How Alibaba Cloud’s AI Infra Innovations Are Transforming Kubernetes Workloads

KubeCon China 2025 recently concluded in Hong Kong, highlighting the rapid growth of Kubernetes and its role as a leading cloud‑native technology conference.

AI Infra Open‑Source Innovation and Optimization

Alibaba Cloud presented several advancements in AI infrastructure, focusing on efficient machine‑learning pipelines on Kubernetes using Argo Workflows. Recent updates include performance and scalability optimizations (Multiple Mutexes, Semaphores, Parallel Artifacts Resolving), Python SDK Hera support for easier workflow authoring, enhanced CronWorkflows for scheduled model training, and expanded AI and big‑data task support with plugins for Spark, Ray, and PyTorch.

Balancing Cost and Performance: Kubernetes Workflow Storage

Choosing the right storage solution is critical for large AI/ML models. Alibaba Cloud discussed the trade‑offs between integrated compute‑storage (e.g., MinIO, Ceph, 3FS) and separated storage (NAS vs. OSS), offering optimization techniques such as service‑side layout changes, lightweight clients, distributed caching, and server‑side acceleration.

Fluid: Cloud‑Native Elastic Data Abstraction

Fluid provides a distributed caching layer for elastic datasets, enabling seamless data access across AI workloads. Key components include Dataset, Runtime, and DataOperation, with recent updates like a lightweight Thinruntime and a generic CacheRuntime that simplify integration with storage vendors.

One‑Click Deployment of Open‑Source PD‑Separated Frameworks

Alibaba Cloud introduced RoleBasedGroup (RBG), a workload that automates the deployment of PD‑separated inference engines such as vLLM, SGLang, and Dynamo. RBG offers unified scheduling, declarative role templates, DAG‑based startup ordering, and dynamic scaling per role.

Multi‑Tenant Security and Stability Best Practices

To address multi‑tenant challenges, Alibaba Cloud demonstrated a fine‑grained ReBAC authorization mechanism using label/field selectors, CEL policies, and OpenFGA for dynamic access control, enhancing protection against lateral node escape attacks.

AI Gateway, Nacos, and Full‑Stack Observability

The AI Gateway, combined with Nacos for service discovery and Alibaba Cloud’s observability stack (SLS, ARMS, Tracing), provides real‑time monitoring, debugging, and reliable deployment for AI agents and model control planes.