CI job to run exisitng libraries
honno opened this issue · 1 comments
In array-api-tests we have a job for NumPy, which runs the test suite (with expected failures piped in) on the publicly released version of NumPy with the numpy.array_api namespace.
For dataframe interchange, we do have say vaex and cudf with public releases that contain NumPy. My initial explorations of these two libraries is that they're highly subject to change, and cudf is a looks like non-starter for CI anyway due to using GPUs.
We could run develop versions on CI, but this seems pretty iffy as they're highly variable to change, and indeed pandas upstream currently breaks things in modin (upstream) and cudf.
So, running just vaex for CI could be an immediate goal at least. It'd a bit annoying right this instance as we're figuring out the test cases, and thus would have to constantly edit the inevitable xfails. Generally we'll need to think of an ergonomic way of specifying xfails.
Then when the ecosystem stabilises, we add more libraries to the pool, which would obviously be more useful.
Got a working CI job for vaex and upstream pandas 🎉 Unfortunately testing modin too is a bit tricky
Might have another stab at it later. Fixed via local mode, so modin's on CI too!