How ACK AHPA and Hybrid ECS/ECI Deployment Supercharge Ride‑Sharing Elasticity

Background and Challenges

Enjoy Mobility, a flagship brand of SAIC Group, supports the automaker’s “new four modernizations” (electrification, intelligent networking, sharing, and internationalization). Rapid business growth has expanded its infrastructure, making efficiency and cost control critical. The platform runs a Java‑based stack on Kubernetes, but native HPA and CronHPA cause minute‑level cold‑starts, cannot absorb unexpected traffic spikes, and require operators to manually script resource calculations, leading to high operational overhead and imprecise scaling.

Proposed Solution Overview

The team introduced two key components: ACK Advanced Horizontal Pod Autoscaler (AHPA) for predictive scaling, and a hybrid deployment of ECS (Elastic Compute Service) and ECI (Elastic Container Instance) to balance cost and performance.

ACK AHPA Predictive Autoscaling

AHPA learns from historical pod ready times and metrics, forecasting the required number of pod instances for each minute of the next 24 hours (1440 data points). By starting pods before traffic rises, the platform eliminates scaling latency, ensuring resources are ready when demand peaks.

For example, if weekend data shows a maximum of 16 pods for Service A, AHPA pre‑creates those pods in advance, preventing cold‑start delays.

Hybrid ECS/ECI Deployment

ECI provides a serverless container runtime that requires no underlying server management, ideal for burst traffic, though its per‑unit cost is higher than bare‑metal ECS. The hybrid model runs baseline services on subscription‑based or pay‑as‑you‑go ECS, while allocating burst workloads to ECI, achieving both cost efficiency and elasticity.

Custom Elastic Resource Priority Scheduling

To enforce deterministic scaling behavior, the deployment prioritizes resources in the order: subscription ECS → pay‑as‑you‑go ECS → ECI. During scale‑down, pods on ECI are terminated first, followed by pay‑as‑you‑go ECS, then subscription ECS. This is achieved by labeling nodes and defining a ResourcePolicy custom resource.

apiVersion: scheduling.alibabacloud.com/v1alpha1
kind: ResourcePolicy
metadata:
  name: DEMO
  namespace: demo-ns
spec:
  units:
  - max: 15
    nodeSelector:
      env: prd
    resource: ecs
  - max: 5
    nodeSelector:
      foo: bar
    resource: ecs
  - resource: eci
  whenExceedMax: NeverEvict

Sidecar Injection for Virtual Nodes

Virtual nodes cannot run DaemonSets, which are used for log collection agents on ECS nodes. To provide equivalent functionality, the team injects sidecar containers into pods scheduled on virtual nodes using OpenKruise’s SidecarSet via an admission webhook. This automatically adds sidecars to matching pods before they are scheduled to virtual nodes.

Note: SidecarSet is a core feature of Alibaba Cloud’s OpenKruise, enabling automatic sidecar injection for pods that meet label criteria, decoupling sidecar lifecycle from business containers.

Custom Elastic Metrics

Since CPU/Memory metrics may not reflect complex business workloads, the solution leverages AHPA’s External Metrics mechanism together with the alibaba-cloud-metrics-adapter to feed custom metrics into the autoscaling decision process.

Business Impact

After scaling the intelligent container elasticity platform, Enjoy Mobility achieved significant cost savings while maintaining stability and handling traffic spikes effectively. With a threshold set at 50, the system proactively scales when metrics approach the limit, delivering both reliability and measurable reduction in resource expenditure.