Re: Add new protocol message to change GUCs for usage with future protocol-only GUCs

On Wed, 17 Jan 2024 at 14:39, Robert Haas <robertmhaas@gmail.com> wrote:
> I think it's hard to say for sure what API is going to work well here,
> because we just don't have much experience with this.

Agreed, but I strongly believe PQunsupportedProtocolExtensions() is
useful regardless of the API choice.

> I also think that the reason why
> the API you're proposing here looks good in this case is because libpq
> itself doesn't really need to do anything differently for these
> parameters. It doesn't actually really change anything about the
> protocol; it only nails down the server behavior in a way that can't
> be changed. Another current request is to have a way to have certain
> data types always be sent in binary format, specified by OID. Do we
> want that to be written as PQsetParameter("always_binary_format",
> "123,456,789,101112") or do we want it to maybe look more like
> PQalwaysBinaryFormat(int count, Oid *stuff)? Or, another example, say
> we want to set client_to_server_compression_method=lz4.

I think from libpq's perspective there are two categories of protocol
extension parameters:
1. parameters that change behaviour in a way that does not matter to libpq
2. parameters that change in such a way that libpq needs to change its
behaviour too (by parsing or sending messages differently than it
normally would).

_pq_.protocol_roles, _pq_.report_parameters, and (afaict) even
_pq_.always_binary_format would fall into category 1. But
_pq_.client_to_server_compression_method would fall into category 2,
because libpq should start to compress the messages that it is
sending.

I think if you look at it like that, then using PQsetParameter for
parameters in category 1 makes sense. And indeed you'd likely want a
dedicated API for each parameter in category 2, and probably have
PQsetParameter error for these parameters. In any case it seems like
something that can be decided on a case by case basis. However, to
make this future proof, I think it might be best if PQsetParameter
would error for protocol extension parameters that it does not know
about.

> Also, I never intended for _pq_ to become a user-visible namespace
> that people would have to care about

I agree that forcing Postgres users to learn about this prefix is
probably unwanted. But requiring client authors to learn about the
prefix seems acceptable to me.