Ayush Tiwari <ayushtiwari.slg01@gmail.com> writes:
> COPY TO FORMAT JSON silently accepts the ENCODING option but doesn't
> perform encoding conversion(?) CopyToJsonOneRow() sends the output of
> composite_to_json() via CopySendData() without calling
> pg_server_to_any(), unlike the text and CSV paths.
> COPY t TO '/tmp/out.json' WITH (FORMAT json, ENCODING 'LATIN1');
> On a UTF-8 server this produces UTF-8 output, not LATIN1.
Seems to me the correct thing here is to make it work like the other
cases, ie perform pg_server_to_any(). I have exactly no sympathy for
the argument about the RFC saying it must be UTF-8, not least because
that's not in fact what is implemented (what if the server encoding
isn't UTF-8?).
Rejecting this option altogether doesn't improve anything, not
functionally, not specs-compliance-wise, nor according to the
principle of least surprise.
> The attached patch rejects the explicit ENCODING option for JSON
> mode, consistent with how DELIMITER, NULL, DEFAULT, and HEADER are
> already rejected. The implicit client_encoding case is a separate
> design question (should COPY TO JSON always emit UTF-8 regardless
> of client_encoding?) that maybe we should address separately and not as
> part of v19.
No, you don't get to punt this till later. Once we ship v19 there's
going to be a strong expectation of backwards compatibility.
The idea of sending UTF-8 to a client that's set client_encoding to
something else would be risible, if it weren't a security hazard.
regards, tom lane
