[MNT] added joblib as core dependency

[yarnabrina]

There are multiple instances of direct imports by joblib without isolation, and it never causes problem as it is a dependency of scikit-learn. It has always been so and unlikely to change in future, so moving it to core dependencies to avoid conda linting warnings.

[fkiraly]

I was working on isolating it, give our conversation - we can discuss which way to go, of course.
This one is always going ot work, in case we can't figure out isolation/conditional.

[fkiraly]

Here: #6385

[yarnabrina]

@fkiraly all jobs got triggered in this PR as pyproject.toml got changed. None of them failed (logical or timeout). Does it mean:

1. at least for once the culprit estimator didn't cause any issues (sporadic behaviour)
2. the culprit estimator got fixed by some package updates of underlying soft dependency
3. the culprit estimator did not get triggered at all

The change in this PR is in core dependencies section, how exactly is it expected to affect the incremental testing?

[fkiraly]

@fkiraly all jobs got triggered in this PR as pyproject.toml got changed. None of them failed (logical or timeout). Does it mean:

1. at least for once the culprit estimator didn't cause any issues (sporadic behaviour)

2. the culprit estimator got fixed by some package updates of underlying soft dependency

3. the culprit estimator did not get triggered at all

The change in this PR is in core dependencies section, how exactly is it expected to affect the incremental testing?

My expectation is, none of the estimator suite tests got run, because joblib is not an explicit dependency of estimators, and core dependency changes do not trigger estimator runs at the current state of the test framework. So, number 3.

Question: would it be good to have a step in the CI that lists the set of estimators tested?

[yarnabrina]

Question: would it be good to have a step in the CI that lists the set of estimators tested?

Yes. Possible issue with that is the subset condition that is there where not all tests for all estimators run in all jobs, but in a pseduo random subset of OS-python combinations.

[fkiraly]

Yes. Possible issue with that is the subset condition that is there where not all tests for all estimators run in all jobs, but in a pseduo random subset of OS-python combinations.

I mean, the pseudo-random subset that is actually tested

[fkiraly]

Should we schedule this for 0.30.0 then?