StarRocks Deployment and Practice at 360: Performance Evaluation, Use Cases, and Future Directions

Introduction: 360 selected StarRocks as its OLAP engine because existing engines (MySQL, Hive, Spark, Druid, ClickHouse) had limitations in scalability, latency, and operational cost.

Performance testing: on a 40‑core CPU, 128 GB RAM environment using the SSB 100 GB dataset, StarRocks showed comparable import speed to ClickHouse, lowest CPU usage, and superior query performance over Doris and ClickHouse in both single‑table and multi‑table tests.

Operational comparison: StarRocks and Doris have simple FE/BE architecture with auto‑scaling, while ClickHouse requires ZooKeeper and is more complex; StarRocks also supports multi‑tenant isolation and a rich set of features such as materialized views and external tables.

Use cases at 360: (1) Data analysis platform replacing MySQL, achieving sub‑2‑second response for billions of rows; (2) User portrait platform moving from Druid/Hive to StarRocks, leveraging bitmap functions and distinct count; (3) Search advertising reporting shifting from Hive/TiDB to StarRocks with pre‑aggregation and materialized views.

Explorations: integration of StarRocks with the cloud‑native data‑lake product “Yunzhou Data Warehouse”, performance comparison with Trino + Iceberg (StarRocks 1‑3× faster), and ongoing work to run StarRocks on Kubernetes by adding a compute‑only BE node and simplifying FE startup.

Conclusion and outlook: StarRocks offers a simple, high‑performance OLAP solution with strong ecosystem integration, but still has areas for improvement such as true compute‑storage separation and better K8s support.

Q&A highlights: smooth migration from Doris, role of Iceberg as storage format, and selection criteria between StarRocks and ClickHouse.