Regarding risks in practice:
> Libraries such as PyTorch, Numba, NumPy, and SciPy, among others, did not seamlessly align with the deferred module loading approach. These libraries often rely on import side effects and other patterns that do not play well with Lazy Imports. The order in which Python imports could change or be postponed, often led to side effects failing to register classes, functions, and operations correctly. This required painstaking troubleshooting to identify and address import cycles and discrepancies.
This isn't precisely the scenario I described above, but it is a concrete example of how deferred imports can cause issues that are difficult to debug.
Regarding performance benefits:
> At Meta, the quest for faster model training has yielded an exciting milestone: the adoption of Lazy Imports and the Python Cinder runtime. ... we’ve been able to significantly improve our model training times, as well as our overall developer experience (DevX) by adopting Lazy Imports and the Python Cinder runtime.
I don't see how. It adds a new, entirely optional syntax using a soft keyword. The semantics of existing code do not change. Yes, yes, you anticipated the objection:
> What if these imports were transitive? ... How could you be sure that adding lazy imports wouldn't break any code downstream?
I would need to see concrete examples of how this would be a realistic risk in principle. (My gut reaction is that top-level code in libraries shouldn't be doing the kinds of things that would be problematic here, in the first place. In my experience, the main thing they do at top level is just eagerly importing everything else for convenience, or to establish compatibility aliases.)
But if it were, clearly that's a breaking change, and the library bumps the major version and clients do their usual dependency version management. As you note, type hints work similarly. And "explicitly calling into typing APIs" is more common than you might think; https://pypistats.org/packages/pydantic exists pretty much to do exactly this. It didn't cause major problems.
> Import statements fundamentally have side effects, and when and how these side effects are applied will cause mysterious breakages that will keep people up for many nights.
They do have side effects that can be arbitrarily complex. But someone who opts in to changing import timing and encounters a difficult bug can just roll back the changes. It shouldn't cause extended debugging sessions unless someone really needs the benefits of the deferral. And people in that situation will have been hand-rolling their own workarounds anyway.
> Too many people in this thread hold the view of "importing {pandas, numpy, my weird module that is more tangled than an eight-player game of Twister} takes too long and I will gladly support anything that makes them faster".
I don't think they're under the impression that this necessarily makes things faster. Maybe I haven't seen the same comments you have.
Deferring imports absolutely would allow, for example, pip to do trivial tasks faster — because it could avoid importing unnecessary things at all. As things currently stand, a huge fraction of the vendored codebase will get imported pretty much no matter what. It's analogous to tree shaking, but implicitly, at runtime and without actually removing code.
Yes, this could be deferred to explicitly chosen times to get more or less the same benefit. It would also be more work.