isinstance check against mutable mappings rather than dicts

[hayd]

atm we're checking dict-like by doing isinstance(x, dict), I think we should be using collections.MutableMapping, slightly more general.

[jreback]

yep

[jtratner]

it really is noticeably slower to use MutableMapping instead of dict (and unfortunately you incur the cost for every comparison). The perf difference isn't huge, but I've found it's enough to cause a problem in tight inner loops (e.g., wouldn't want to do this in type inference code from apply). mostly an FYI

[nicktimko]

Using the more general collections.Mapping would be preferable in the basic read cases (e.g. making a DataFrame). I have an lazy object that inherits from collections.abc.Mapping and I can't use it to initialize a DF, but if I pd.DataFrame(dict(mymapping)) it works just fine. Bit strange.

[jreback]

I think this is possible. You are welcome to submit a pull-request, with tests for this functionaility

[cpcloud]

Are there a lot of mappings that are not dicts? Having to call dict is kind of a minor inconvenience compared to a performance hit for which there is no user facing solution. On the flip side I can see why this is useful since maybe you don't want to be coupled to dicts and you just want to implement the interface. Need performance tests as well.

[immerrr]

I have always thought that isinstance(x, CLS) (or rather issubclass(type(x), CLS)) lends itself to caching pretty well, given that user types don't change that frequently.

I was in fact going to suggest to use singledispatch to improve structure of multi-branch ifs that implement overloading on parameter types and I'm pretty sure it caches the type-to-function-overload lookup somewhere under the hood, so if not borrowing actual code we could borrow some inspiration from it.

[mroeschke]

Closing in favor of #58803