Using the following code

ctx := context.Background()

c := h.db.Collection("col")
cs, err := c.Watch(ctx, []bson.M{})
if err != nil {
	log.Println("cannot watch for changes", err)
}
defer cs.Close(ctx)

for cs.Next(ctx) {
	doc := bson.M{}
	if err := cs.Decode(&doc); err != nil {
		continue
	}

	// do something with doc
}

if cs.Err() != nil {
	log.Println("error reading changestream", cs.Err())
}

 
When the MongoDB server goes down or gets restarted, the for loop keeps hanging and doesn't receive any more data or error.
Working with a time-constrained context seems wrong, since I actually want to listen endlessly.
IMHO either the driver should silently handle the reconnect and setup the change stream again, or it should properly signal the lost session/connection by returning from the Next() call with an error, so I have a chance to reestablish the change stream myself.