Apache Hudi vs Apache Iceberg

Verdict

Both are open-source table formats solving the same base problem — ACID transactions and schema evolution on data-lake files — but they were built around different write patterns. Apache Hudi was built at Uber specifically for high-volume, record-level upserts and deletes: the pattern behind change-data-capture ingestion and regulatory deletion requirements. It ships built-in indexing to accelerate those lookups and incremental queries so downstream jobs process only what changed. Apache Iceberg optimizes instead for broad, engine-agnostic adoption — its manifest-based metadata and independent Apache governance are why it has become a common interchange format across Snowflake, Databricks, Dremio and Starburst.

Choose Apache Hudi if

• Your pipeline is dominated by upserts and deletes — CDC ingestion from operational databases, or deletion-driven compliance work.
• You want incremental queries built in, so downstream consumers process only changed records without extra logic.
• You value Hudi's built-in table-management services (compaction, clustering, indexing) working out of the box rather than being assembled separately.
• Managed support matters: Onehouse, built by Hudi's original creators, is available if you'd rather not operate it yourself.

Choose Apache Iceberg if

• Multiple engines across your organization — Spark, Trino, Flink, or a cloud warehouse — need to read and write the same tables.
• Your workload is more append- and batch-heavy than upsert-heavy, so Hudi's upsert-optimized indexing buys you less.
• You want the format with the broadest current adoption across catalog and vendor ecosystems, which matters for long-term optionality.
• Partition evolution without rewriting existing data is a requirement you expect to hit.

What they share

Both are Apache-2.0 licensed, free, self-hosted-by-nature specifications and libraries rather than managed databases. Both provide ACID transactions, time travel and schema evolution, and both integrate with Spark, Flink, Presto and Trino.

The honest caveat

Both projects release frequently and are actively closing feature gaps with each other — Iceberg has expanded partition evolution, Hudi has continued improving indexing speed — so treat any specific performance claim as time-bound rather than permanent. If you're choosing for a new project with no strong upsert requirement, Iceberg's broader current adoption is the safer default for optionality; if CDC-style upserts are the core of the workload, Hudi's purpose-built tooling is worth the narrower ecosystem. See how to choose a lakehouse platform for how this decision fits into the bigger picture.