Probably not still an issue for op, but was having this issue on a self-hosted Docker swarm service with three replicas. The resolution was to limit to 1 replica.