Partitioning and coordination

Partitioning spreads data and work across nodes, but it also turns local operations into distributed operations when a request spans partitions. Choose a key that distributes load while keeping common queries local. Watch for hot keys, uneven growth, and rebalancing cost.

Coordination boundaries

Leader election, membership, and consensus are useful when many nodes need one authoritative decision. They do not remove failure, and they do not make arbitrary multi-step business workflows atomic. Keep coordination scopes small and make retries idempotent.

Rebalancing

Adding capacity should not require moving the entire dataset at once. Use staged migration, dual reads or writes when needed, verification, and a rollback plan. Measure the effect on latency and background bandwidth while the move is running.