Need faster split of ratings to sets

[ukrbublik]

1h is too much!
Try MySQL instead of Pg?
Try raw files (for sets splitting data only)?

[ukrbublik]

Get rid of malrec_ratings.dataset_type.
Create 2 tables malrec_users_datasets, malrec_items_datasets with 4 cols user_id/item_id, train, validate, test (last 3 are arrays of ids).