In order to provide scalability, Kudu tables are partitioned into units called tablets, and distributed across many tablet servers. A row always belongs to a single tablet. The method of assigning rows to tablets is determined by the partitioning of the table, which is set during table creation.

Choosing a partitioning strategy requires understanding the data model and the expected workload of a table. For write-heavy workloads, it is important to design the partitioning such that writes are spread across tablets in order to avoid overloading a single tablet. For workloads involving many short scans, where the overhead of contacting remote servers dominates, performance can be improved if all of the data for the scan is located on the same tablet. Understanding these fundamental trade-offs is central to designing an effective partition schema.

Kudu provides two types of partitioning: range partitioning and hash partitioning. Tables may also have multilevel partitioning, which combines range and hash partitioning, or multiple instances of hash partitioning.