Their answer clearly, explicitly, assumes a C-style `union`. Despite the essay literally using a proper sum type as example.

> The argument for union types seems to get weak when one asks: how do we index their components?

With the system creating partial indices under the cover? I fail to see what's complicated about it.

> SQL need not be "one language to rule them all".

SQL has "won" the relational battle and is literally the only way to query relational databases. Any time SQL is unable to do the job and you have to move that job to application code, you're making the schema less reliable and less of a source of truth, because parts of the schema's information have to be embedded in each application instead.

That doesn't seem desirable to me, unless you assert SQL should just be a trivial data storage and retrieval interface, which it has not been… possibly ever, but at the very least since the introduction of window functions.