Cost estimation may add unnecessary overhead if it uses sampling because it may read (nearly) the whole collection, then produce a plan, which evaluated after that only to read more or less the same amount of data.

At the same time multi-planning could evaluate the query directly.

This ticket is based on some performance assumptions about the overhead of various components, and in the end may turn out to be unnecessary. Consider first implementing a prototype and perform some measurements before getting deep into the implementation.

The ticket must also implement a benchmark to determine what collection size is small enough, and to validate the approach. The final implementation should implement a benchmark to guard from regressions.