I wrote:
> ISTM that an appropriate fix would be to remember if any command
> failed, not just the last one.
Alternatively, one could argue that once we've decided to issue
ROLLBACK, there's not really any reason to continue performing
additional -c/-f actions, so that a sufficient fix would be to
get rid of the loop's ON_ERROR_STOP conditionality:
- if (successResult != EXIT_SUCCESS && pset.on_error_stop)
+ if (successResult != EXIT_SUCCESS)
break;
But I'm not sure that that argument is bulletproof. If we are
considering a client-side failure then the server may still think
the transaction is fine, in which case we should continue to
perform actions that could have client-side side effects; or
at least, not doing so is a potentially significant behavioral
change.
In any case, I now agree with Robert's upthread objection that
this should not have been back-patched. It's a nontrivial
behavioral change and it's not clear that it's 100% without
downsides. I particularly do not want to ship 14.4 with this,
because we really need a clean release with no new regressions.
regards, tom lane