Inferred dtype property tests

[billylanchantin]

Description

We've had a few issues with the inferred dtype logic (unsurprising, it's tricky!). I thought we may benefit from some property tests moving forward.

Basic Idea

I've written a generator for dtypes and a generator for series of a specific dtype. I'm currently testing roughly the following property:

dtype |> series_of_dtype |> Series.dtype |> sub_dtype_of?(dtype)

where sub_dtype_of? deals with the fact that empty lists result in the :null dtype.

First hit

I believe I've already found a bug with these tests. I've included it and the code which generated it as a demonstration.

Considerations before merging

• More scalar types?
• Fix bugs before merging?
• How to deal with slow tests?

[josevalim]

👍

How to deal with slow tests?

Skip them by default (I think exclude: [test: :property] works) and then always include them on CI.

[billylanchantin]

Ok I think I've got the tests sorted. Remaining questions:

• Do we want to add more scalar dtypes?
  • I think sticking with just {:s, 64} is probably fine for now.
• Is the logic of sub_dtype_of?/2 correct?
  • It caught a real bug, so I'm inclined to think yes.

[josevalim]

Is the logic of sub_dtype_of?/2 correct?

It looks so to me!

[josevalim]

I am happy with the PR. Feel free to merge or ask others opinions. :)