Failure with numpy 1.18.4

We’re experiencing a bug with pyxDamerauLevenshtein 1.5.3 (any version above, and including, 1.5.2, really), when using it with numpy 1.18.4 (or any version between 1.18.4 and 1.20.0).

Here’s how to reproduce (using docker):

docker run --rm -ti python:3.7.6 bash
pip install numpy==1.18.4
pip install pyxdameraulevenshtein==1.5.3
python -c 'from pyxdameraulevenshtein import damerau_levenshtein_distance'

This will produce the following error:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "__init__.pxd", line 242, in init pyxdameraulevenshtein
ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 88 from C header, got 80 from PyObject

This was working fine until at least last Friday, but started failing without any change on our part.

I’m not sure if anything can be done (seems pretty hard to debug), but if someone has the same issue, upgrading numpy to 1.20.0+ or downgrading pyxDamerauLevenshtein to 1.5.1 works.

[patatepartie]

I’m not quite sure where the incompatibility comes from.

The weird thing is that it happened all of a sudden, without anything related (versions of python, numpy or pyxdameraulevenshtein) changing. The only thing I can think of is the system packages (build-essential, git, curl, unzip, in our debian-based docker image).

Part of me just wants to remove the NumPy dependency entirely

I can only speak for our case: numpy is required by many libraries we use, so having to upgrade it is a pretty big risk. Having one less library depending on numpy lower both the risk of having to upgrade numpy and the risk of breakages when we do have to upgrade numpy.

[patatepartie]

Thank you, @gfairchild 👍🏻 , especially for your responsiveness!

I tested with the reproduction steps above (without installing numpy first), and by running our unit test: everything seems ok.

We can close this issue!