The lakehouse pattern is what happens when warehouses get too expensive and lakes get too messy.
For a decade the choice was data warehouse (structured, expensive, fast) or data lake (cheap, raw, slow). The lakehouse pattern unifies them: open table formats on cheap object storage that warehouses can query directly. By 2026 it is the default architecture for new enterprise data platforms.
What changed
Apache Iceberg and Delta Lake matured to the point where they are credible storage layers for production analytical workloads. ACID transactions, schema evolution, time travel, and partition evolution all work. Engines like Snowflake, BigQuery, Databricks, Trino, and Athena can all read these formats.
Apache Iceberg
Strengths: vendor-neutral, supported by every major engine including Snowflake (since 2024) and BigQuery (since 2025). Open governance under the Apache Foundation. Becoming the de facto standard.
Best for: enterprises that want zero vendor lock-in, multi-engine analytics, or cross-cloud strategies.
Delta Lake
Strengths: more mature ecosystem (predates Iceberg by 2 years), tight integration with Databricks, excellent CDC and merge-on-read support. Now also open-source under Linux Foundation.
Best for: Databricks-centric stacks, Spark-heavy ETL workloads.
The cost story
Storage on S3 is roughly $23/TB/month. Storage in Snowflake is roughly $40/TB/month. For petabyte-scale data, the lakehouse pattern saves real money. Add the fact that compute scales independently and you get the modern data lakehouse architecture: cheap raw storage in Iceberg on S3, transient compute via Trino or Spark or Snowflake on top.
The pattern that works
- Raw and intermediate data in Iceberg tables on S3 / GCS.
- dbt or Spark transforms run against the lakehouse.
- Curated mart tables either remain in Iceberg or land in Snowflake / BigQuery for the BI layer.
- Each engine queries what it is good at.
What we recommend
For new platforms in 2026: start with Iceberg. The vendor neutrality is worth the slightly less polished tooling. For Databricks shops with existing Delta investment: stay on Delta and watch the convergence (Databricks ships Iceberg support too).
The takeaway
The next five years of data platform architecture is the lakehouse architecture. Pure warehouse is too expensive at scale. Pure lake is too painful operationally. The lakehouse delivers both — and the open table formats finally make it portable.