Failing to run a dataset's SQL can leave the database in an inconsistent state

[toolness]

Last night the auto-update of hpd_registrations failed. The actual cause of this error is documented in #61... a side-effect of the error, though, is that some of the functions created by hpd_registrations up to that point (specifically, anyarray_uniq) were removed from the public schema, and once this error was encountered, they were not re-created.

I think this happened because nycdb dataset SQL has code like the following:

DROP FUNCTION IF EXISTS blah();
CREATE OR REPLACE FUNCTION blah()

The behavior of code like this is not straightforward because of our connection's SCHEMA PATH. At the time that this code executes, it's set to a temporary randomly-named schema, but falls back to the public schema (see #4 for why this is done). I think the first line of the above code is actually dropping the function from the public schema (as it's not in the newly-created temporary one), while the second line is creating the function in the new temporary schema. Thus, the function in question now exists in the temporary schema, but doesn't exist in the public one. (It is actually made to exist in the public one further down the line, when we re-run the dataset's SQL in the context of the public schema, which was added in #22.)