes, albeit with a cairo error...

/build/buildd/cairo-1.10.2/src/cairo.c:173: _cairo_error: Assertion `(status != CAIRO_STATUS_SUCCESS && status <= CAIRO_STATUS_LAST_STATUS)' failed

I will post an issue on the pgi github site and see what they say...

Even with this failing, I still like this PR as we test Travis, but I do have strong doubts about advertising it until pgi becomes a bit more stable. Perhaps we advertise it as "experimental".

Hmm, travis seems to fail with:

ERROR: Failure: AttributeError ('module' object has no attribute 'test_backend_gtk3')

Any idea why?

@fariza I have now fixed that SegFault, it occured becausepgi usescairocffi whereaspygobject usespycairo. In the GTK3 code we have an ugly hack (using C pointer manipulation, hence ugly hack) that converts thepycairo object to acairocffi object, naturally as we expected to have to do the conversion we didn't check first what kind of object we had, and because we used an ugly pointer-manipulation hack, it worked without error until we later came to use that code.

Right now, it seems to segfault occasionally which I assume comes from a race condition over garbage collection,pgi still exists as experimental. Turning off garbage collection seems to prevent segfaults:

importgcgc.disable()

What do people think about turning of GC for now for the GTK tests?

Closed by#9932.

Thanks for this@anntzer I didn't realise that you had reviewed this. Thanks for getting it in.