Turns out it isn't safe to use aggregation in 2.4 at all.

2.4 fails if the total size of the aggregation is too big,
while in 3.2 it only fails if the size of an individual doc
gets too big.
Running mgopurge against big 2.4 databases was failing
(2.4M txns * 20 bytes per ID is >16MB).

oracle.go

@@ -99,14 +99,13 @@ type DBOracle struct {
 // populates the working set by inserting them from the client. This is less
 // efficient that a $out in the pipeline, but must be used when Mongo doesn't
 // support pipelines.
-func (o *DBOracle) prepareWorkingDirectly(pipeline []bson.M) error {
+func (o *DBOracle) prepareWorkingDirectly() error {
 	logger.Debugf("iterating the transactions collection to build the working set: %q", o.working.Name)
 	// Make sure the working set is clean
 	o.working.DropCollection()
-	pipe := o.txns.Pipe(pipeline)
-	pipe.Batch(maxBatchDocs)
-	pipe.AllowDiskUse()
-	iter := pipe.Iter()
+	query := o.txns.Find(bson.M{"s": bson.M{"$gte": taborted}}).Select(bson.M{"_id": 1})
+	query.Batch(maxBatchDocs)
+	iter := query.Iter()
 	var txnDoc struct {
 		Id bson.ObjectId `bson:"_id"`
 	}
@@ -149,8 +148,13 @@ func (o *DBOracle) prepareWorkingDirectly(pipeline []bson.M) error {
 
 // prepareWorkingWithPipeline adds a $out stage to the pipeline, and has mongo
 // populate the working set. This is the preferred method if Mongo supports $out.
-func (o *DBOracle) prepareWorkingWithPipeline(pipeline []bson.M) error {
-	pipeline = append(pipeline, bson.M{"$out": o.working.Name})
+func (o *DBOracle) prepareWorkingWithPipeline() error {
+	pipeline := []bson.M{
+		// This used to use $in but that's much slower than $gte.
+		{"$match": bson.M{"s": bson.M{"$gte": taborted}}},
+		{"$project": bson.M{"_id": 1}},
+		{"$out": o.working.Name},
+	}
 	pipe := o.txns.Pipe(pipeline)
 	pipe.Batch(maxBatchDocs)
 	pipe.AllowDiskUse()
@@ -167,16 +171,11 @@ func (o *DBOracle) prepare() (func(), error) {
 	// Load the ids of all completed and aborted txns into a separate
 	// temporary collection.
 	logger.Debugf("loading all completed transactions")
-	pipeline := []bson.M{
-		// This used to use $in but that's much slower than $gte.
-		{"$match": bson.M{"s": bson.M{"$gte": taborted}}},
-		{"$project": bson.M{"_id": 1}},
-	}
 	var err error
 	if o.usingMongoOut {
-		err = o.prepareWorkingWithPipeline(pipeline)
+		err = o.prepareWorkingWithPipeline()
 	} else {
-		err = o.prepareWorkingDirectly(pipeline)
+		err = o.prepareWorkingDirectly()
 	}
 	if err != nil {
 		o.cleanup()