What Iceberg Standardized, and What It Didn't
Apache Iceberg is a table format: a metadata specification that tracks a table's schema, partitioning and file history so that multiple compute engines, among them Spark, Trino, Flink and Snowflake, can read and write the same underlying data files without duplicating them. By 2026 it is the closest thing the data platform market has to a settled standard, with native support across AWS, Google Cloud, Snowflake, Databricks, Dremio and Cloudera (Fivetran, 2026).
The specification keeps moving. Iceberg v3 added deletion vectors, row-level lineage tracking and native support for semi-structured VARIANT data. Databricks brought Iceberg v3 to general availability inside Unity Catalog on May 28, 2026, alongside open data sharing over the Iceberg REST protocol (Databricks, 2026).
What Iceberg standardizes is the file and metadata layer, not who is allowed to read or write it. That permission layer belongs to a catalog. In 2026, the catalog is where the competition actually sits.
Why the Catalog Fight Replaced the Format Fight
A catalog tracks which tables exist, who owns them, and which engine currently holds the lock to write. For years, each vendor ran its own, incompatible version: Databricks built Unity Catalog, Snowflake built its own metadata layer, AWS ran Glue. A table written through one vendor's catalog was difficult for another engine to safely write to, even when both spoke Iceberg underneath.
Both major vendors tried to fix this by open-sourcing their catalog technology. Snowflake introduced Polaris Catalog, an implementation of the Iceberg REST Catalog API, and donated it to the Apache Software Foundation's incubator in August 2024. Apache Polaris graduated to a top-level Apache project on February 18, 2026 (Snowflake, 2026). Databricks took a parallel path, open-sourcing Unity Catalog itself in June 2024 through the LF AI & Data Foundation, including its own implementation of the Iceberg REST Catalog API (Databricks, 2024).
Both moves were marketed as ending lock-in. Neither, on its own, made the two platforms interoperate with each other.
How Snowflake and Databricks Diverge in Practice
At Snowflake Summit 2026, held June 1 to 4 in San Francisco, Snowflake said its Horizon Catalog now runs on Apache Polaris and supports bidirectional read and write access. Under this model, an external engine such as Spark or Trino can write to a Snowflake-managed Iceberg table through Horizon's Iceberg REST Catalog implementation. Snowflake vends short-lived storage credentials so the external engine writes directly to the underlying object storage, while Snowflake itself commits the transaction and applies its governance rules to the result (Snowflake, 2026).
Databricks reached a comparable milestone earlier. At its Data + AI Summit in June 2025, the company announced that external Iceberg-compatible engines could read and write Unity Catalog-managed Iceberg tables through its own Iceberg REST Catalog API implementation (Databricks, 2025).
The two paths are not identical in practice. According to Onehouse, a vendor that builds lakehouse tooling independent of both platforms, tables registered as "Managed Iceberg" inside Unity Catalog cannot currently be pointed at an external Iceberg REST catalog such as Polaris, Nessie or AWS Glue. That means every external catalog stays read-only from Databricks' side. Onehouse also reports that several Unity Catalog features, Feature Store among them, still require a Delta-format table underneath, because they depend on primary-key constraints that managed Iceberg tables do not yet support (Onehouse, 2026). Onehouse competes in exactly this market, so it has a commercial reason to highlight the gap. The specific constraints it names are checkable against Databricks' own documentation, which is why they are cited here rather than the general claim of bias.
The result: two platforms that both now support open Iceberg REST catalogs, and each still has a version of its own managed data that behaves fully open only inside its own walls.
What to Check Before Assuming Portability
A vendor's claim of "open Iceberg support" is not a single fact. Buyers evaluating a warehouse or lakehouse on that basis should check four separate things before assuming their data is portable:
- Direction of write access. Can an external engine write to the vendor's managed Iceberg tables, or only read them? Read-only interoperability is a materially weaker claim than full read-write portability.
- Governance sync. Row- and column-level permissions set inside one platform's catalog do not automatically carry over to another catalog. Keeping access rules consistent across engines requires an explicit sync process, not just a shared table format.
- Feature parity on managed tables. Some platform features are still built against a specific underlying format. Ask which capabilities work on Iceberg-managed tables specifically, not on the platform as a whole.
- Catalog of record. Decide deliberately whether an external catalog such as Polaris or Glue is the source of truth, or whether the primary platform's own catalog is, and confirm the vendor's sync direction matches that choice.
The Adoption Gap Behind the Interoperability Race
The interoperability race is running ahead of mainstream production adoption. A January 2026 survey of 252 senior data and IT leaders, commissioned by lakehouse vendor Ryft and fielded by research firm TrendCandy, looked only at organizations already running Iceberg in production. Among that group, 79% planned to move their remaining data onto it within twelve months and 58% were using it for business-critical analytics (Ryft/TrendCandy, 2026). That sample was drawn entirely from existing Iceberg users, so it measures depth of commitment among adopters, not how much of the market has adopted at all, and Ryft sells tooling in this category.
A broader, vendor-independent read comes from the Practical Data Community's 2026 State of Data Engineering Survey, which found lakehouse architectures, spanning Iceberg, Delta Lake, Hudi and related formats, had reached a 27% foothold among respondents' production data platforms (Practical Data Community, 2026).
Put together, the table format question is settled and the catalog question is being fought hard, but mostly among vendors who already won the first argument. Most of the broader market has not yet had to pick a side.
Sources:
- Fivetran. "What is Apache Iceberg? Open table format explained." 2026. https://www.fivetran.com/learn/apache-iceberg
- Databricks. "Advancing Apache Iceberg on Databricks: Iceberg v3 GA, Open Sharing, and Unified Governance." May 28, 2026. https://www.databricks.com/blog/unity-catalog-and-next-era-apache-icebergtm
- Snowflake. "Apache Polaris: The End of Data Vendor Lock-In." 2024. https://www.snowflake.com/en/blog/engineering/apache-polaris-iceberg-rest-catalog/
- Snowflake. "Apache Polaris Graduates to Top-Level Project." February 2026. https://www.snowflake.com/en/blog/apache-polaris-top-level-project/
- Databricks. "Open sourcing Unity Catalog, creating the industry's only universal catalog for data and AI." June 2024. https://www.databricks.com/blog/open-sourcing-unity-catalog
- Snowflake. "Bidirectional Interoperability for Snowflake Iceberg Tables." June 2026 (Snowflake Summit 2026, June 1-4). https://www.snowflake.com/en/blog/engineering/bidirectional-interoperability-iceberg-snowflake-horizon-catalog/
- Databricks. "Databricks Eliminates Table Format Lock-in and Adds Capabilities for Business Users with Unity Catalog Advancements." June 11, 2025. https://www.databricks.com/company/newsroom/press-releases/databricks-eliminates-table-format-lock-and-adds-capabilities
- Onehouse. "Databricks Iceberg Support Has a Catch. It's Called Unity Catalog." June 2026. Vendor blog; Onehouse sells competing lakehouse tooling, disclosed in text. https://www.onehouse.ai/blog/databricks-iceberg-support-has-a-catch-its-called-unity-catalog
- Ryft / TrendCandy. "The State of Apache Iceberg in the Enterprise (2026)." January 2026. Vendor-commissioned survey of 252 senior data/IT leaders already running Iceberg in production, disclosed in text. https://www.ryft.io/blog/the-state-of-apache-iceberg-in-the-enterprise-2026
- Practical Data Community (Joe Reis). "2026 State of Data Engineering Survey." 2026. https://joereis.github.io/practical_data_data_eng_survey/