We should consider the number of documents scanned between different plans being evaluated. At the very least, we should use that to resolve ties between plans.

This is critical for Big Data systems reading data via Spark Connector that partitions data by _id for reading. Example:

For the query

{date:{$gte:A}, _id:{$gte:B}, email:{$gte:C}}

all indexes below will tie, although it is beyond obvious which one should be selected (the difference is of course dramatic when we're talking about TB's of data):

{date:1, _id:1}
{date:1, _id:1, email:1}
{date:1, _id:1, some_other_email:1}