Github user orhankislal commented on the issue:
https://github.com/apache/madlib/pull/325
@iyerr3 @jingyimei I'll put the following in the commit message before
merging.
This commit fixes the following potential issue.
1. User deploys MADlib on the schema `madlib1`.
2. User creates a table named `vertex` in the `madlib1` schema.
3. User runs install-check.
4. The install check creates a new role and a new schema for each module in
the database.
5. The install check sets the `search_path` to
`madlib_installcheck_<module>, madlib1`.
6. The graph IC calls `DROP TABLE IF EXISTS vertex` and fails because the
vertex table does exist but it is not owned by the install-check role.
This commit removes the madlib installation schema from the search path so
that it only uses its own schema. This means every madlib function call, type
and operator has to be called directly using the madlib schema name.
One alternative solution is eliminating the `drop table` commands from the
tests but that would require a very complicated refactoring work since most of
the tests are written to reuse the same output table names. Another alternative
is changing the `drop table` and `create table` commands to use the newly
created test schema. However, this is very tricky to test; if a developer
forgets to put the schema name, the test will still work unless she also
creates a table of the same name in the madlib deployment schema.
---
