For now, we include all fields for sampling estimation used in join ordering & single table access path planning, but we actually could be smarter by extracting relevant fields to use as a projection. This is tricky in cases where we have multiple canonical queries per namespace, for instance.