We currently test in Travis against Python versions 3.7 and 3.8. The current "latest" version is 3.9, which was released in October 2020. It's now mature and all our dependencies are available on Python 3.9.
(a) I suggest we begin testing and supporting Python 3.9. Doing so will also help with automating PYPI release deployment via travis, as travis can only deploy what it tests. This should be uncontroversial.
(b) The current full suite of tests takes 4 hours to run on travis. If we add 3.9 and don't drop 3.7, that will increase the testing time to 6 hours. In the long run once agency production models are in use we probably won't want to drop support for older Pythons until the PSF drops them (for 3.7 that will be June 2023). But given that no one is yet maintaining a legacy production ActivitySim model, an argument can be made that we can (any maybe should) drop support for Python 3.7 now, just to keep a lid on the number of different versions we need to maintain.
(c) Should we push that theory to the limit and require Python 3.9 or better starting now?
We currently test in Travis against Python versions 3.7 and 3.8. The current "latest" version is 3.9, which was released in October 2020. It's now mature and all our dependencies are available on Python 3.9.
(a) I suggest we begin testing and supporting Python 3.9. Doing so will also help with automating PYPI release deployment via travis, as travis can only deploy what it tests. This should be uncontroversial.
(b) The current full suite of tests takes 4 hours to run on travis. If we add 3.9 and don't drop 3.7, that will increase the testing time to 6 hours. In the long run once agency production models are in use we probably won't want to drop support for older Pythons until the PSF drops them (for 3.7 that will be June 2023). But given that no one is yet maintaining a legacy production ActivitySim model, an argument can be made that we can (any maybe should) drop support for Python 3.7 now, just to keep a lid on the number of different versions we need to maintain.
(c) Should we push that theory to the limit and require Python 3.9 or better starting now?