Adding some more granular options to influence the $sample behaviour to improve:

• Consistency in sample size (as a percentage vs a fixed size)
• Sampling method (rnd cursor vs collscan)

For example:

db.example.aggregate([{
  $sample: {
    percent: 10,
    method: "cursor"
  }
}]);

• The parameter for "size" and "percent" would be exclusive options
• The [optional] parameter for "method" would take a value of "cursor", "scan" or "auto", defaulting to auto (representing the current implementation