Kudu

Apache Kudu is the hot-tier store: low-latency upserts and analytical scans on weekly range partitions. Iceberg on RustFS holds settled weeks. See Query Engine.

Role in the Stack

Kudu stores the primary (hot) copy of analytical tables. Tables are created via Impala’s STORED AS KUDU syntax and managed by Kudu’s master server. In HMS-free mode, table metadata is registered in a PostgreSQL catalog registry rather than the Hive Metastore.

Impala DDL → KuduCatalogOpExecutor → Kudu Master → Tablet Servers
                                   → PG catalog registry

Services

Table Naming

Kudu tables created through Impala use the naming convention impala::<db>.<table>. This is stored as the kudu.table_name property in the metastore table object and in the PostgreSQL catalog registry.

Data Lifecycle

Tables begin in Kudu (hot tier) and can be migrated to Iceberg (warm tier) via CREATE TABLE ... AS SELECT through Impala:

-- Hot tier: Kudu (low-latency reads/writes)
CREATE TABLE db.events (
  event_id INT, ts BIGINT, data STRING,
  PRIMARY KEY(event_id)
) STORED AS KUDU;

-- Warm tier: migrate to Iceberg via Polaris
CREATE TABLE db.events_archive
  STORED AS ICEBERG
  TBLPROPERTIES('iceberg.catalog'='polaris')
AS SELECT * FROM db.events WHERE ts < 1710000000;

Build

Kudu requires both C++ binaries and a Java client:

• C++ binaries: Built from components/kudu/ with cmake/ninja. The Impala toolchain includes a pre-built Kudu client for ABI compatibility.
• Java client: kudu-client via devenv tasks run kudu:install-java into .devenv/m2 (project Maven repo), not ~/.m2.

Build dependencies: cmake, ninja, gcc, protobuf, flatbuffers, cyrus_sasl, openssl.