If the flushing server can't be reached for a bit, a large batch of stats may build up. We've had an instance recently where the data size after encoding (6MB) was too big for the endpoint it was being sent to.
Suggestion
Chunk the and submit it in batches