Building a Data Lakehouse: Architecture Guide (2026)

A data lakehouse combines the low-cost storage of data lakes (S3 + Parquet) with the reliability and performance of data warehouses (ACID transactions, schema enforcement, SQL queries). In 2026, the standard lakehouse stack is object storage + Apache Iceberg + a query engine (Trino, Spark, or DuckDB).

Lakehouse Architecture

┌─────────────────────────────────────────────────────┐
│                  Query Engines                       │
│   Trino  │  Spark  │  DuckDB  │  Snowflake  │  BigQuery │
├─────────────────────────────────────────────────────┤
│              Apache Iceberg (Table Format)           │
│   ACID  │  Schema Evolution  │  Partition Evolution  │
├─────────────────────────────────────────────────────┤
│              Object Storage (S3 / GCS / ADLS)        │
│   Parquet Files  │  Metadata Files  │  Manifest Files │
└─────────────────────────────────────────────────────┘

Adding Real-Time with a Streaming Layer

Kafka / CDC ──→ RisingWave ──→ Iceberg Tables ──→ Query Engines
                    │
              Materialized Views (real-time serving)

RisingWave provides the streaming layer: ingest from Kafka/CDC, transform with SQL, serve real-time queries via PostgreSQL, and sink to Iceberg for historical analytics.

Lakehouse vs Data Warehouse

Aspect	Data Warehouse	Data Lakehouse
Storage	Proprietary	Open (S3 + Parquet)
Cost	High	Low (S3 pricing)
Format lock-in	Yes	No (open formats)
Multi-engine	No	Yes
Real-time	Limited	Via streaming layer
ML/AI workloads	Limited	Native (Python, Spark)

Frequently Asked Questions

Is a lakehouse cheaper than a data warehouse?

Yes, typically 5-10x cheaper for storage. S3 costs ~$0.023/GB/month vs proprietary warehouse storage. Compute costs depend on query engine choice.

Can I add real-time to an existing lakehouse?

Yes. Add RisingWave as a streaming layer that sinks to your existing Iceberg tables. No changes to your query engines or dashboards — they continue reading from Iceberg.