Re: cached plan must not change result type

On Fri, 29 Mar 2024 at 12:21, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Dave Cramer <davecramer(at)postgres(dot)rocks> writes:
> > This is really an issue that needs to be solved in the backend. The error
> > is coming from PostgreSQL and what should happen is that when you alter a
> > table that a server prepared statement relies on the backend should send
> a
> > message to tell us that all of the prepared statements that rely on are
> now
> > invalid and we can reprepare them.
>
> This is something that can't change without a wire protocol change.
> There is nothing in the protocol that allows the backend to send out
> a message like "oh, that Describe I sent you awhile back? It might
> be a lie now" at random times.

I agree, but it's a known issue. I'm just pointing that it would be nice to
have.
We'd have to figure out the details.

> Also, what do you want to do about
> race conditions --- that is, what if you fire off an Execute only
> to find that one of those messages was already in flight to you?

> A non-racy way to handle it might be for Bind/Execute to refuse to
> run the query if its output has changed since the last Describe,
> which we could check after acquiring table locks during Bind.
> But we'd want to define "refuse" in a way that doesn't abort the
> transaction, and that's a concept that doesn't exist in the
> protocol at all.
>

This actually sounds like the best option as we wouldn't have to fire off a
message, just refuse to run the Execute in a way that doesn't abort the
transaction.

Dave