Design and Evolution of Baidu Comment Platform: Architecture, Performance Optimization, and Stability

Baidu Comment Platform provides a unified, high‑throughput comment capability for Baidu’s ecosystem, handling daily traffic at the hundred‑billion level and serving more than 20 products.

Background: Comments are a core user‑generated content channel that drives interaction and product stickiness; the platform must support diverse product forms and rapid feature iteration while maintaining stable performance.

Concept Introduction: The comment lifecycle includes production, moderation, storage, and retrieval. Data is sharded across distributed databases, with vertical and horizontal partitioning to meet read‑write scalability, and later aggregated into a big‑data warehouse for analytics.

Service Design: Originally built for Baidu App, the service was re‑positioned as a middle‑platform. A unified gateway layer offers traffic isolation, flow coloring, rate limiting, and fault detection. Interfaces are abstracted for low coupling and high cohesion, and a plug‑in framework supports access control, traffic billing, and after‑sales support.

4.1 Improving Performance and Iteration Efficiency: The list‑type API was refactored by separating responsibilities, parallelizing downstream calls, and introducing a DAG‑based scheduler that executes independent nodes concurrently, dramatically reducing 99th‑percentile latency.

4.2 Building High‑Performance, Low‑Latency Ranking: An offline scoring service computes comment scores using configurable formulas and stores results in Redis ZSET indexes. Online fine‑tuning adds personalized boosts and a fan‑out model in Go for full‑re‑ranking, enabling rapid A/B testing without impacting service latency.

4.3 Ensuring Continuous Stability: Multi‑level caching (local LRU, Redis, empty cache) mitigates cache penetration and breakdown. Hot‑spot detection automatically promotes heavily commented resources to a special pipeline, while disaster‑recovery and degradation mechanisms (active/passive triggers, circuit breakers, chaos engineering) protect the system during traffic spikes or downstream failures.

Conclusion: The platform now delivers peak QPS >400k, daily PV >100 billion, and SLA >99.995%. Ongoing work focuses on further innovation, middle‑platform expansion, and stability improvements to sustain a high‑quality community experience.