Using del on uncalled cached_property throws exception.

[Matthew Schinckel]

It's explicitly documented that you may use del (or delattr) to force a cached_property to recalculate the next time it's called.

However, if you try to del a cached_property that has never been evaluated, then an AttributeError is raised.

There doesn't seem to be a way to detect if a cached_property has been evaluated or not, as hasattr(foo, 'bar') results in the evaluation of the cached_property.

It's entirely possible that this is something that is beyond the capability of the way cached_property is implemented, but it still feels wrong to me.

(Likewise, if you attempt to del foo.bar twice without evaluation in between, you will raise an AttributeError).

[Curtis Maloney]

The problem here is (a) fundamental to how Python works, and (b) complicated by cached_property "cheating".

So, cached_property works by saving the value it calculates into the instances __dict__.

This means next time that attribute is accessed, _Python_ will check __dict__ first, and not even try the property.

Calling del will remove the entry from the instances __dict__. However, if it's not there, it will fall back to the class's property.

This causes the attribute error, because the property class does not have a __delete__ method.

If it did have a __delete__ method, Python's lookup would differ, and it would _not_ check __dict__ before calling the property. (the official documentation mentions this about __get__, but a simple test shows it applies to __delete__ equally)

​https://docs.python.org/3/reference/datamodel.html#invoking-descriptors

So, we can't have _both_ - if we support del on uninvoked cached_property we lose the utility of cached_property (or, at least, the performance)

[Matthew Schinckel]

Gotcha.

(In practice, it's possible to catch an exception raised by code that is calling del foo.bar when it's possible foo.bar has not been referenced yet, so I guess that will have to do).

[Carlton Gibson]

Great discussion. Thank you both!

Just before this disappears over the horizon, is it worth adjusting the del example in the docs here do we think?

[Matthew Schinckel]

Replying to Carlton Gibson:

Just before this disappears over the horizon, is it worth adjusting the del example in the docs here do we think?

I did wonder that: maybe it is worth having a note in the docs.

[Tim Graham <timograham@…>]

In c3c2ec54:

Refs #30278 -- Doc'd behavior of del on an unaccessed cached_property.

Thanks to Curtis Maloney for the description of the problem.

[Tim Graham <timograham@…>]

In f141704:

[2.2.x] Refs #30278 -- Doc'd behavior of del on an unaccessed cached_property.

Thanks to Curtis Maloney for the description of the problem.
Backport of c3c2ec54f59428cdf0a35abce594fd2ada35c209 from master

[Tim Graham <timograham@…>]

In b9455b01:

Refs #30278 -- Fixed link in cached_property docs.

[Tim Graham <timograham@…>]

In fc708f3:

[2.2.x] Refs #30278 -- Fixed link in cached_property docs.

Backport of b9455b010e41d1c6e68faa11115212d50de3c231 from master.