Apache Iceberg

Use it when

• Several engines must read and write one copy of the data: Spark, Trino, Flink, and most warehouses speak Iceberg through a catalog.
• Schemas and partition layouts will change over time; field IDs and hidden partitioning let both evolve without rewriting files.
• You need snapshots for time travel, rollback, and reproducible reads.
• Vendor neutrality matters: it is the convergence target other formats translate into, with the broadest catalog ecosystem.

Think twice when

• You do not want to run or choose a catalog; Iceberg requires one (REST, Hive, Glue, or Nessie), unlike Delta's self-contained log.
• Nobody will own table maintenance: snapshot expiry, orphan cleanup, and compaction are jobs you schedule.
• The table takes continuous high-frequency updates by key; an LSM-based format like Paimon, or Hudi, is built for that shape.
• You need secondary indexes; Iceberg prunes with min/max statistics and partitions only.

How it runs

Iceberg is a specification with libraries inside the engines, so there is nothing to deploy except the catalog that resolves table names. Data and metadata files sit on object storage; commits swap a metadata pointer atomically, giving serializable isolation by default. Maintenance runs as engine jobs you schedule. Nearly every lakehouse vendor offers managed Iceberg tables, catalogs, and maintenance.