Why architecture matters here

Kudu fails on wrong partitioning + missing compactions. Architecture matters because tablet + Raft + columnar layers combine.

Advertisement

The architecture: every piece explained

The top strip is core. Table. Tablets. Raft replication. Columnar storage.

The middle row is usage. Upsert / update MVCC. Impala integration. Spark connector. Scan performance.

The lower rows are ops. Metrics. vs HBase / Iceberg. Ops — schema + partition + compact.

Kudu — tablet + Raft + columnar + Impala + upsertcolumnar store with upsertsTableprimary key + colsTabletsrange or hashRaft replication3-way defaultColumnar storageDiskRowSet + deltaUpsert / updateMVCCImpala integrationSQL on KuduSpark connectorread/writeScan performancepredicate pushdownMetricstablet statsvs HBase / Icebergdifferent sweet spotOps — schema + partitioning + compactionsmvccsqlsparkpushwatchcomparecompareoperateoperate
Kudu tablet + Raft + columnar storage.
Advertisement

End-to-end flow

End-to-end: time-series table partitioned by hash(user) + range(time). Upserts via Kudu client. Impala queries with predicate pushdown. Compaction merges DRSs + deltas. Raft ensures 3-way replication.