Version Overview

Version Change Description

New Features

Table 1 openFuyao v26.03 New Feature Information

Enhanced Component	Component Change Type	New Features	SIG
Cluster-API	Enhancement	1. Provide component health check interface bootstrap node image repository self-dependency 2. Installation and deployment supports node pre/post operations 3. Installation and deployment BKECluster CR split, supports individual CR O&M for each node 4. Installation and deployment architecture optimization 5. Support deploying chart form plugins	sig-Installation
env check	New	Provide installation and deployment pre-check tool	sig-Installation
hermes-router	Enhancement	New disaster recovery design	sig-ai-inference
Eagle eye	Enhancement	New network static metrics (A2/A3 generation host side & card side RDMA, host side PCIe bandwidth, etc.), implement load - NPU device ID association, and some device sub-health metrics (e.g., overload downclocking).	sig-ai-inference
Elastic scaler	New	1. Time-based tidal scheduling decision algorithm 2. General scaling decision framework 3. Dynamic scaling scheduling for PD separation scenario	sig-ai-inferenc
Basic Container Platform	Enhancement	Build openFuyao unified logging framework	sig-container-platform
npu-dra-plugin	New	1.Support Ascend NPU device resource discovery and reporting. 2.Support device filtering through DeviceClass/CEL. 3.Support using ResourceClaim/ResourceClaimTemplate for resource requests, implementing binding between business Pod and ResourceSlice. 4.Support injecting devices into containers through CDI.	sig-orchestration-engine
ub-network-device-plugin	New	1.Support URMA device management and on-demand allocation. 2.Support setting URMA device network namespace.	sig-orchestration-engine
matrix	New	1. Memory borrowing: Based on UB memory pooling mechanism, when node or numa memory usage rate in bare metal container scenario reaches preset value, trigger memory borrowing, sharing part of memory pressure to borrowed memory. 2. Memory sharing: Support importing/exporting memory blocks in UBS Server cluster through memory pooling capability, implementing cross-node and multi-process memory sharing on bare metal, while ensuring resource security and QoS through directory isolation and proxy layer.	sig-orchestration-engine
confidential-containers	New	Built based on Kunpeng TEE technology, through complete software stack of k8s+containerd+Kata+QEMU+KVM+CoCo, implementing confidential container deployment	sig-container-platform

Removed Features

Table 2 openFuyao v26.03 Removed Feature Information

Component Name	sig	Removed Features	Removal Reason
AI Inference Software Suite	sig-ai-inferance	Full features	AI Inference Software Suite aims to provide one-stop AI inference service deployment capability. This capability is already covered in Infernex project. Meanwhile, Infernex can provide more flexible deployment and other inference functions. Therefore, AI Inference Software Suite exits extension components in v26.03 version.

API Change Description

None

Version Feature Introduction

openFuyao master main functions are shown in Table 3. For specific information about functional features, please refer to User Guide.

Table 3 openFuyao Component Feature List

Category	Component Name	Feature Description
Container Platform	Installation and Deployment	Installation and deployment tool docking with standard Cluster-API, supporting one-click installation of business clusters. Management cluster provides multi-scenario interactive business cluster lifecycle management capabilities on unified management interface, including single/multi-node installation (including high availability), online/offline installation, cluster scaling, Kubernetes in-place upgrade, etc.
	Container Orchestration Core	Provides openFuyao Kubernetes, compatible with K8s 1.34, providing enhanced functions such as high-density deployment, startup acceleration, log enhancement, certificate management enhancement, etc.
	Management Interface	Provides out-of-the-box console, supporting application management, application marketplace, extension component management, resource management, repository management, monitoring, alerting, user management, command-line interaction, etc. 1.Authentication and authorization: Built-in OAuth2-Server, supporting OAuth2.0 protocol, supporting application authentication, authorization, password reset, password policy, etc. And provides unified authentication and authorization access scheme for frontend interface applications and non-frontend interface applications. 2.User management: Provides cross-cluster multi-user management capability, supporting binding of platform and cluster-level users with roles such as administrators, operators, and observers. 3.Command-line interaction: Provides command-line interaction popup for cluster administrators on cluster management interface, enabling administrators to conveniently manage clusters directly through background kubectl commands on console. 4.Application marketplace: Application marketplace supports browsing, finding, and deploying Helm-based extension components and applications, and provides computing power acceleration kits to unleash surging computing power. 5.Application management: Integrates Helm v3 application package manager, enabling quick deployment, upgrade, rollback, and uninstallation of applications. Can view Helm Chart details, resources, logs, events, and monitoring information. 6.Repository management: Provides built-in Harbor repository, supporting uploading and managing Helm Chart packages. Can add and delete remote Harbor repositories, and synchronize Helm Chart packages from remote Harbor repositories. 7.Extension component management: Dynamic pluggable framework developed based on ConsolePlugin CRD, supporting seamless integration of extension component frontend interface into openFuyao management interface, quick deployment through Helm Chart, and convenient upgrade, rollback, start/stop frontend interface and uninstallation operations; meanwhile supports convenient access of extension components to platform authentication and authorization system to ensure security, achieving plug-and-play of components. 8.Resource management: Resource management includes all Kubernetes core resources and custom resource definitions, facilitating user management (add, delete, query, modify). 9.Events: Reflect changes occurring to Kubernetes native resources such as Pod, Deployment, StatefulSet, etc. 10.RBAC management: Implement permission control for various cluster resources by setting ServiceAccount, Role, RoleBinding. 11.Monitoring: Provides out-of-the-box metric collection and visualization capabilities, supporting monitoring of resources such as clusters, nodes, workloads, etc., and providing out-of-the-box monitoring dashboards. 12.Alerting: Used to monitor various states in the cluster and trigger alerts when specific conditions are met,及时发现 problems, and take necessary measures to ensure system stability and reliability.
Independently Released Components	Online-Offline Colocation	Supports mixed deployment of online/offline businesses, guarantees scheduling of online businesses during peak usage periods and suppression of offline businesses, while enabling offline businesses to use oversold resources during online business trough periods to improve cluster resource utilization. Utilization improves by 30%~50%, with no significant impact on QoS and jitter below 5%.
	NUMA Affinity Scheduling	Implements hardware NUMA topology awareness at cluster and node levels, and performs NUMA affinity scheduling for applications based on NUMA affinity, improving application performance. Average throughput improvement reaches 30%, with redis performance improvement averaging 30%.
	Many-Core Scheduling	Implements anti-affinity scheduling and multi-dimensional resource scoring based on business type at cluster level. Container deployment density increases by 10% with performance degradation less than 5%.
	Ray	Provides high-usability, high-performance, high computing power utilization solution for Ray in cloud-native scenarios, supporting full lifecycle management of Ray clusters and jobs, reducing O&M costs, and enhancing cluster observability, fault location and optimization practices, achieving efficient computing power scheduling and management.
	KAE Operator	Implements minute-level automated management capability for Kunpeng KAE hardware, including KAE hardware feature discovery, automated management and installation of components such as drivers, firmware, hardware device plugins, etc. Can complete KAE deployment to availability within five minutes.
	NPU Operator	Implements minute-level automated management capability for Ascend NPU hardware, including NPU hardware feature discovery, automated management and installation of components such as drivers, firmware, hardware device plugins, metric collection, cluster scheduling, etc. Can complete NPU deployment to availability within ten minutes.
	Custom Monitoring Dashboard	Supports users customizing monitoring metrics according to their own business needs, achieving precise data observation and analysis.
	AI Inference Hermes Router	Intelligent routing EPP (Endpoint Picker) component built based on K8s GIE framework, supporting kv cache aware, bucket and other routing strategies, optimizing efficiency and performance of large language model (LLM) inference services by routing inference requests to most suitable backend service instances.
	PD Orchestration	Integrates three major capabilities: tidal algorithm, scaling decision framework, and dynamic PD scaling, covering multiple scenarios such as independent PD instance scaling, PD instance group proportional scaling, metric-driven scaling, tidal business scheduled scaling, etc., ensuring service availability during business traffic surges.
	AI Inference Eagle Eye	Observability system construction for AI inference scenarios (including hardware resource observability, business running state and system running state), supporting basic hardware disaster recovery and LLM acceleration routing capabilities.
	Multi-Cluster Management	Multi-cluster management can upgrade current cluster to management cluster, achieving multi-cluster management federation.
	Log	Collects various types of logs in the cluster, providing capabilities to view and download logs, and providing functionality to report alerts based on preset alert rules.
	URMA Network Device Plugin	Enables businesses to use URMA devices for communication, reducing communication latency and improving business performance.
	Ascend Dynamic Resource Allocation Plugin	Based on Kubernetes native DRA architecture, completed deep adaptation for Ascend NPU devices, enabling users to not only request NPU devices but also make scheduling decisions based on device attributes, computing power specifications and other metadata, thereby achieving higher performance and higher quality heterogeneous computing resource scheduling.
	UB Memory Pooling Component	For cross-node big data processing scenarios, avoids data replication through memory sharing capability, improving processing efficiency; for high memory density computing scenarios, improves node memory utilization and reduces hardware costs through memory oversubscription and borrowing mechanisms.
	High-Density Container	Provides strong isolation similar to traditional virtual machines, avoiding security issues between different containers.

Table 4 openFuyao Solution List

Solution	SIG	Description
Large-Scale Cluster	sig-large-scale-cluster	Completely includes intelligent routing, elastic scaling and decision system, observability, distributed KVCache management, and end-to-end one-click deployment capability, can flexibly configure installation of Hermes-router, Elastic-scaler, Eagle-eye, Mooncake components
inferNex	sig-ai-inference	Provides stable ultra-large-scale cluster, single cluster focuses on Kubernetes core component optimization, AI job scheduling optimization and TCP store key chain optimization, breaking through Kubernetes declared management upper limit, supporting 128k cards (16k nodes) cluster.

Version Overview ​

Version Change Description ​

API Change Description ​

Version Feature Introduction ​

Version Overview

Version Change Description

API Change Description

Version Feature Introduction