If you have an elemMatch with a conjunction:

db.c.find({a: {$elemMatch: {x: 1, y: 1}}})

Then we could generate a plan that scans a multikey index on {a: 1}, and apply the {x: 1, y: 1} predicate to each index key. (And then do the "is array" check after fetching.)

Bonsai does not consider this plan, because we relax the predicate to two separate Traverses:

Filter ... Get [a] Traverse ComposeM (Get [x] ...) (Get [y] ...)
->
Sargable
  Get [a] Traverse Get [x] ...  perfOnly
  Get [a] Traverse Get [y] ...  perfOnly

This means we consider plans such as:

• Full index scan, filter one predicate, fetch, filter on whole predicate
• Index intersection of {a: 1} with itself
• Collection scan

(This is still an improvement over classic, which only considers a collection scan. Even with hint({a: 1}), classic does not consider evaluating either predicate on the index side to reduce fetching.)

Maybe this would break down into sub-features:

• an exploration rewrite that turns Traverse predicates into Unwind
• a rewrite that matches Unwind with a multikey index