db.adminCommand({ setParameter: 1, internalQueryCBRCEMode: "samplingCE" , internalQueryPlanRanker: "costBased" , internalQueryStatsSampleRate: 1, }); for ( const coll of [ "a" , "b" ]) { db[coll].drop(); db[coll].insertMany(Array.from({length: 1000}, () => ({a: 0, b: 0}))); db[coll].createIndex({ a: 1 }); db[coll].createIndex({ b: 1 }); } db.adminCommand({ setParameter: 1, internalQueryStatsCacheSize: "0MB" }); // drop all the existing entries db.adminCommand({ setParameter: 1, internalQueryStatsCacheSize: "1%" }); // the default db.a.aggregate([{$match: { a: 0, b: 0 }}, {$lookup: {from: "b" , localField: "a" , foreignField: "a" , as: "c" , pipeline: [{$match: {b: 1}}]}}]).toArray(); printjson( db .getSiblingDB( "admin" ) .aggregate([ { $queryStats: {} }, //{ $match: { "key.queryShape.cmdNs.coll" : "a" } }, { $project: { key: 1, cbr: "$metrics.queryPlanner.costBasedRanker" } }, ]) .toArray(), );   Reports: cbr: { cardinalityEstimationMethods: { Histogram: Long('0'), Sampling: Long('3'), Heuristics: Long('0'), Mixed: Long('0'), Metadata: Long('0'), Code: Long('0'), HyperLogLog: Long('0') }, nDocsSampled: { sum: Long('380'), max: Long('380'), min: Long('380'), sumOfSquares: Decimal128('144400') } } So it reports only the size of one of the samples despite sampling three times (as evidenced by the counter in cardinalityEstimationMethods + with the help of a debugger). The first sample is to cost the access path on the first collection. The other two are for the second collection. This is due to the plan cache taking at least two queries to start using an entry.