Container Service for Kubernetes:AHPA Controller release notes

How it works

AHPA runs two complementary scaling paths in parallel:

• Proactive prediction: AHPA analyzes historical metric data using machine learning algorithms from DAMO Academy and forecasts pod demand up to 24 hours ahead. Pods are pre-provisioned before a predicted traffic surge arrives. This path is effective for applications with periodic traffic patterns — daily peaks, weekly cycles, and similar rhythms.

• Passive prediction: When real-time metrics deviate from the predicted baseline — for example, during an unexpected traffic spike — AHPA adjusts the pod count immediately based on current data. This path provides a safety net for unpredictable load changes.

Service degradation lets you set maximum and minimum pod counts within one or more specific time windows.

The following figure shows the AHPA architecture.

The architecture comprises the following modules:

Supported metrics

AHPA supports the following metric types: CPU, memory, QPS (queries per second), RT (response time), and external metrics.

Scaling methods

AHPA applies scaling decisions through three target types:

For ACK Serverless clusters where all pods run on elastic container instances, AHPA enables zero-node auto scaling — the cluster scales at the pod level without managing node capacity.

Design principles

AHPA is built around three principles:

• Stable: Scaling triggers only when the application is in a stable state.

• O&M-free: No additional controllers are required on the client side. AHPA's configuration syntax is simpler than HPA's.

• Serverless-oriented: The focus is on pods, not on node resource utilization, making AHPA a natural fit for serverless and elastic container instance environments.

Usage notes

For configuration details, prerequisites, and YAML examples, see AHPA overview.

Release notes