Iceberg Meetup Japan 4th Recap

Jan 21, 2026
iceberg

Iceberg Meetup Japan #4 Recap

Apache Iceberg continues to gain momentum in Japan. In 2025, two Iceberg-related books were published [1][2], and our community meetups regularly attracted over 200 participants.

To kick off 2026, we hosted the Apache Iceberg Meetup Japan #4 in January, featuring four presentations ranging from beginner introductions to advanced features.

Getting Started with Apache Iceberg

Saki Kitaoka (Databricks) gave a clear introduction to Open Table Formats and Apache Iceberg. She explained how Iceberg brings reliability and manageability to data lakes by organizing metadata at the file level, enabling ACID transactions, time travel, and schema evolution.

Her talk covered Iceberg’s architecture, highlighting its three logical layers and how metadata is managed for efficient read/write operations.

Key features:

  • ACID reliability and historical data management
  • Metadata design for large-scale analytics
  • Open, engine-agnostic architecture

Apache Iceberg V3: Current Status and Migration

Tomohiro Tanaka (AWS) discussed new data types and features in Iceberg V3, including unknown, variant, timestamp nanos, and geometry/geography types. He also covered Deletion Vectors, Row Lineage, Default Values, Multi-Argument Transforms, and Table Encryption.

The session focused on the variant type and Deletion Vectors. With Variant, you can use parse_json at write time and variant_get at read time. V3 also improves efficiency by reducing Puffin file size during DELETE operations.

Merge-on-Read performance gains depend on workload size, so benchmarking is recommended.

Iceberg Streaming Write at Repro

Tomohiro Hashidate (Repro) shared his experience evaluating Iceberg for streaming writes under mixed workloads, including queries fetching millions of records and single-record lookups.

Repro’s existing architecture uses Cassandra for write scalability and per-record updates, but it’s not ideal for bulk loads. After evaluating Hudi, they found Trino’s support limited and Spark compaction difficult to control.

To replicate Cassandra queries, they created a Trino view that UNIONs Kafka and Iceberg connectors. Single queries performed well, but further evaluation is needed for high concurrency scenarios.

High-Performance Analytics on Iceberg with StarRocks

Marko SUN (CelerData) introduced StarRocks, a high-performance MPP database for real-time analytics.

He explained how metadata planning can be a bottleneck with Iceberg, and how StarRocks optimizes planning by switching between distributed and local strategies based on metadata size.

×