[Akuli]

Fixes #5567 so that Cursor and DictCursor no longer return annoying unions. I also got rid of several Optionals. In those methods, I looked at the source code and I tried select * from foo on a local test database, and couldn't get it to return None.

I also made Connection generic, so that when you create a connection, it remembers which cursor class you want to use. You can still customize this by passing a custom class to .cursor().

[srittau]

I haven't compared with the implementation, but the changes look reasonable. Especially getting rid of the response unions. Two remarks below.

Also, could you also update the version in METADATA.toml?

[alch-emi]

I'm having some problems with this, but I'm posting here instead of creating a new issue in case I'm doing something dumb. I'm currently writing some typed code that uses a Connection[Cursor], but when I include this in my function signature, I get the error TypeError: 'type' object is not subscriptable, presumably since the underlying library doesn't expect Connection to be a generic.

I was able to fix this either by patching pymysql to use a generic Connection, or by removing the generic from my type signature. The former solution of course makes the code much harder to maintain, and would require a fork of pymysql specifically for this project, at least until pymysql adopts python 3.10 or 3.7 (See PyMySQL/PyMySQL#953). The later causes my typechecker to complain very loudly about how Connection is supposed to be a generic, and also removes the benefit of having the generic in the first place.

Is there a solution to this that doesn't require either leaving a type error in the code or forking pymysql?

[Akuli]

On 3.7+, add to beginning of file:

from __future__ import annotations

This tells Python to not evaluate the type annotations, so that they are just for the type checker. It also means that you can also use Foo | None instead of Optional[Foo], for example.

[alch-emi]

Oh thank you so much! That's perfect!