Re: savepoint improvements

On Sun, 2007-01-21 at 13:28 -0500, Tom Lane wrote:
> "Jaime Casanova" <systemguards(at)gmail(dot)com> writes:
> > On 1/21/07, Simon Riggs <simon(at)2ndquadrant(dot)com> wrote:
> >> - continue on error i.e. COMMIT can/might succeed - though there are
> >> still cases where it cannot, such as a serializable exception.
>
> > and what should be the behaviour of that? the same as rollback?

No. The behaviour is to continue the transaction even though an error
has occurred, i.e.

BEGIN;

1. INSERT...
success

2. INSERT .... VALUES () () ()
--fails with error on 3rd VALUES statement

dynamically re-construct INSERT statement with remaining 2 VALUES
statements

3. INSERT VALUES () ();
success

COMMIT;
work done by 1 and 3 is committed

Behaviour needs to support any error at (2) except serializable
exceptions.

> The only conceivable implementation is an implicit savepoint issued
> before each statement.

Perhaps the only acceptable one.

> By and large that seems to me to be most easily
> handled on the client side, and many of our client libraries already
> have the ability to do it.

PL/pgSQL supports EXCEPTIONs, but no other clients support it, AFAICS.

> (For instance, psql has ON_ERROR_ROLLBACK.)

Thats not the same thing, regrettably.

> If we tried to do it on the server side, we would break any client
> software that wasn't prepared for the change of behavior --- see the 7.3
> autocommit fiasco for an example.

Only if we changed the default behaviour, which I am not suggesting.

> So as far as the server is concerned, I see no TODO here.

If the server team won't allow it, we must document that this behaviour
must be a client-side function in the *server* TODO, so that all the
various client projects can read the same TODO item and implement it.

"Implement continue-on-error transactional behaviour for each client
library".

--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com