How 58 Group Scaled AP Analytics with StarRocks: Benchmarks, Ops Tools, and Cloud Deployment

58 Group, a leading Chinese internet services provider, needed a high‑performance analytical platform to support diverse AP‑heavy scenarios such as security analysis, real‑time data warehousing, and business reporting, handling billions of rows and petabyte‑scale data.

Why StarRocks?

Initial tests of TiFlash revealed write‑throughput and read‑latency bottlenecks due to TiKV dependencies, making TiFlash unsuitable for pure‑AP workloads. Comparative evaluations showed StarRocks outperforming TiFlash, ClickHouse, and TiDB in both single‑table and multi‑table queries, with StarRocks 3‑7× faster than ClickHouse when low‑cardinality optimizations were enabled.

StarRocks Advantages

High write throughput (Broker Load >1.7M rows/s) and fast read performance.

Simplified operations: only FE, BE, and Broker nodes; easy cluster creation and scaling with a single command.

Rich data‑ingestion ecosystem: supports MySQL protocol, Spark, Flink, Hive, Kafka (Routine Load), Stream Load, and external tables.

Multiple table models (detail, aggregate, primary‑key) and strong vectorized execution.

Business Practice Cases

Security analysis: Ingested billions of log entries daily, switched from detailed to aggregated models with 15‑minute intervals, reducing storage by 75% and query latency from >2 s to ~300 ms.

DBA slow‑SQL monitoring: Collected fe.audit.log slow‑query entries via Filebeat → Kafka → StarRocks, built dashboards for daily trends, real‑time hot queries, and time‑range summaries.

Business reporting: Replaced a custom Infobright columnar store with StarRocks, achieving >90% query speed improvement and migrating 700+ tables.

Operational Tools Developed

Topology tool (qstarrocks): Displays cluster roles, IPs, ports, domains, VIPs, business lines, owners, and health status.

Cluster management (starrocks_manage): Deploy, scale, upgrade, start/stop, and monitor clusters with a unified CLI.

Kafka import task manager (starrocks_kafka): Create, rebuild, monitor status, latency, consumption, and SQL of Routine Load tasks; supports alerts.

Slow‑SQL manager: Aggregates slow‑query logs, provides day‑level trends, real‑time hot queries, and detailed SQL inspection.

Monitoring Architecture

Implemented Prometheus + Grafana for node health (SHOW FRONTENDS/BACKENDS) and performance metrics, feeding data into a central Zabbix instance for alerting. Custom dashboards visualize key server and instance metrics.

Cloud‑Native Deployment

Moved FE nodes from limited‑capacity VMs to Docker containers, defining two FE packages (8‑core/32 GB and 16‑core/64 GB) and four BE packages (up to 32‑core/128 GB/3.2 TB). Deployed ~7 cloud clusters, enabling elastic scaling and simplifying resource management.

Conclusion & Outlook

After a year of production use, StarRocks has proven stable, performant, and easy to operate for 58 Group’s AP analytics, delivering up to 90% query acceleration and significant storage savings. Ongoing work includes expanding cloud‑native tooling, intelligent host diagnostics, and broader resource‑pool management.