- There is no need for the initial DROP ROLE commands as those already get dropped at the end of the tests.
Removed.
- There is already rolenames.sql which has a tiny coverage for default roles, why not just using it?
Moved changes to rolenames.sql.
+-- should fail because regress_role_nopriv has not CONNECT permission on this db +SELECT pg_database_size('regression') > 0 AS canread; +ERROR: permission denied for database regression +-- should fail because regress_role_nopriv has not CREATE permission on this tablespace +SELECT pg_tablespace_size('pg_global') > 0 AS canread; +ERROR: permission denied for tablespace pg_global Why is that part of a test suite for default roles?
Just to check if changes broke something. I haven't find these checks in other
regress tests. In other way we get only positive tests. If this is not needed then
should I remove all the checks for regress_role_nopriv role or negative
regress_role_nopriv tests only?
2) is easy to be triggered as a negative test (which fails), less as a positive test. In order to make a positive test failure-proof with installcheck you would need to have a parameter which can be changed by a superuser at session level which gets updated to a certain value, and would fail to show for another user, so you could use one which is GUC_SUPERUSER_ONLY and of category PGC_SUSET, like session_preload_libraries or dynamic_preload_libraries. Still that's pretty restrictive, and would only test one out of the three code paths available.
