Confusing error message when setting item in a heterogeneous dict

[timabbott]

This simple piece of code throws a mypy error:

foo = dict(junk=1, bar=dict(yo=3))
foo['bar']['yo'] = 5

I've only tested on the last release (0.471)

[JukkaL]

The dictionary is heterogeneous, as it has both integer and dictionary values. In this case mypy infers foo to be a Dict[str, object]. Mypy gives the error because object does not support indexing.

The error message from mypy isn't very helpful though. For example, Dict[str, object] can only be indexed once would be a better message. Alternatively, mypy could require an annotation for foo, since it can't infer a precise type automatically. I'm leaving this issue open because the current behavior is confusing.

TypedDict is a work-in-progress mypy feature that will let you give a precise type for foo. It's still unclear if mypy will be able to automatically use a TypedDict type for foo -- it's possible that you'd still need an annotation.

[showell]

I'm not sure there's an easy way to convey this in the error message, but the mypy failure, despite the cryptic warning, led us to better code:

bar = dict(yo=3)
bar['yo'] = 5
foo = dict(junk=1, bar=bar)

Basically, we built the inner dict first before embedding it into the outer dict.

[benkuhn]

EDIT: I was confused and this is wrong. Original:

Isn't it a bug to infer the type as Dict[str, object] instead of Dict[str, Any] in this case? After all, the dicts contain values that are not compatible with type object. The following seems like it shouldn't issue a type error:

d = {'b': 1, 'c': 'foo'}
i = 1
i = d['b']

[gvanrossum]

@benkuhn Sounds like you have the definition of compatibility backwards. Being compatible is similar to being a subclass, with some added rules regarding Any thrown in. So, since every type is a subclass of object, every type is also compatible with object.

The difference between object and Any is not that they are at opposite ends of the type hierarchy. Instead, object is at the top, but Any is not in the type hierarchy at all. The rules that distinguish the two state that Any is compatible with everything and everything is compatible with Any. (See https://www.python.org/dev/peps/pep-0483/#summary-of-gradual-typing for a longer read on the subject.)

Regardless, inferring Dict[str, Any] in this case would be sub-optimal, because it would imply that as soon as you have a value of an unintended type in your dict, mypy would shut up instead of pointing out the mistake. (After all, who says you didn't mean d to contain only string values?)

[benkuhn]

Oh, whoops! You're right (and I'm convinced the code example probably should issue a type error).

[OrenLeaffer]

For posterity, here's the error that I get running the latest (38453c0):

$ cat z.py
foo = dict(junk=1, bar=dict(yo=3))
foo['bar']['yo'] = 5
$ python -m mypy z.py
z.py:2: error: Unsupported target for indexed assignment

I agree that this is a confusing error.

[sid-kap]

@OrenLeaffer are you planning to upstream your changes in OrenLeaffer@0eedd06?

[wkschwartz]

In the mean time, is there a canonical work-around for this problem?

[ilevkivskyi]

In the mean time, is there a canonical work-around for this problem?

Probably you might use TypedDict.

[ggirelli]

How about Dict[str, Any]?

foo: Dict[str, Any] = dict(junk=1, bar=dict(yo=3))
foo['bar']['yo'] = 5