If you hard-kill a test process or the test suite hits the (10m default) timeout, then the postgres database used by the test is leaked into the postgres database server.

By default this server is reused between test runs (i.e. in engineer's workspaces, not CI), so it can cause the server to require ever-expanding resources. On our dogfood servers I have seen this result in excess paging which then kills performance and spirals out of control.