On Tue, May 31, 2022 at 11:39:27AM +0900, Kyotaro Horiguchi wrote:
> The code looks like just a thinko that "COMMIT" works fine even if the
> given commands have ended in failure. But actually it doesn't for
> client-side failure.
>
> diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c
> index ddff903915..2261f78f81 100644
> --- a/src/bin/psql/startup.c
> +++ b/src/bin/psql/startup.c
> @@ -426,7 +426,9 @@ main(int argc, char *argv[])
>
> if (options.single_txn)
> {
> - if ((res = PSQLexec("COMMIT")) == NULL)
> + res = PSQLexec(successResult == EXIT_SUCCESS ?
> + "COMMIT" : "ROLLBACK");
> + if (res == NULL)
> {
> if (pset.on_error_stop)
> {
Yeah, it seems a bit strange to commit the changes if an error happens
on the client side, and the docs are a bit blurry about that because
it has never been considered, I guess. This would not happen with a
failure in the backend as COMMIT would just map to a ROLLBACK
automatically.
The change that you are sending would enforce this policy as Christoph
would like. Some tests would be nice to check such behaviors, say in
001_basic.pl, but we also need to be careful when sending down queries
with psql expected to fail because of SIGPIPE (c757a3d, 6d41dd0). The
docs need a refresh, they mention now that COMMIT is sent after the
last command but that would not be the case anymore with this patch if
there is a client-side error.
--
Michael