Core currently includes bindings for a small subset of the Arb C library. These are currently implemented by bundling a dynamically-linked library into the Core wheels and accessing its contents using ctypes. There are a couple of reasons that this isn't a good long-term solution:

• It makes Core wheels extremely large -- 40+ MB for Linux wheels, though it's possible they could be slimmed down somewhat -- and slow to build the first time due to the lengthy C compilation step. This is inconvenient for:

  • Users: slow to download, multiple versions bloat pip cache
  • PyPI: we've already had to request additional storage for the Core project
  • Core contributors: the Arb bindings are rarely updated, and less-frequent contributors likely won't ever touch them, but the C library compilation still needs to happen as part of a local developer install

  Plus, we aren't actually getting any benefit from this: new versions of Arb are added relatively infrequently, so many versions of Core will ship with the same version of Arb.

• Using ctypes this way isn't especially robust. We've already hit one bug that came about from it silently accepting the wrong number of arguments to a C function, and this approach also requires some kind of hacky platform-specific logic to actually locate the compiled library.

The solution to this first problem is relatively straightforward: split out the Arb bindings into a new package with a slower release cadence. We could do that without major changes and it would solve our immediate problems without significant risk of breaking anything.

The second problem is trickier. We could:

• Do nothing, on the theory that the Arb bindings are a relatively small piece of the codebase and they change so infrequently that switching to another way of writing bindings would create more bugs than it would prevent.
• Look at using Cython for the bindings. This should be more robust and better-supported, and might also have other benefits, but it's a fair amount of work.
• Explore what other options are available for doing things like this.