High Availability Practices: From Taobao to Cloud Migration

The speaker introduces the topic "High Availability Practices: From Taobao to Cloud Migration", describing how stability was achieved in Alibaba's e‑commerce platform and how those lessons apply to public‑cloud environments.

In the traditional Taobao shop system, peak traffic during Double 11 reached 20 million page requests, generating 4 million QPS with heavy RPC usage. The architecture relied on careful cache, database, and RPC design to balance performance, stability, and cost, avoiding over‑provisioning while maintaining user experience.

Performance tuning at the OS level used perf to identify hotspots such as excessive exception handling and JVM constant‑pool issues. Tools like BTrace helped replace inefficient third‑party libraries, and warm‑up strategies pre‑compiled hot methods (e.g., using Azure Zing ReadyNow or IBM J9 AOT) to eliminate start‑up latency.

Cache design emphasized high hit rates (e.g., 98 %) using a rich client that first checks a distributed cache before falling back to backend services, dramatically reducing backend load. Two deployment models were discussed: shared‑cluster across data centers for cost efficiency, and independent dual‑data‑center clusters for higher availability.

Limiting and degradation mechanisms were highlighted as essential for fault tolerance. Regular disaster‑recovery drills, such as cutting network connections or simulating database failures, validate that systems can survive component outages without user impact.

When migrating to Alibaba Cloud, the speaker presented classic failure cases (GitLab database deletion, AWS S3 command mishap) and stressed the need for regular backup verification and understanding third‑party dependencies. Cloud‑native high‑availability design incorporates multi‑AZ deployments, health‑checked load balancers, DNS failover, and cross‑region data replication.

Containerization and micro‑services enable horizontal scaling; by making front‑end and back‑end services stateless, traffic spikes can be handled by automatically provisioning additional instances, provided underlying storage remains healthy.

Overall, the talk stresses that high availability is achieved through a combination of robust architecture, proactive fault drills, intelligent caching, capacity planning, and leveraging cloud elasticity, while always preparing for the possibility that any component—including cloud services—may fail.