Christophe Pettus <xof@thebuild.com> writes:
> That's an interesting question. It appears to be a no-op, although a quick scan of the code doesn't reveal why.
Here'san illustrative test case:
This test case would be less confusing if the outer handler did
RAISE NOTICE 'in outer exception handler: %', sqlerrm;
With that, the test shows
regression=# call outer();
NOTICE: in outer exception handler: invalid transaction termination
CALL
What is happening is that inner() does PERFORM 1/0, fails and bounces
out to its exception handler, and then the ROLLBACK throws an error
because we're still inside outer()'s subtransaction. So inner()'s
first INSERT has been rolled back and the second one is never
reached. Back at outer()'s exception handler, we trap the error
from ROLLBACK, abort that subtransaction, and go on our merry way,
allowing the original INSERT (which was outside both subtransactions)
to complete.
regards, tom lane
