We can help to maintain a balanced graph by applying domain-specific knowledge to place nodes on shards; we can use insert-time algorithms to help us select the most appropriate shard to place a node; and we can use re-balancing algorithms periodically (sort of like a graph defrag) to help maintain a good node-to-shard mapping at runtime. We can also keep heuristic information on which relationships are traversed most frequently and optimise around those.
For graph databases, the problem is that what is optimal for a scenario can be a huge issue for other scenarios. If only you could rebalance on a scenario basis and that without killing the inter-node communication.
Original title and link: On Sharding Graph Databases (NoSQL databases © myNoSQL)