How Baidu’s Search & Recommendation Engines Achieved Cloud‑Native Efficiency

Background and Motivation

Since 2022, Baidu’s Mobile Ecosystem Group (MEG) has been progressively refactoring its user‑facing products toward a cloud‑native architecture. The goal is to improve both resource‑utilization efficiency and development efficiency for large‑scale, complex services such as search and recommendation.

Core Focus on Two Efficiencies

Resource utilization efficiency : By containerizing services and applying dynamic resource scheduling, Baidu aims to achieve finer‑grained mixing of workloads, leading to more balanced cluster usage and lower hardware costs.

Development efficiency : Micro‑service decomposition reduces inter‑team coupling, allowing independent iteration. Standardizing common infrastructure capabilities on a cloud‑native platform raises the baseline for new services.

Service Design Standards

To move from simple cloudification to true cloud‑native, Baidu defined three design principles:

Micro‑service granularity : Each service must stay within a defined size limit.

Container encapsulation : A service should depend only on the underlying platform and components inside its own container.

Dynamic management : Services must be adjustable at runtime without breaking SLA commitments.

Evaluation metrics include resource overhead, deployment time, package description compliance, exclusive resource dependency, naming‑service accessibility, fault tolerance, and service‑state observability.

Phased Cloud‑Native Migration

The migration is divided into five stages:

Micro‑service transformation : Split monolithic services (e.g., Feed video recommendation) into smaller modules, reducing CPU and memory usage, cutting latency by >50 ms, and improving iteration speed.

Containerization : Package all dependencies inside containers, enabling one‑click deployment and eliminating physical‑machine constraints.

Dynamic management : Implement elastic scaling, automated failover, and resource‑pool sharing to support rapid capacity adjustments for traffic spikes.

Advanced cloud‑native : Introduce Serverless and Function‑as‑a‑Service (FaaS) to further lower cost and push development efficiency to its limit.

Declarative architecture : Move toward a model where business logic is declared and the system automatically derives and executes the necessary functions, separating architecture from application code.

Key Outcomes

After micro‑service refactoring, the Feed recommendation service was split into over ten modules, reducing the largest module’s memory footprint to <40 GB, cutting overall CPU usage by 23 % and memory usage by 84 %, and improving stability from <99.9 % to 99.97 %.

Dynamic management enables on‑demand scaling within a shared resource pool, eliminating the need for manual provisioning of physical machines and allowing idle resources to boost recommendation quality during low‑traffic periods.

Future Directions

Baidu plans to deepen declarative architecture, letting developers describe desired business outcomes while the platform automatically orchestrates functions, services, and resource allocations. This vision aims to further decouple business logic from infrastructure, achieving maximal flexibility and scalability.