Rebalancing Partitions

Rebalancing moves load from overloaded nodes to underloaded nodes.

Strategies

1. Fixed Number of Partitions: Create many more partitions than nodes (e.g., 1000 partitions for 10 nodes). If you add a node, it steals a few partitions from existing nodes. (Used by Riak, Elasticsearch, Couchbase).
2. Dynamic Partitioning: Split partitions when they exceed a size (e.g., 10GB). Merge them when they shrink. (Used by HBase, MongoDB).
3. Proportional to Nodes: Have a fixed number of partitions per node. When you add a node, it splits a random existing partition and takes half. (Used by Cassandra).

Request Routing

How does the client know which IP address to connect to?

1. Round-Robin: Ask any node; if it doesn't have the data, it forwards the request.
2. Routing Tier: A separate load balancer (e.g., Mongos) knows the topology.
3. Client-Aware: The client library knows the topology (e.g., Riak Java Client). ZooKeeper is often used to keep track of this cluster metadata.