本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://jumping-code.com/2024/10/12/data-lakehouse-architecture/
Summary
Chinese-language explainer on the three generations of data analytics architecture: Data Warehouse (Gen 1), Data Lake + Data Warehouse two-tier (Gen 2), and Data Lakehouse (Gen 3). Based on the Databricks/UC Berkeley/Stanford Lakehouse paper (CIDR 2021), explaining why each generation emerged and what problems it solved.
關於三代數據分析架構的中文說明：數據倉庫（第一代）、數據湖+數據倉庫雙層架構（第二代）和數據湖屋（第三代）。基於 Databricks/UC Berkeley/Stanford 的 Lakehouse 論文（CIDR 2021），解釋每一代如何出現以及解決了哪些問題。
Key Points
- Gen 1 (Data Warehouse): structured data only, schema-on-write, compute+storage coupled; can’t handle unstructured data at modern scale
- Gen 2 (Data Lake + DW): data lake stores raw Parquet/ORC files (schema-on-read) + downstream warehouse for BI; Fortune 500 standard; problems: data staleness (86% analysts use stale data), reliability (ETL bugs), high cost (duplicated data)
- Gen 3 (Data Lakehouse): merge lake and warehouse; ACID transactions, schema enforcement, BI + ML from same platform; open formats (Parquet), compute-storage separation; implemented by Delta Lake, Apache Iceberg, Apache Hudi
- Key Lakehouse properties: metadata layer enables ACID over object storage, time travel, indexing for fast queries
- Still has optimization space: stream/batch unification, ML feature stores, governance
Insights
The Gen 2 critique is sharp: the two-tier lake+warehouse architecture solved the unstructured data problem but introduced new problems worse than Gen 1 in some dimensions (data staleness, reliability). Lakehouse is compelling specifically because it eliminates the ETL pipeline between lake and warehouse — the same Parquet files that serve ML jobs also serve BI queries, with transaction guarantees via the metadata layer. The Chinese tech community’s engagement with this topic (citing the original 2021 paper) reflects the architecture’s growing dominance in enterprise data platforms.
Connections
Raw Excerpt
Data Lakehouse 的核心概念是透過 metadata layer 來提供 ACID transaction 的支援，並且可以直接對 cloud storage 上的 open format 進行高效的查詢，同時支援 BI 和 ML 工作負載。