How Alibaba Cut Double‑11 Transaction Costs to One‑Quarter with Cloud‑Native Architecture

In celebration of the tenth Double 11, Alibaba Technology released the "Ten‑Year Code Chronicle" series, inviting senior engineers who helped prepare for each Double 11 to review the evolution of its core infrastructure.

Three Evolution Phases (2012‑present)

1. Technology Catch‑up (≤2012) : The focus was on solving immediate problems and ensuring system stability.

2. Technology Maturity (2013‑2014) : Massive technical debt was turned into assets, introducing full‑link stress testing, multi‑active deployments across regions, and early adoption of Docker and Kubernetes concepts (AliDocker/Zeus).

3. Technology Explosion (≥2015) : Rapid iteration accelerated, with large‑scale data‑center rollouts, full containerization of online services, unified scheduling, hybrid‑cloud elastic architecture, and storage‑compute separation, dramatically improving scalability and reducing costs.

Key Milestones

2013: Full‑link stress testing launched, providing deterministic stability verification across the entire business chain.

2014: Unit‑level multi‑active deployment enabled remote, cross‑region scaling for Double 11.

2015: Hybrid‑cloud elastic architecture and self‑developed infrastructure introduced.

2016: Containerization and unified scheduling deployed.

2017: Hybrid deployment and storage‑compute separation piloted.

2018: Base‑type data centers and full‑scale storage‑compute separation realized, achieving "datacenter as a computer".

Cost Reduction Model

The cost of a large‑scale promotion can be expressed as:

Promotion Cost = Resource Holding Time × Resource Holding Scale

By optimizing unified scheduling and cloud‑native architecture, the incremental cost per ten‑thousand transactions dropped by 50% each year from 2015 to 2018, ending at one‑quarter of the 2015 baseline.

Key formulas:

Daily Transaction Capacity = Peak Transaction Capacity – Total Elastic Capacity

Total Elastic Capacity = Hybrid Cloud + Hybrid Deployment + Scheduling Optimization

The strategy is to keep daily capacity low while leveraging abundant low‑cost elastic resources during peak events.

Unified Scheduling Cloud‑Native Architecture

Alibaba’s internal Sigma scheduler interacts with Alibaba Cloud via OpenAPI to request ECS resources managed by Houyi, isolates them in VPC networks, and runs PouchContainer workloads on large‑scale ECS instances. One‑click site creation deploys temporary transaction units, while hybrid deployment mixes online services with offline compute tasks, maximizing resource reuse.

Technical challenges include:

Storage‑compute separation (Pangu storage) to mask heterogeneous hardware differences.

Kernel‑level resource isolation (CPU, memory, I/O, network) with millisecond‑level priority scheduling.

Resource profiling and interference detection to protect high‑priority tasks.

Outcomes and Future Outlook

Unified scheduling and hybrid‑cloud deployment increased overall resource utilization, achieving over 45% CPU usage for online servers and reducing daily IT costs by roughly 30%. The combined effect of containerization, orchestration, and cluster management has become a cornerstone of the cloud‑native era, with further potential to drive industry‑wide standardization and enable near‑zero incremental cost for future large‑scale promotions.