Why those harsh `==` requirements?

[yandex]

Why those harsh `==` requirements?

[arogozhnikov]

Hi, Maximilian.

Thanks for question, this is an important point we didn't cover in the documentation.

REP is trying to incorporate two contradictory things:

1. Uniform support of rapidly evolving ML packages
2. Reproducibility for years. This means, once the code was written based on REP, we expect it to work (with the same version of REP!) years after.

We cannot guarantee that next versions of library will be back-compatible (and in practice, they aren't).
Specially this is related to pandas, it's results were changing over time when pandas communicating with matplotlib or numpy.

So we have to use fixed versions for each release.

There is nothing special about 0.14 (or any other versions we use), we will update libraries from time to time after testing that updated versions work fine (or changing necessary pieces of code).

If you need some newer versions of libraries, you can install them (and almost surely this will work).
A convenient way to have custom reproducible environment is described here.

If you feel real need in updating some dependencies, open an issue.