From a three nodes cluster, our head devops expanded the cluster size to 6 nodes (2x 3) and we still recorded compacting fatal errors (on the leader too, so the whole cluster was unresponsive). Good news: five days ago the cluster was expanded to 10 nodes (2x 5) and finally we had some stability, no compacting errors in these 5 days, and CPU usage went down during off-peak hours. It made think that for stability reasons, the whole data center may run one day a huge dgraph cluster. Super horizontal scaling may be the answer! 
: