The vectorized engine currently only handles DISTINCT on fully-ordered columns. It doesn't yet know how to deal with distincts on unordered (or partially ordered) columns.

This issue tracks implemented an unordered distinct operator. Unordered distinct is implemented with a hash table. The DistSQL implementation of unordered distinct could serve as a guide - but it's possible that we could get away with reusing much of the unordered aggregation code rather than doing something from scratch.