👍

The experiment in#5307 shows that there are now no text differences, even on Linux vs. Mac. There are other sorts of differences that happen when you turn the tolerance all the way down to 0, however, that I'm now trying to get to the bottom of. Some of it appears to just be random number effects that we've never noticed before with the higher tolerances.

BTW, which tests were different, out of curiosity? Is it mostly Linux vs Mac or even between Linux distros?

They mostly seem to be differences from one run to the next (all other things being equal). Some things just need an explicit random seed. The others are caused by a very deep and long standing bug in the test runner -- it's possible to write a test returning multiple figures all using the same output filename. Race conditions abound! I hope to have a test for that soon.

Sorry, I meant which tests prompted this PR specifically?

Oh -- different versions of Freetype have always been a problem since we've had our test suite. It's the whole reason we do "fuzzy" rather than exact matching. But the goal here is to nail down the version of Freetype such that we can do exact matching (and catch more bugs that way).

OK, I'd just like to know whether FreeType eventually stabilized and we could drop this once all the older distros were EOL, or whether it's a more cross-platform issue.

In testing Cartopy, I've never run into anything caused by Freetype and not changes in matplotlib, so perhaps we are all running new (or close) enough versions that it does not trigger anything.

Freetype has stabilized somewhat in recent years, but we ran into this as recently as 2014 the last time. The cross-platform and cross-distro issues still exist though -- it can be compiled with different hinting settings etc. and this is an attempt to normalize that.

@QuLogic: Just as an example that this is still a live problem, see#5305.@zLbZ generated a new set of images on Debian testing, and the version of Ubuntu on Travis (I think it's 12.04, maybe 14.04), with freetype 2.6.0 and 2.4.8 respectively, mismatched the images by as much as 20% (notably, these are text-heavy images).

#5307 is now passing (except for some PEP8 stuff), so I'm convinced this approach to use a fixed version of freetype is sound. This is ready for final review.

Ah, I had forgotten about all the configurable hinting settings in FreeType; those are quite likely to cause a sizeable difference (even if mostly imperceptible).

Is there any particular reason to use Freetype '2.5.2' According tohttp://www.freetype.org/ several security fixes has gone in since. It seems better to me to use 2.5.5 or 2.6.1?

@jenshnielsen: I started with the version that is currently on Travis, and the version many (but not all) of the current images were created with, in an attempt to avoid updating too many of the images. Alternatively, we could just move to the latest and start fresh, but that will add more stuff to the repository.

AFAIK the version on Travis is 2.4.8

fromhttps://travis-ci.org/matplotlib/matplotlib/jobs/87692893

 freetype: yes [version 2.4.8]

We could upgrade Travis from the existing 12.04 image to the "new" beta 14.04 images to get 2.5.2

This is of cause slightly more complicated because the versions in ubuntu have backported some of the security fixes from 2.5.x

@jenshnielsen: Ah -- that was my bad -- I just just for what 14.04 had rather than what Travis was actually using. Maybe it would be best to put this at the latest now just for longevity -- the idea should be to rarely, if ever, update this again, except for zero-visible-changes security updates.

@QuLogic: I've addressed all of your minor points in the latest commits here.

So, I think before merging, we should have a discussion about which version of Freetype to use here. I could be persuaded either way, I think.

Rebased.

Things look good to me. Just waiting for Travis to go green.

The plan is to backport this to 2.0.x right? In that case feel free to merge#5413

Yes. This should be merged to 2.0.x. I'll do that.

Cherry-picked to 2.0.x asd730481