If, information in the filter is not contained in any of the indexes, then I don't see how it could avoid reading all the records in the scope.
Filter information is never stored in indexes.
RDD builds a record list (whatever name it is called) of which each bit represents one record. RDD sets on and off the bits evaluating the filter condition based on index expressions and index values. Where the filter condition is not fully optimized, RDD scans the remaining records (maybe's) and eliminates filtered records. This is normally done as and when required. OrdKeyCount() also forces this operation.
Finally counting filtered records is only counting "on" bits of the internal list.
DBFCDX RDD (and some other advanced RDDs) basically has good optimization capabilities. It is for the programmer to make the best use of these capabilities.
1. Do not use aliases in index experssions or filter expressions.
INDEX ON STATE is better than INDEX ON CUST->STATE or INDEX ON FIELD->STATE
2. Do not use memory variables or UDFs in the filter experssions. Instead use literal constants.
3. Left side of a comparison expression should match exactly an index expression.
When there is an index on UPPER(CITY), better we use UPPER(CITY) < "M" than CITY < "M"
4. If the filter expression contains optimzable and not optimizable expressions, let us start with optimizable expressions first. If some expressions result in fewer filtered records then let us have such experssions first. (I am not very sure how our RDD works, but this is by long years of experience in building where/filter clauses on large RDBMSs)
5) If targetting networked environments, keep the size of index files to the minimum. This is very important
