I have a *VAST* database of 10's of millions records.
THere are two kinds of records in it, A an B.
An A-record may be referencing B records. B-records should be
referenced by one or more A-records,
But some B records may not be refererenced by any A record. That is
considered an error in the database that must be corrected.
I want to know how many of such (unused, surplus) B-records there are
in the database.

However: random access to the database is expensive and therefor not
allowed. Only sequentially reading the records is allowed.
Performance is important here.

How would I do it ?