Lakehouse Architecture
The lakehouse adds warehouse-grade transactions, schema, and governance to low-cost lake storage, aiming to serve both raw and curated workloads from one layer.
Closing the warehouse-lake gap
The lakehouse is an architecture that puts warehouse-like management on top of data-lake storage. It keeps the lake's cheap, open, any-format storage while adding the transactional guarantees, schema enforcement, and governance that were previously exclusive to warehouses. The goal is one system that serves both the flexible raw workloads of a lake and the reliable analytical workloads of a warehouse.
Open table formats
The enabling technology is the open table format: a metadata layer over columnar files that turns a directory of Parquet into a managed table. It maintains a transaction log recording which files make up the table at each version. This log is what provides the guarantees a bare lake lacks: readers see a consistent snapshot, and writers commit atomically.
What the table layer adds
- ACID transactions: concurrent readers and writers see consistent state
- Schema enforcement and controlled schema evolution
- Time travel: query the table as of a past version via the log
- Efficient upserts and deletes, not just appends
- Compaction of small files and pruning statistics for fast scans
Time travel and reproducibility
Because the transaction log records every version, a lakehouse table can be queried as of a past point, and a data-quality regression can be diagnosed by comparing versions. This built-in versioning of the table state complements dataset-level versioning and supports reproducible analysis: a result can pin the exact table version it read, and that version remains retrievable.
Where it fits
The lakehouse suits organizations that want to avoid maintaining a separate lake and warehouse with brittle copying between them. It pairs naturally with the medallion architecture, whose bronze, silver, and gold layers become managed tables in one system. It is not a universal replacement, very low-latency operational serving still favors specialized stores, but for unified analytics over both raw and curated data it removes a longstanding split. See warehouse vs lake, the medallion architecture, and columnar storage.