Markus Schiltknecht <markus@bluegap.ch> writes:
> I've just found the stumbling block: the -c option of psql wraps all in
> a transaction, as man psql says:
> ...
> Thank you for clarification, I wouldn't have expected that (especially
> because CREATE DATABASE itself says, it cannot be run inside a
> transaction block... A transaction block (with BEGIN and COMMIT) seems
> to be more than just a transaction, right?)
Hm, that's an interesting point. psql's -c just shoves its whole
argument string at the backend in one PQexec(), instead of dividing
at semicolons as psql does with normal input. And so it winds up as
a single transaction because postgres.c doesn't force a transaction
commit until the end of the querystring. But that's not a "transaction
block" in the normal sense and so it doesn't trigger the
PreventTransactionChain defense in CREATE DATABASE and elsewhere.
I wonder whether we ought to change that? The point of
PreventTransactionChain is that we don't want the user rolling back
the statement post-completion, but it seems thatpsql -c 'CREATE DATABASE foo; ABORT; BEGIN; ...'
would bypass the check.
regards, tom lane