A warning about freezegun on full unit test classes

[yajo]

I just wanted to warn the wider community about a little corner problem I found recently with freezegun.

There's definitely something weird in the way it patches setUpClass or other classmethods.

In short, the class-level cleanups are bound to a different in-memory object. This produces that a class decorated with @freezegun.freeze_time(some_date) losses class cleanups. A particularly important one is the one that closes the cursor. Thus, these classes leave a dangling cursor after exercising all tests within.

This can be or not a problem, depending on what happens after the tests are run. If Odoo just finishes, no problem; the cursor will get reaped by other means. But if you keep testing other modules whose installation means a DB lock in some way, you'll get to clunky, hard-to-diagnose Postgres locks.

This was notified upstream in spulec/freezegun#485 but:

• Don't be fooled by the title. It does affect other python versions too.
• There's zero activity to fix that.
• Odoo still requires very old freezegun versions in their requirements.txt files.

So, the most sensible approach for us here downstream at OCA, for now, is to replace such decorators like I did on OCA/sale-workflow#2747.

Take this example as a rule of thumb:

-@freeze_time("2021-10-02 15:30:00")
 class ExampleCase(TransactionCase):
     @classmethod
     def setUpClass(cls):
         super().setUpClass()
+        # HACK https://github.com/spulec/freezegun/issues/485
+        freezer = freeze_time("2021-10-02 15:30:00")
+        freezer.__enter__()
+        cls.addClassCleanup(freezer.__exit__)

Thanks!

[simahawk]

Interesting! What about decorating test methods? (Which btw is better from my POV in any case 😉 )

[yajo]

That works fine, but if you're creating frozen data in setUpClass, it won't work (unless you decorate that method too).