Add NewType() to create simple unique types with zero runtime overhead

[gvanrossum]

The description is best gleaned from the following mypy issue: python/mypy#1284 (comment) and following. We're going with option (A).

Really brief example:

from typing import NewType
UserId = NewType('UserId', int)

Now to the type checker UserId is a new type that's compatible with int, but converting an int to a UserId requires a special cast form, UserId(x). At runtime UserId instances are just ints (not a subclass!) and UserId() is a dummy function that just returns its argument.

[ilevkivskyi]

I think this is a useful feature.

I have added PR #226 with a simple minded implementation of NewType, some tests, and a short discussion in the PEP.

Also I added NewType, Type and ContextManager to the list of things provided by typing.py.

I do not discuss whether NewType('EURtoUSD', Callable[[float], float]) etc. are allowed. If necessary, we could clarify this later.

[gvanrossum]

I would like to limit the arg to a class. Also note that that class must have a constructor that is compatible with that assumed by the type checker -- because whenever you write UserId(42) the runtime will actually see int(42) -- which is fine, and ditto if you use e.g. strings, since str('x') is fine, but if you have your own custom class that you want to diversify this way, you have to make sure that your constructor accepts instances of itself -- even though the customary way of constructing instances might be something totally different. This won't work:

class PacketId:
    def __init__(self, major: int, minor: int) -> None:
        self._major = major
        self._minor = minor
    # Etc.

TcpPacketId = NewType('TcpPacketId', PacketId)

a = TcpPacketId(127, 0)
x = PacketId(100, 100)
b = TcpPacketId(x)  # ???

UPDATE: As pointed out below this is backwards. In order not to confuse future readers I'm not changing my words but adding this disclaimer.

[ilevkivskyi]

I would like to limit the arg to a class.

OK

This won't work: ...

Actually, at runtime, UserId(42) will be just 42 not int(42). So that actually b = TcpPacketId(x)
will work, but a = TcpPacketId(127, 0) will not work. This is intentional (as discussed in mypy issue),
TcpPackedId works as a cast function, that should be fed by PacketId. Do I need to clarify this in the PEP?

[gvanrossum]

I guess you need to clarify this -- using int or str as examples doesn't make this clear at all, since they have the same signature. (And sorry for having such a frazzled memory that I didn't remember this -- but maybe it's been helpful teasing out the ambiguities in the description. :-)

FWIW I think we should take this to python-dev for a formal review period.

[ilevkivskyi]

I guess you need to clarify this -- using int or str as examples doesn't make this clear at all, since they have the same signature.

OK, I will do this and then will post a link to the PR to python-dev.

[gvanrossum]

To summarize, the essence here is that to the type checker, UserId is a subclass of int with a constructor that takes an int, while at runtime it is the identity function that returns its argument (hopefully an int :-) unchanged. (ADDED:) From this it follows that e.g. adding 1 to a UserId instance is legal and produces an int, not a UserId.

[ilevkivskyi]

From this it follows that e.g. adding 1 to a UserId instance is legal and produces an int, not a UserId.

I totally agree with this, there is already this example in the text

[gvanrossum]

I just wrote that down to summarize the proposal in the fewest words possible (and because I had never quite realized before that this is how the proposal works out -- it's a very sweet proposal that can be summarized so briefly).