čt 10. 8. 2023 v 14:05 odesílatel Jelte Fennema <postg...@jeltef.nl> napsal:
> On Tue, 8 Aug 2023 at 07:20, Pavel Stehule <pavel.steh...@gmail.com> > wrote: > > The reason why I implemented separate flow is usage from psql and > independence of transaction state. It is used for the \set command, that > is non-transactional, not SQL. If I inject this message to some other flow, > I lose this independence. > > I still don't understand the issue that you're trying to solve by > introducing a new flow for handling the response. What do you mean > with independence of the transaction state? That it is not rolled-back > in a case like this? > > BEGIN; > \set PROMPT '%N' > ROLLBACK; > surely not. \set is client side setting, and it is not transactional. Attention - "\set" and "set" commands are absolutely different creatures. > That seems more like a Postgres server concern, i.e. don't revert the > change back on ROLLBACK. (I think your current server-side > implementation already does that) > Postgres does it, but not on the client side. What I know, almost all environments don't support transactions on the client side. Postgres is not special in this direction. > > I guess one reason that I don't understand what you mean is that libpq > doesn't really care about "transaction state" at all. It's really a > wrapper around a socket with easy functions to send messages in the > postgres protocol over it. So it cares about the "connection state", > but not really about a "transaction state". (it does track the current > connection state, but it doesn't actually use the value except when > reporting the value when PQtransactionStatus is called by the user of > libpq) > > > Without independence on transaction state and SQL, I can just implement > some SQL function that sets reporting for any GUC, and it is more simple > than extending protocol. > > I don't understand why this is not possible. As far as I can tell this > should work fine for the usecase of psql. I still prefer the protocol > message approach though, because that makes it possible for connection > poolers to intercept the message and handle it accordingly. And I see > some use cases for this reporting feature for PgBouncer as well. > Maybe we are talking about different features. Now, I have no idea how the proposed feature can be useful for pgbouncer? Sure If I implement a new flow, then pgbouncer cannot forward. But it is not too hard to implement redirection of new flow to pgbouncer. > However, I think this is probably the key thing that I don't > understand about the problem you're describing: So, could you explain > in some more detail why implementing a SQL function would not work for > psql? > I try to get some consistency. psql setting and some features like formatting doesn't depend on transactional state. It depends just on connection. This is the reason why I don't want to inject dependency on transaction state. Without this dependency, I don't need to check transaction state, and I can execute prompt settings immediately. If I implement this feature as transactional, then I need to wait to idle or to make a new transaction (and this I have not under my control). I try to be consistent with current psql behaviour. It Would be strange (can be very messy) if I had a message like "cannot set a prompt, because you should do ROLLBACK first" When this feature should be implemented as an injected message, then I have another problem. Which SQL command I have to send to the server, when the user wants to set a prompt? And then I don't need to implement a new message, and I can just implement the SQL function pg_catalog.report_config(...).
