Apache Iceberg
Open table format for huge analytic tables, with time travel and schema evolution; the lakehouse layer that AI agents query through MCP servers
Overview
Apache Iceberg is "the open table format for analytic datasets": a free, open specification and library that turns files in object storage into reliable SQL tables that many engines (Spark, Trino, Flink, Presto, Hive, Impala) can read and write safely at the same time. It adds row-level updates and deletes, full schema evolution, hidden partitioning, time travel and rollback to past snapshots, and built-in compaction. Ryan Blue and Dan Weeks started it at Netflix in 2017; it entered the Apache Incubator in November 2018 and later became a top-level project. The v3 spec, ratified in 2025, adds deletion vectors, default column values, row lineage and VARIANT, geospatial and nanosecond timestamp types. It is on this list as a standard AI-pipeline component: an open lakehouse storage layer that analytics, ML and agent tools can share, with snapshots that make a dataset reproducible. AWS launched an MCP server for S3 Tables (managed Iceberg) in July 2025 so AI assistants can list and query Iceberg tables in plain language, and community MCP servers expose Iceberg catalogs through PyIceberg. Managed Iceberg is now offered by Databricks (which bought Tabular, the company of Iceberg's creators, in 2024), Google BigQuery, Snowflake, AWS S3 Tables and Cloudflare R2 Data Catalog.
The Verdict
Who Should Use Apache Iceberg?
Best For
- Data teams that want one copy of data readable by many engines
- AI and ML teams that need reproducible snapshots of training data
- Avoiding lock-in to a single warehouse vendor
Not Ideal For
- Small datasets that fit in a plain database
- Teams with no one to run compaction and snapshot clean-up
What's Great
- Open spec supported by Databricks, Snowflake, Google and AWS
- Time travel and rollback to any snapshot
- Schema and partition changes without rewriting data
Watch Out For
- Compaction, snapshot expiry and orphan clean-up must be scheduled
- Many small files slow queries
- Catalog choice matters, and managed catalogs can add lock-in
Pricing
View all features & details
Key Features
- Merge, update and targeted delete in SQL
- Full schema evolution and hidden partitioning
- Time travel and rollback by snapshot or timestamp
- Built-in compaction (bin-packing or sorting)
- v3: deletion vectors, row lineage, VARIANT and geospatial types
AI and Agent Access
- AWS S3 Tables MCP server (Jul 2025), read-only by default
- Community MCP servers built on PyIceberg
- v3 row lineage for auditing and change data capture
Real-World Usage
Industry Backing
- Created and run at scale at Netflix
- Databricks bought Tabular in June 2024 for over $1B, per its CEO
- Cloudflare R2 Data Catalog: managed Iceberg, beta April 2025
Community
- Releases 1.11.0 (May 2026) and 1.12.0 (Sep 2026)
- Repo pushed daily
How It Compares
| Feature | Apache Iceberg | Proprietary warehouse tables |
|---|---|---|
| Format | Open spec, many engines | Tied to one vendor |
| Cost | Free; pay for storage and compute | Bundled in warehouse pricing |
| Agent access | MCP servers (AWS, community) | Vendor tools |
| Maintenance | You schedule compaction, or use a managed catalog | Handled by the vendor |
Sources: Apache Iceberg · Releases · Incubator proposal · v3 (Aug 2025) · S3 Tables MCP server (Jul 2025) · MCP server docs · see also Cube and dbt
Is this your tool? Add a badge