The previous implementation skipped empty fields, which meant targets
differing only in which fields were populated could produce the same
string (e.g. {UserId:"A"} and {DomainId:"A"} both gave "A"). This
caused key collisions when the string was used in storage index keys.