feat: bulk save objects ignoring conflicts

[victorgarcia98]

Previously, it wasn't possible to bulk save beliefs with the allow_overwrite = True. With the help of ON CONFLICT DO NOTHING, we can bulk save beliefs and ignore duplicated key conflicts.

[Flix6x]

Looks good! I only have two minor requests, and it seems the pipeline still fails on pre-commit.

Was this option introduced recently in SQLAlchemy? If so, perhaps we should specify a minimum requirement, too.

[nhoening]

One naming comment, one idea for complete support of allow_overwrite

[victorgarcia98]

Was this option introduced recently in SQLAlchemy? If so, perhaps we should specify a minimum requirement, too.

Good point. This feature was introduced in SQLAlchemy 1.1 and the project requires SQLAlchemy >= 2.0 so we don't need to constraint it.

[Flix6x]

Great catch by @nhoening. By now I would prefer to use on_conflict_do_nothing over raising an exception, in case allow_overwrite is False.

And also I think we can now default to the faster bulk saving in the function signature. I think we used the slower option as a default because it was less prone to raise errors.

[victorgarcia98]

Good idea! I've implemented the suggested changes and update the tests.

[Flix6x]

Thanks for addressing the failing tests.

Note that I had first resolved one of the failing tests differently. You revised add_to_session to make sure it does not fail if an empty frame is being saved, which was the right thing to do, while I first thought the test wasn't supposed to save an empty frame in the first place, so I had started revising the test and conftest. I reverted my changes, although there is still something odd going on there, namely this:

Due to the use of params=[1, 2] in the rolling_day_ahead_beliefs_about_time_slot_events fixture, each test that uses that fixture is being called twice as many times (in the case of test_adding_to_session, two sets of 4 times, due to the use of @pytest.mark.parametrize), where in the first round of calls, the fixture set up data from source A and none from B, and the second round the fixture set up data from source B and none from A. The test then selects only data from source B. All in all, the test setup and test cases could be clearer. But I'm leaving this for another day.