consider two collections  collectionA and collectionB. When we do this aggregation

// code placeholder
collectionA.aggregate(…)

with a pipeline like

[
{
   $match: {foo: "bar"}
},
{
  $merge: {
    "into": "collectionB",
    "on": "_id",
    "whenMatched": "merge",
    "whenNotMatched" : "fail"
} 
]

Then the business code is returned a cursor that iterates over the whole collectionB.  

From my experiments, it looks as if the command executed by the java driver is equivalent to:

db.collectionA.aggregate(…) // wait until the command is over
db.collectionB.find({}) // return this cursor to the caller.

The cursor returned by the aggregation is supposed to yield no document, by design of the $merge stage. Somehow, the cursor returned yield the content of the whole collection that is the target of the $merge.

• It does not matter if the $match matches something or not.
• In the case there is a match, then the merge is observed in collectionB : the write does occur.
• this is not reproducible in the mongo shell or with the golang driver
• this is reproducible with a bunch of mongodb server including
  • mongodb-community/6.0.4
  • atlas cluster 8.0.4
  • docker image mongodb/mongodb-atlas-local:8.0.4