Having to return `NotImplemented` (which counts as a successful function call) in `functools.lru_cache` and `functools.cache` decorated binary dunder methods has the potential to fill the LRU cache with calls that will in ultimately result in errors (read "garbage").

There are a few ways to avoid this IMO:

1. Change `functools.lru_cache` not cache calls that result in `NotImplemented`.
2. Allow a user to easily extend `functools.lru_cache` (the implementation is not extensible right now) to do the previously mentioned operation.
3. Add the ability to *throw* `NotImplemented` in binary dunder methods so the function call does not complete. This would work exactly like returning `NotImplemented` and be handled by the runtime.

And my current solution...

4. Copy-paste `functools.lru_cache` and add in changes for solution 1 :)

> has the potential to fill the LRU cache with calls that
> will in ultimately result in errors

How would a call result in an error?  The worst that can happen is a cache miss and the underlying function gets called.


> Change `functools.lru_cache` not cache calls that 
> result in `NotImplemented`.

That would be a backwards incompatible change.


> Allow a user to easily extend `functools.lru_cache`

The implementation is intentionally closed off to support thread-safety, having a C implementation, ease of use, and making it performant.  

We've provided OrderedDict to make it relatively easy to roll your own cache variants.  Essentially the only algorithmically interesting part of the lru_cache is the linked list. OrderedDict has that logic already built in (the move_to_end() method does the equivalent of the linked list update).


> Add the ability to *throw* `NotImplemented` in binary dunder
> methods so the function call does not complete. This would 
> work exactly like returning `NotImplemented` and be handled 
> by the runtime.

Deep changes to the language would require a PEP.

> And my current solution...
> 4. Copy-paste `functools.lru_cache` 

The usual technique is to cache only what you want cached.  

Consider factoring-out the expensive logic rather than caching the NotImplemented logic:

    def __add__(self, other):
        if not isinstance(other, type(self)):
            return NotImplemented
        self._expensive_method(other)

    @lru_cache
    def _expensive_method(self, other):
        ...

Thanks for the suggestion, but I think it is better to use the tool as designed without any special cases.