In 79518fa the constraint of the numpy dependency was bounded to numpy < 1.19.0 from numpy < 2.0 due to breaking ABI changes from numpy/numpy#15355.

The next quarterly release of pip, 20.3, in October 2020 will change the default dependency resolution behavior to ensure that this constraint is more strictly followed. Specifically, it means that installing the latest version of TensorFlow will have the effect of asking the user to downgrade to numpy < 1.19.0 if they have numpy >= 1.19.0 installed.

It's likely that there are currently many TensorFlow users who are also using numpy >= 1.19.0 without issue, who will start to experience this behavior when the new version of pip is released. This is already being reported by users who have opted-in to the new version of the resolver (pypa/pip#8076 (comment)).

I couldn't find an issue about resetting this constraint to numpy < 2.0 and migrating to the new ABI (sorry if I missed it) so I wanted to create this to capture that, and also give you some early warning about how the change pip's behavior will affect users of this project. Thanks!