After my experience in not using them for a large schema, using an int pk is pretty much a default decision for me, unless the table's PK is meaningful compared

After my experience in not using them for a large schema, using an int pk is pretty much a default decision for me, unless the table's PK is meaningful compared to another table like an association table. Doing big data migration jobs where the primary keys have to change is very painful. An identity-map ORM like SQLAlchemy is keying everything on primary key - SQLA supports mutable PKs and cascading updates (both via the DB, as well as in memory if you're on sqlite/MyISAM) fully, but the changing in-memory identities is fundamentally painful to deal with when you're shoveling around a lot of data and mutations.