Design and Implementation of a Cloud‑Native Recommendation System Architecture

The presentation introduces a cloud‑native recommendation system architecture, outlining three main content areas: the cloud‑native technology stack, the recommendation system architecture, and key design considerations for cloud‑native recommendation systems.

Cloud‑Native Technology Stack – The CNCF‑defined stack consists of four layers: Provisioning, Runtime, Orchestration & Management, and App Definition & Development, plus observability and analytics infrastructure. The speaker emphasizes using these layers to build foundational capabilities for recommendation systems, noting early adopters often built custom infrastructure before cloud‑native components matured.

Recommendation System Architecture – The system is divided into online and offline components. Offline pipelines handle content modeling, data ingestion, feature extraction, and vectorization for both items and users. Online services perform recall, ranking, and presentation, while traffic exhibits clear diurnal patterns.

Design Priorities for Cloud‑Native Recommendation Systems – Three hierarchical layers are proposed: (1) foundational cloud‑native infrastructure (PaaS, event mechanisms, service orchestration, profiling, metrics); (2) cloud‑native capabilities (ALM lifecycle, capacity management, SaaS resource management, scheduling, traffic control, chaos engineering); and (3) business value (cost reduction, development efficiency, stability, performance). The talk then dives into four focus areas:

1. Virtualization and Micro‑service Refactoring – Describes hardware‑assisted virtualization (HVM, KVM, Xen, VMware) and GPU virtualization for high‑density workloads. It explains why large monolithic services are split into fine‑grained micro‑services to improve resource utilization, enable automatic migration, and support self‑healing.

2. Service Governance and Elasticity – Introduces Application Lifecycle Management (ALM) for standardized service governance, capacity planning based on service‑specific load profiles, and dynamic quota resizing. Elasticity is achieved by building multi‑dimensional service portraits and adjusting resources according to traffic predictions.

3. Cloud‑Native Business Applications – Highlights how freed resources can be used for asynchronous computation, Nearline recall, and dynamic parameter tuning, improving recommendation effectiveness while leveraging idle capacity.

4. Stability Construction – Chaos Engineering – Explains the CNCF‑originated practice of injecting controlled failures to validate system resilience, using red‑blue experiments, fault libraries, and a resilience index to quantify stability and drive continuous architectural improvement.

The session concludes with a summary of the discussed topics and thanks the audience.