How Alibaba Conquered Double 11: Scaling to 17.5k TPS with High‑Availability Architecture

Background

Alibaba's e‑commerce platform grew exponentially over eight years, increasing transaction volume from 0.59 billion RMB in 2009 to 120.7 billion RMB in 2016 and peak QPS from 400 to 175 000. This created a world‑class “zero‑point” challenge: delivering maximum throughput and optimal user experience at the midnight peak with limited cost.

Technical Challenges

Key challenges included horizontal scalability, precise capacity planning, rapid cost growth, fine‑grained runtime control, and stability governance. The lack of external references forced Alibaba to innovate its own high‑availability middleware across several architecture generations.

Architecture Evolution

From a centralized 2007‑2008 design, the system migrated to a layered distributed architecture with shared services, caching, and storage clusters. Over time, the architecture incorporated multi‑datacenter active‑active deployment, cross‑region unitization, and a custom distributed traffic engine capable of generating tens of millions of QPS.

Capacity Planning & Online Stress Testing

Capacity planning began with baseline performance measurement, evolving from offline load tests to online traffic siphoning that gradually redirects traffic from many servers to a few to observe load‑response curves. Full‑chain online stress tests expose bottlenecks only under real‑world traffic, enabling thousands of issues to be discovered each year.

Unit‑Based Deployment & Hybrid Cloud

Alibaba partitioned the system into “units” that host complete buyer‑side services, allowing independent scaling, rapid failover, and seamless data synchronization. Hybrid‑cloud elasticity lets Alibaba provision massive resources for Double 11 and release them afterward, dramatically reducing cost.

Runtime Control, Flow Scheduling & Stability Governance

Dynamic throttling, downstream degradation, and traffic routing based on real‑time health metrics protect the cluster from overload. Middleware tracks call chains, aggregates stability metrics, and guides automated degradation decisions.

Future Directions

Future work aims for finer‑grained, data‑driven, and intelligent operations: per‑machine kernel allocation, predictive traffic modeling, and self‑decision making to achieve deterministic capacity and resource usage.