Hello,
The attached patch changes the location of the dumpUserConfig call in
the dumpRoles function of pg_dumpall.
This is related to this thread:
http://archives.postgresql.org/pgsql-hackers/2011-02/msg02359.php
Currently if you use 'ALTER ROLE rolename SET ROLE', pg_dumpall will
dump an 'ALTER ROLE' out right after the 'CREATE ROLE' statement.
Sometimes this will cause a conflict when a dependent role is not yet
created:
--
-- Roles
--
CREATE ROLE a;
ALTER ROLE a WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOLOGIN
NOREPLICATION;
ALTER ROLE a SET role TO 'b';
CREATE ROLE b;
ALTER ROLE b WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOLOGIN
NOREPLICATION;
CREATE ROLE postgres;
ALTER ROLE postgres WITH SUPERUSER INHERIT CREATEROLE CREATEDB LOGIN
NOREPLICATION;
As you can see, role a is set to role b before role b is created.
This patch moves the call to dumpUserConfig to after the loop where
all the roles are created. This produces output like the this:
--
-- Roles
--
CREATE ROLE a;
ALTER ROLE a WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOLOGIN
NOREPLICATION;
CREATE ROLE b;
ALTER ROLE b WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOLOGIN
NOREPLICATION;
CREATE ROLE postgres;
ALTER ROLE postgres WITH SUPERUSER INHERIT CREATEROLE CREATEDB LOGIN
NOREPLICATION;
ALTER ROLE a SET role TO 'b';
Now this dump will succeed upon restore.
This passed all regression tests.
Thanks.