We just disabled warnings on travis in #9840.
I don't think we should do that. I've been a bit absent lately, but I think the current state of the warnings is pretty bad. Many of these seem recent changes that require cleanup.

• catch and ignore all ConvergenceWarnings raised in sklearn.utils.estimator_checks done in [MRG+2] catch more expected warnings in common tests #11151.
• catch deprecations of min_impurity_split in tree and tree ensembles, make sure it's not used anywhere else.
• catch deprecations of random_state in SVC tests, make sure it's not used anywhere else.