conftest: put transaction manager in its own fixture

[woodruffw]

This moves the tm local from the webtest fixture into its own fixture.

This should have no impact on its own, but it will hopefully make it easier to tease apart whatever session/TM mismatch is causing CI failures like https://github.com/pypi/warehouse/actions/runs/11057423074/job/30721387457?pr=16778.

See: #15634 (comment)

[di]

@woodruffw Looks like this breaks some tests (probably because it's session-scoped).

[woodruffw]

@woodruffw Looks like this breaks some tests (probably because it's session-scoped).

Whoops, fixed!

[miketheman]

I'm 👍 generally, but I'm curious about the timing of this - since the tm.begin()/tm.doom() calls now happen after get_db_session_for_app_config() is called - does that change the behavior at all?

[woodruffw]

I'm 👍 generally, but I'm curious about the timing of this - since the tm.begin()/tm.doom() calls now happen after get_db_session_for_app_config() is called - does that change the behavior at all?

I might not be following, but I think the order is correct here -- if my understanding of pytest fixture scoping is correct, the ordering here should be:

database (session-scoped)
  app_config_dbsession_from_end (session-scoped)
    tm (function-scoped)
      get_db_session_for_app_config -> _db_session (context manager within function scope)
        actual test code

But it's possible I'm missing where the scope has changed 🙂

[miketheman]

Maybe I didn't phrase the question well?
It's specifically that currently, tm.begin()/tm.doom() are explicitly called prior to calling get_db_session_for_app_config and creating the TestApp, and with this change, the call to begin/doom happens later.

I don't think it will matter, so let's ship this and find out.

[woodruffw]

It's specifically that currently, tm.begin()/tm.doom() are explicitly called prior to calling get_db_session_for_app_config and creating the TestApp, and with this change, the call to begin/doom happens later.

Ah, I see what you mesn! Yeah, I think that won't matter, but that is indeed a slight difference.