Should HTTPMessage return `Optional[str]` for `__getitem__`?

[asottile]

>>> from urllib.request import urlopen
>>> urlopen('https://example.com')
<http.client.HTTPResponse object at 0x7fc608f39fd0>
>>> resp = urlopen('https://example.com')
>>> resp.headers['foo']
>>> resp.headers['connection']
'close'
>>> type(resp.headers['connection'])
<class 'str'>

import urllib.request
from typing import Optional

resp = urllib.request.urlopen('https://example.com')
x: Optional[str] = resp.headers['connection']

$ mypy t.py
t.py:5: error: Incompatible types in assignment (expression has type "Union[str, Header, None]", variable has type "Optional[str]")

If this sounds sane, I'd like to add __getitem__ here?

[JelleZijlstra]

Does it never return Header? I think we discussed this in some issue before.

[asottile]

I can't seem to find a situation where it returns Header, it's possible I'm overlooking something.

[OJFord]

In a model generated from bytes, any header values that (in contravention of the RFCs) contain non-ASCII bytes will, when retrieved through this interface, be represented as Header objects with a charset of unknown-8bit.

from the last paragraph of intro, just above https://docs.python.org/3.5/library/email.message.html#email.message.Message.__len__

[asottile]

Interesting! Though it doesn't appear to be true for HTTPResponse:

 while true; do echo -e "HTTP/1.1 200 OK\r\nContent-Length: 0\r\nWat: $(python -c 'print("\u2603")')\r\n" | nc -l localhost 1500; done

$ python3
Python 3.6.5 (default, Apr  1 2018, 05:46:30) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from urllib.request import urlopen
>>> x = urlopen('http://localhost:1500')
>>> x.headers
<http.client.HTTPMessage object at 0x7f51af37ae80>
>>> x.headers['Wat']
'â\x98\x83'
>>> type(x.headers['Wat'])
<class 'str'>

Some pretty sick mojibake though :)

[srittau]

I think the right solution is to override __getitem__() et al. in http.client.HTTPMessage to not return Header.

[asottile]

this seems worse now, it's Any so all type checking is gone -- this appears to be due to #3045

[Avasam]

Taking a look at this test run: https://github.com/Avasam/typeshed/pull/36/files#diff-ced02b01b42b244e9b7bc2e12ef7a215bde89a29072c99ade903f58fb83a3bc7R36

spack would have to change their assert to first store the value in a local variable then assert that variable: https://github.com/spack/spack/blob/develop/lib/spack/spack/oci/oci.py#L138-L142

twine implicitely uses header_factory. https://github.com/pypa/twine/blob/main/twine/commands/check.py#L73C40-L73C54 They could replace with

    _, header = msg.policy.header_store_parse("content-type", value)`
    return msg.get_content_type(), header.params

urllib3 is a false-positive, and an interesting one: https://github.com/urllib3/urllib3/blob/main/src/urllib3/contrib/emscripten/fetch.py#L382C24-L382C24

---

I'm tempted to say that, from a user code PoV, you should just use .get instead of accessing through [] to be type-safe. And that usual expectation with ["key"] is that you're assuming that the value is present otherwise it throws (except that it returns None here instead of throwing).

---

As for the type currently being Any. I don't think it could really be resolved without something like AnyOf, or making Message generic with a default value (wasn't there a PEP for that?)

If HTTPMessage really is always strings, we could override all Message methods that use _HeaderType