StarRocks 3.0 Storage‑Compute Separation Architecture: Design, Implementation, and Evaluation

The article introduces StarRocks 3.0’s storage‑compute separation architecture, beginning with three industry examples (AWS S3‑based shared storage, Snowflake SaaS architecture, and AWS Redshift) that illustrate different approaches, benefits, and limitations of separating storage and compute.

It then describes StarRocks’ own design, including the cloud‑native StarOS platform, the transformation of the traditional BE node into a stateless Compute Node (CN), and the role of StarManager in managing tablets, shards, services, files, and logs.

Key implementation details are covered: caching remote data locally to mitigate high I/O latency, enforcing immutable files for version control, ensuring write idempotency to avoid data corruption during node failures, and employing multi‑version management for consistency.

The article analyzes challenges such as heterogeneous storage support, low‑latency I/O over distributed storage, metadata consistency, cache and compute scheduling, and compaction/GC coordination in a stateless environment.

Benefits of the separation are quantified, highlighting cost reduction (fewer replicas and cheaper storage media), improved reliability and availability (leveraging cloud storage SLAs), and enhanced elasticity through resource isolation similar to Snowflake’s warehouse model.

Performance evaluations include StreamLoad ingestion throughput, TPC‑DS query benchmarks on a 1 TB dataset, and real‑world user case studies that demonstrate comparable latency to integrated architectures while achieving lower cost and better scalability.

The article concludes that StarRocks’ storage‑compute separation addresses many of the limitations observed in existing solutions, offering a flexible, cost‑effective, and high‑performance platform for modern analytical workloads.