Rokid

Solution

Rokid used Alibaba Cloud ACK with Knative to serverless-ize the online service layer behind its AR exhibition-guidance applications. Knative gives the team request- and concurrency-based scaling, traffic-based gray release, rollback, and multi-version management so new application versions can ship without rebuilding a large set of Kubernetes primitives for each rollout. To reduce the lag of conventional HPA scaling, Rokid also adopted ACK's AHPA prediction so pods can be prepared ahead of expected traffic increases rather than only after metrics cross a static threshold. The deployment mixes ECS and ECI capacity and applies a custom ResourcePolicy so normal load prefers long-lived ECS capacity while burst traffic spills into ECI, and scale-down releases ECI first. Together, those controls let Rokid pursue GPU efficiency, faster delivery, and more deterministic real-time service behavior for AR scene experiences.

Data flow

AR scene-experience requests enter Knative services on ACK, where traffic levels drive scaling and version-routing behavior. AHPA predicts capacity needs ahead of spikes, while ResourcePolicy determines whether workloads land on ECS or ECI resources before responses are returned to location-aware AR clients.