This is partially a bug report and partially a notification/awareness message.

The [arrow] package is a dependency for [python-pyarrow] because pyarrow is built using the Apache Arrow libraries that are installed by [arrow] into /usr/lib (notably, /usr/lib/libarrow.so.1000 as of this writing). A very specific example is the "lib.cpython-310-x86_64-linux-gnu.so" file built (as part of [python-pyarrow]) which is dynamically linked against libarrow.so.1000 ... ("ldd /usr/lib/python3.10/site-packages/pyarrow/lib.cpython-310-x86_64-linux-gnu.so")

As a result, when [arrow] is updated/recompiled, this completely breaks [python-pyarrow] *AND* other packages that might auto-import pyarrow... Including, and perhaps most importantly, [python-pandas]. (And when I say "breaks", python dies with an illegal hardware instruction error whenever one tries to import pandas or pyarrow.)

There are two possible solutions to this that I can see:

1. Ensure that updating [arrow] requires/cascades an update of [python-pyarrow]
2. Change [python-pyarrow] to not depend on the system arrow libraries and instead bundle its own

Solution #2 is what the "pyarrow" binary package from PyPi does - it bundles the required libraries in a self-contained way that then get installed within the python pacakge directory itself. Perhaps not the most elegant solution, but it does ensure that a major part of the Python ecosystem doesn't break when a single package is updated.

Interestingly, Debian currently doesn't even package arrow (or pyarrow) at all - there is experimental support for it though.