How Alibaba Engineers Capacity Planning and Full‑Link Load Testing for Massive Sales Events

The Origin of Capacity Planning

Alibaba runs many distributed business systems; during large‑scale events such as Double‑11 it must determine how many machines each system needs. Capacity planning was created to answer when to add or remove machines, ensuring stability while minimizing cost.

Four‑Step Capacity Planning Process

1. Traffic forecasting – estimate future request volume using historical data.

2. System capacity assessment – compute a preliminary machine count for each system.

3. Fine‑tuning via full‑link load testing – simulate user behavior to adjust capacity.

4. Traffic control – set rate‑limit thresholds and protection measures to keep services responsive even if actual traffic exceeds forecasts.

Single‑Machine Load‑Testing Methods

Alibaba obtains per‑machine service capacity by conducting load tests directly in production. Four approaches are used:

Simulated requests (e.g., Apache ab, JMeter, LoadRunner).

Request replication – copying a request multiple times to a designated test machine.

Request forwarding – redirecting traffic from many machines to a single test machine.

Load‑balancer weight adjustment – increasing the weight of a test machine to receive more traffic.

Each method balances realism, data isolation, and operational impact.

Why Full‑Link Load Testing Is Needed

Single‑machine tests ignore inter‑service dependencies, which can cause cascading failures during peak events. Alibaba therefore built a full‑link load‑testing platform that reproduces Double‑11 traffic (over 10 million requests per second) in production while isolating test data in a shadow zone.

The platform consists of a control node and thousands of worker nodes, each running a custom load‑testing engine capable of handling the required request rate.

By extracting anonymized production data (buyers, sellers, items, promotions) and applying predictive models, Alibaba creates a realistic business model with more than 100 factors, generates corresponding requests, and feeds them to the engine.

Full‑link testing has uncovered hundreds of system issues each year, dramatically improving site stability during Double‑11 and other large‑scale promotions.

Unexpected Traffic Spikes and Flow Control

Even with accurate capacity models, actual traffic can exceed predictions, leading to overload, increased latency, request retries, and an avalanche effect where many machines become unresponsive.

Alibaba’s flow‑control framework monitors runtime status, call relationships, and applies flexible control strategies (drop, degrade, blacklist, queue) to protect the system.

The framework operates in three dimensions: runtime health, dependency chain, and control method, allowing dynamic balancing of response time, load, and QPS.

After applying the new algorithm, the system stabilizes within a defined range, and traffic recovers quickly once problematic machines are restored.

Impact and Future Availability

Full‑link load testing has become a cornerstone of Alibaba’s promotion‑readiness, significantly raising site reliability for Double‑11, Double‑12, and smaller events. The service will be launched on Alibaba Cloud in June, offering million‑plus requests‑per‑second capability, nationwide CDN‑based traffic generation, production‑environment isolation, and comprehensive performance diagnostics.