In general, I agree, though pg's JSONB is effectiuvely hstore2 so its performance is surprisingly good.

This example didn't involve a JSON column in the database at all though, I was talking about turning native database rows into json data - so given tables like users(id, name) and posts(post_id, author_id, title, body) (apologies for pseudocodish rendering) instead of

    <author_id>, <author_name>, <post_id_1>, <post_title_1>, <post_body_1>,
    <author_id>, <author_name>, <post_id_2>, <post_title_2>, <post_body_2>,

you'd instead get

    <author_id>, <author_name>, json[
      { post_id: <post_id_1>, title: <post_title_1>, body: <post_body_1> },
      { post_id: <post_id_2>, title: <post_title_2>, body: <post_body_2> }
    ]

so on the application side you'd just deal with (for each row) something like

    let id = author.id;
    let name = author.name;
    let posts = author.posts;

and the 'posts' variable would be an array of objects with 'post_id', 'title' and 'body' keys.