Currently this is the Appveyor build matrix:

  matrix:
    - PYTHON: "C:/Python27"
    - PYTHON: "C:/Python27-x64"
    - PYTHON: "C:/Python33"
    - PYTHON: "C:/Python33-x64"
    - PYTHON: "C:/Python34"
    - PYTHON: "C:/Python34-x64"
    - PYTHON: "C:/Python35"
    - PYTHON: "C:/Python35-x64"
    - PYTHON: "C:/Python36"
    - PYTHON: "C:/Python36-x64"

I'm not sure I can recall a time when we had a failure on x64 that didn't also fail on x86, and Appveyor is often the slow step in the CI (I think they don't do concurrent builds). I'm thinking we should probably cut down and re-organize the build matrix, like this:

  matrix:
    - PYTHON: "C:/Python27-x64"
    - PYTHON: "C:/Python36-x64"

    - PYTHON: "C:/Python27"
    - PYTHON: "C:/Python33"
    - PYTHON: "C:/Python34"
    - PYTHON: "C:/Python35"
    - PYTHON: "C:/Python36"

We'd be testing on all supported versions (except 3.2, which we've droped from these builds for whatever reason ages ago) on x86, and only test the latest two builds on x64, but the latest two x64 builds would be executed first.

The other option is to test x64 on all versions and only test the latest two with x86. The downside there is that I think 1. x86 is more likely to have problems than x64 and 2. you're more likely to be using x86 if you're also using an old version of Python.

I imagine when we add a binary component, we'll want to turn the full matrix back on again at least temporarily, since that's way more likely to have architecture + version corner cases.