I tried this locally and added and it seems to be working correctly. Haven't benchmarked it either though.

The cached methods seem to stick around in the cache until the gc fires. I haven't been able to find a way to make them stick around.

Does CPython re-useid values after instances are destroyed?

SO says that yes, but it was in 2018. Is it still true?

If so, then can identity be reliably cached at all?

Maybe similar approach tocached_property could be used here? Although it would restrict to classes that haveinstance.__dict__.

Does CPython re-useid values after instances are destroyed?

SO says that yes, but it was in 2018. Is it still true?

If so, then can identity be reliably cached at all?

Maybe similar approach tocached_property could be used here? Although it would restrict to classes that haveinstance.__dict__.

Yes, cpython re-uses the id (the id is equal to the memory address, and memory is reused). But the id is unique during the lifetime of an object, seehttps://docs.python.org/3/library/functions.html#id.

The_method in theWeakValueDictionary has a reference toobj, so the keyid(obj)is valid and unique as long as_method exists. Once the garbage collector detects a cycle containing the_method it will remove it from the dictionary. But sinceobj is part of this cycle, it also meansobj no longer exist andid(obj) is free to re-use.

But your remark got me thinking: the PR right now is reliable, but maybe not very efficient. Runninggc.collect clears the entire cache. The benchmarks I executed showed an improvement, but that might very well be because the garbage collector is not running often.

Runninggc.collect clears the entire cache.

Could you clarify what do you mean by this?

Runninggc.collect clears the entire cache.

Could you clarify what do you mean by this?
Take for example:

@dataclassclass A:    value: int    @singledispatchmethod    def dispatch(self, x):        return id(self)t1 = A(1)t1.dispatch(10)

Right after running this code the cache (e.g._method_cache) will contain a single key-value pair: the key isid(t1) and the value is the_method object created in the dispatcher. The_method has a positive reference count (since_method.__self_reference holds a reference to_method).
When the garbage collector runs (seePython internal documentation on GC for details) the_method is part of a cycle of objects (e.g._method ->_method.__self_reference ->_method) and is removed. Since there the_method object is no longer alive, theWeakValueDictionary cache_method_cache will remove the key-value pair.

We can address the issue by creating a reference fromobj to the entry in the cache. Something like:

        try:            setattr(obj, f'_singledispatchmethod_cache_{id(self)}', _method)  # create strong reference to _method which is not an isolated cycle. see gh-127751        except Exception as ex:            # frozen object, disable the cache            self._method_cache = None

This works, but has the disadvantage that we are adding some (private) attributes in the object instance.

Note: the documentationhttps://docs.python.org/3/library/weakref.html#weakref.WeakValueDictionary states:

Mapping class that references values weakly. Entries in the dictionary will be discarded when no strong reference to the value exists any more.

That text is a bit misleading: there is a reference to_method, but still it is removed. I will make a separate PR to address this.

I see now.
I think simplifying this might be worthwhile.

Also, I would like the method to be cached for the lifetime of my instance, doesn't it currently garbage collect perfectly good cached methods?

This works, but has the disadvantage that we are adding some (private) attributes in the object instance.

What about the same way ascached_property works?

classsingledispatchmethod:def__set_name__(self,owner,name):self.attrname=namedef__get__(self,instance,owner=None):try:cache=instance.__dict__exceptAttributeError:# disable caching going forwardelse:method=cache.get(self.attrname)ifmethodisnotNone:returnmethodelse:                ...

I see now. I think simplifying this might be worthwhile.

Also, I would like the method to be cached for the lifetime of my instance, doesn't it currently garbage collect perfectly good cached methods?

Yes, that is why I did not like the approach in this PR.

What about the same way ascached_property works?

Nice idea. I created an alternative PR for this variation.

Issue resolved in other PRs, closing