Meituan-Dianping Kubernetes Cluster Management and Optimization Practices

Background: Meituan‑Dianping, a leading Chinese lifestyle service platform, experiences extreme traffic peaks during holidays and promotions, requiring highly elastic and cost‑effective cluster resources to maintain user experience.

The company introduced a custom cluster management and scheduling system called HULK, evolving from Docker‑based elastic scaling (HULK 1.0) to a Kubernetes‑integrated solution (HULK 2.0) to improve resource utilization and operational efficiency.

Architecture Overview: HULK abstracts business‑level resource requests through a unified HULK API, decoupling the upper‑layer services from the underlying Kubernetes platform while preserving compatibility with native Kubernetes APIs.

Why Kubernetes? The platform offers a robust, extensible foundation for large‑scale, elastic workloads, enabling rapid deployment, dynamic scaling, and sophisticated scheduling policies.

Cluster Operations Status: Over 100,000 online instances across multiple regions, with comprehensive monitoring, health alerts, automated inspections, visual dashboards, and capacity planning powered by rule‑based and machine‑learning predictions.

Kube‑Scheduler Optimizations: Upgraded from version 1.6 to 1.10+, achieving >400% performance gain; introduced the alwaysCheckAllPredicates option to abort pre‑filter checks early, reducing scheduling latency by up to 40%.

Local‑Optimal Scheduling: Adopted a “best‑of‑N” approach (e.g., evaluating 100 candidate nodes instead of all 1000) to obtain near‑optimal placement with significantly lower computation.

Kubelet Enhancements: Added a “Reuse” restart policy alongside the native “Rebuild” policy, implemented a custom CNI plugin for IP reuse, and constrained aggressive eviction to improve risk controllability.

Resource Management & Optimization: Employed service profiling, affinity/anti‑affinity rules, scenario‑based priority, elastic scaling, fine‑grained resource allocation (NUMA binding, CPU‑Set), image distribution optimizations (cross‑site sync, base‑image pre‑distribution, P2P sharing), and online‑cluster techniques such as Numa binding, application‑level throttling, and re‑scheduling.

Conclusion: Ongoing efforts focus on hybrid online‑offline deployment, intelligent scheduling, and high‑performance, secure container technologies to further boost resource efficiency and service SLA.