How Taobao Scaled from LAMP to Cloud‑Native: Architecture Evolution and Migration Best Practices

Early LAMP Architecture (2003)

When Taobao was founded, it adopted the popular LAMP stack to launch quickly: PHP for development, Linux as the OS, Apache as the web server, and MySQL as the database. Within three months the site went live with about ten application servers and a MySQL deployment using master‑slave replication (one master, two slaves).

Transition to Oracle + IBM Mini‑Mainframe (2004‑2006)

To support rapid growth, Taobao migrated to an enterprise‑grade solution based on Oracle databases running on IBM mini‑mainframes with EMC storage. Although costly, the performance was excellent, but the system began to feel the pressure of increasing traffic.

Key concerns at the time were how to design the system architecture, choose databases, select caching solutions, and build business services as traffic and transaction volume continued to rise.

Adoption of Java Open‑Source Stack (2006‑2007)

Inspired by eBay’s architecture, Taobao built a Java‑centric solution using many open‑source components:

JBoss as the application server.

Spring as the IOC container for business logic.

iBATIS as the ORM tool for database access.

Self‑developed ISearch engine to replace Oracle‑based product search, reducing database load by dumping Oracle data nightly and building indexes on a single server.

Self‑Built CDN and Distributed Cache (2006‑2007)

To improve user experience, Taobao built its own CDN for static assets such as product images and descriptions, bringing content closer to users and speeding up page loads.

In 2007, with daily transaction volume exceeding 100 million RMB and over one million transactions per day, Taobao introduced:

TDBM (the predecessor of Tair) as a distributed cache for hot static data.

TFS, a self‑developed distributed file system deployed on dozens of x86 servers, replacing commercial NAS for storing images, descriptions, and transaction snapshots.

A distributed version of ISearch with 48 nodes for horizontal scaling.

Service‑Oriented Decomposition and HSF (2008)

To overcome the bottlenecks of a centralized Oracle architecture (connection limits, I/O performance), Taobao split the system by business domains (user, product, transaction, shop, etc.) into more than 20 business centers. Direct MySQL access was prohibited; all calls had to go through remote interfaces provided by the business centers via HSF (high‑speed service framework). Inter‑service communication used asynchronous Notify messaging.

Migration to Alibaba Cloud (2010‑Present)

Starting in 2010, Taobao unified its infrastructure on Alibaba Cloud, leveraging services such as SLB, ECS, RDS, OSS, ONS, and CDN. High‑availability features of the cloud enabled active‑active data center disaster recovery and modular deployment across regions.

Key technical challenges during the migration from the IOE (IBM‑Oracle‑EMC) stack to the cloud were:

Availability : Can a PC‑server‑based distributed cloud achieve the same redundancy as mainframes and high‑end storage?

Consistency : Can MySQL on RDS provide the same physical‑level consistency as Oracle RAC with shared storage?

Performance : Can RDS on commodity servers match the I/O throughput of high‑end storage, and how does MySQL performance compare to Oracle?

Scalability : How to split business logic, service‑ify components, decide on database/table sharding dimensions, and enable future re‑splitting?

Best‑practice solutions adopted on Alibaba Cloud include:

Stateless application design with extensive caching layers (browser, reverse‑proxy, page, fragment, object cache) and read‑write separation.

Service atomization and database partitioning.

Asynchronous processing to alleviate performance bottlenecks.

Minimizing transaction scope and, when necessary, relaxing strict consistency.

Automated monitoring and operations: alerting, unified configuration management, server/URL/network/module‑level monitoring, intelligent analysis, fault‑management platform, and capacity planning.

Practical Migration Patterns

File Storage : Replace EMC with OSS, supporting up to 40 PB of distributed storage; use multipart upload for large files.

Application Services : Replace IBM mini‑mainframes with SLB + multiple ECS instances; alternatively deploy middleware services such as ACE, ONS, OpenSearch.

OLTP Workloads : Use Alibaba Cloud RDS (up to 48 GB memory, 14 000 IOPS, 1 TB SSD) to replace Oracle + mainframe; consider adding OCS (Open Cache Service) for high‑traffic queries.

Read‑Heavy Scenarios : Deploy multiple RDS instances with read‑write separation; scale read replicas as needed.

Sharding : Horizontally split large tables across multiple RDS instances to improve performance and capacity.

OLAP Workloads : Adopt ODPS + OTS + RDS/ADS to replace the traditional Oracle‑based OLAP stack.

Overall Outcome

By migrating to RDS, introducing caching, applying database sharding, and employing read‑write separation, Taobao achieved a scale‑out architecture that surpasses the previous IOE setup in performance, scalability, and cost efficiency.