---

Project Overview

This project is built around three core components: Elastic Scaler provides general scaling decision capabilities, ResourceScalingGroup provides workload group scaling and resource orchestration capabilities, and Tidal provides time-rule-based tidal scaling capabilities.

Elastic Scaler - Scaling Decision Framework

Elastic Scaler is a general scaling decision framework that adopts a plugin mechanism to support custom algorithm and resource management logic extensions, while simultaneously supporting metric-driven and event-driven scaling semantics.

• Supports MetricsTrigger (metric-driven) and StateTrigger (event-driven) scaling.
• Supports custom scaling algorithm plugin development.
• Supports custom resource plugin development (extensible via ResourceScalingGroup).
• Compatible with Kubernetes native HPA capabilities.
• CRD API: elasticscaler.io/v1alpha1.
• MetricsTrigger supports four metric types: Resource, External, ExternalServer, and Custom (HPA algorithm path does not support ExternalServer).
• Built-in APA algorithm as the default scaling algorithm; when scalingAlgorithm is left empty, it defaults to APA.
• StateTrigger associates external resources (e.g., Tidal) via elasticscaler.io/name and elasticscaler.io/namespace labels, synchronizing their status.desiredReplicas.
• Stabilization strategies such as metric windows, aggregation, and missing values can be configured via Annotations (e.g., elasticscaler.io/metrics.windowSeconds).

ResourceScalingGroup - Resource Group Scaling

RSG targets Kubernetes workload group scaling scenarios, abstracting multiple resources into logical groups for unified orchestration, supporting two core strategies: GroupReplication and InplaceScaling.

• Manages a set of target resources via the standard CRD ResourceScalingGroup.
• Supports managing Deployment, StatefulSet, and LeaderWorkerSet.
• GroupReplication: Replicates, restores, and scales down by group, suitable for full-chain group scaling.
• InplaceScaling: Links follower replicas proportionally with leader replicas.
• Exposes the scale sub-resource under GroupReplication, which can be driven by upper-level strategies such as HPA/ElasticScaler/Tidal.

Tidal - Tidal Algorithm

Tidal provides time-driven scaling decision capabilities, generating target replicas at specified time points through the Tidal CRD, with scaling executed uniformly by Elastic Scaler.

• Provides time-driven scaling decisions via the standard CRD Tidal.
• Supports combination of multiple time rules and 6-field CRON (including timezone) configuration.
• When multiple rules conflict at the same time, resolves by the "maximum replicas" strategy.
• Provides observability information such as desiredReplicas, activeRule, lastTriggerTime, and nextTriggerTime through the Status field.

User Guide

• Elastic Scaler User Guide: AI Inference Elastic Scaling User Guide
• ResourceScalingGroup User Guide: AI Inference ResourceScalingGroup User Guide
• Tidal User Guide: AI Inference Tidal Scheduling Algorithm User Guide

Component Versions

Component	Version	Description
elastic-scaler	latest	Decision and trigger controller
resourcescalinggroup	latest	RSG resource group scaling controller
tidal	latest	Tidal scheduling controller
pd-orchestrator	0.0.0-latest	Three-component integrated deployment

Quick Start

Option 1: Local Helm Deployment

Execute at the repository root directory:

cd charts/pd-orchestrator
helm dependency build
helm install pd-orchestrator .

charts/pd-orchestrator/values.yaml supports enabling/disabling the three components as needed:

elastic-scaler:
  enabled: true
resourcescalinggroup:
  enabled: true
tidal:
  enabled: true

Option 2: OCI Registry Helm Deployment

No need to clone the repository; use the published Chart directly:

helm install pd-orchestrator oci://cr.openfuyao.cn/charts/pd-orchestrator --version 0.0.0-latest

To override default values as needed, append -f values.yaml or --set key=value.

Applicable Scenarios

• General elastic strategy scenarios: Implement metric-driven and event-driven scaling via elastic-scaler.
• Multi-resource coordinated scaling scenarios: Uniformly manage a set of target workloads via resourcescalinggroup.
• Business tidal fluctuation scenarios: Implement scheduled scaling via tidal.