[rhinoceraptor]

I have to tell the database which columns I want in either case. The only difference (in my opinion) is now it's in a more convenient form, especially if you have a Node/Ruby/etc app where json/jsonb types is put into a native hash object by the PG library, and you would have to build that object anyways.

[kbenson]

Are you telling me there's not a way in Ruby or Node to query the database and put each result directly into the native hash object equivalent? That seems like the sort of thing that gets optimized fairly early in a database interface's existence. That JSON ends up being faster would surprise me, as there's still a deserialization step that needs to take place (JSONB might indeed help with this aspect if available though).

Now, if you are indeed just passing the JSON directly through to some client further up the stack, there probably is some benefit to generating JSON on the DB, as long as your DB isn't CPU constrained. For any other purpose, I imagine generating JSON would be less space efficient and require additional and/or more onerous steps to converting the record to the appropriate language construct.