Distribute data using vnodes

In single-token architecture clusters, you must calculate and assign a single token to each node in a cluster. Each token determines the node’s position in the cluster, also known as the ring, and the portion of data assigned to the node according to the hash value. With vnodes, each node owns many small partition ranges distributed throughout the cluster. Partition ranges are based on the partitioner.

Virtual nodes use consistent hashing to distribute data without requiring new token generation and assignment each time you add, remove, or replace a node.

With a single-token architecture, you assign each node a single token that represents a location in the ring. A node owns exactly one contiguous partition range in the ring space. Each node stores the data mapped from the partition key to a token value within a range spanning from the previous node to the current node’s assigned value. Each node also contains copies of each row from other nodes in the cluster. For example, if the replication factor is 3, each range replicates to three nodes on the ring.

In contrast, when the ring has vnodes, the database randomly selects vnodes within the cluster, which makes the partition ranges non-contiguous. The hash of the partition key determines the placement of a row within many smaller partition ranges belonging to each node.

Make sure the number of vnodes is appropriate for your cluster and workloads; more vnodes doesn’t always mean better performance.

While vnodes provide considerable operational benefits, be aware that the number of vnodes you assign to any one node can impact cluster-wide operations. For example, when you increase the number of vnodes, you also increase the number of repairs that run during a repair cycle, which increases the duration of full cluster repairs.

For most workloads, DataStax recommends 8 or 16 vnodes. In performance tests, 8 vnodes distributed token ranges between nodes with approximately 10 percent variance and minimal impact on performance.