👍 Unfortunatly it doesn't work with the Trusty environments in #5730

how badly does this overlap with what you have been doing (other than trying trusty?)

with ccache + pip cache + wheelhouse it seems to go from boot -> running tests as < 2minutes.

Unfortunately the trusty environment only support caching for paid instances (at the moment?) If this speeds up things significantly I am +0.5 on doing this rather than the trusty environments. The only real gain from trusty at this stage is a more up to date Texlive (2013 vs 2009) which enables us to run the pgf/xelatex tests on Travis. As I said in that PR I wish I knew more about what Travis's long term strategy is.

Can you try adding a pip install wheel --upgrade after pip install --upgrade setuptools that should enable it to build and cache wheels of numpy on python 3.6 where the wheel is not available on the wheelhouse

I didn't read your comment carefully enough. I have not looked into any more travis work beyond the trusty stuff and silencing some warnings here and there in the tests.

@jenshnielsen I am inclined to self merge this after this round of travis passes, I want to see how much this helps master.

Feel free to do that and regard this as a review if Travis is green

Looks like installing wheel worked

  Running setup.py bdist_wheel for numpy
  Stored in directory: /home/travis/.cache/pip/wheels/09/99/27/1eae69e0d6be3d744317fe1da672278fe91ab8c15dcced4248
  Running setup.py bdist_wheel for pillow
  Stored in directory: /home/travis/.cache/pip/wheels/4a/50/3d/d5d9f8a46e67a55269632e5d2adc9c04c5254c2d708e9fbe9d
  Running setup.py bdist_wheel for snowballstemmer
  Stored in directory: /home/travis/.cache/pip/wheels/26/52/99/e0eb6b825c16715026ef290c13f4a0fbb86f1a5eab51bb10a0
  Running setup.py bdist_wheel for MarkupSafe
  Stored in directory: /home/travis/.cache/pip/wheels/94/a7/79/f79a998b64c1281cb99fa9bbd33cfc9b8b5775f438218d17a7
Successfully built numpy pillow snowballstemmer MarkupSafe

It might be worth rerunning the 3.6 job after this completes before merging to ensure that the caching of wheels works correctly and taking a note of any speed improvement for reference.

I retriggered the 3.6 job. Doing bdist_wheel it took "9 min 54 sec"

Doesn't look like any significant difference installing numpy etc takes 25 with and without wheel thanks to ccache but that is still much better than without caching approx 180 sec.

This is cool. I didn't know Travis had this feature now.

We could probably also include .cache/matplotlib not because I would expect any great speedup, but it might help us catch instances where the font cache id needs to be updated.