TechPrufer
Get started
Coreweave
Coreweave logoCoreweave · Feature trail

Serverless Kubernetes

1 post · Mar 2026 – Jul 2026

What this trail covers

CoreWeave SUNK (Slurm on Kubernetes) is enhanced with self-service capabilities and SUNK Anywhere for multi-environment deployment. This simplifies the creation and operation of AI research clusters by integrating Kubernetes primitives, schedulers, identity systems, and observability tools. New features include Automated User Provisioning (AUP) and SUNK User Provisioning (SUP) for streamlined access management, and GPU straggler detection for improved observability of distributed training performance. This post details the production-grade AI training capabilities of SUNK, emphasizing its intelligent, topology-aware scheduling, automated health management and failure mitigation, deep operational visibility, and unified lifecycle management. It highlights metrics such as up to 96% goodput, 10x longer MTTF for thousand-GPU clusters, and ~90 seconds to job restart, demonstrating its effectiveness at scale. The integration with CoreWeave Mission Control for continuous health management and observability is also emphasized.

All Coreweave trails
  1. CoreWeave SUNK adds new capabilities for modern AI research clusters | CoreWeave BlogThis post introduces self-service capabilities for CoreWeave SUNK, enabling faster cluster bring-up and more consistent operational behavior. It also details SUNK Anywhere, extending the unified training system to external infrastructure. New user provisioning features (AUP/SUP) automate identity synchronization and internal user setup. Enhanced observability includes GPU straggler detection to identify performance outliers in distributed training.
  2. A Slurm on Kubernetes Implementation for HPC and Large Scale AI | CoreWeaveThis post introduces SUNK (Slurm on Kubernetes), an implementation that integrates Slurm as a Kubernetes scheduler, allowing Slurm jobs to run within Kubernetes. It details the motivation for combining Slurm's batch job orchestration with Kubernetes' production application management, highlighting the efficiency gains from a unified compute pool. The post describes SUNK's features, including configuration via Helm charts, Kubernetes integration benefits (fast scheduling, containerization, HA, dynamic scaling), and state management through a Slurm syncer and cluster-wide operators. Key custom-built components – Nodeset, Syncer, and Scheduler – are explained, along with their roles in enabling two-way state syncing between Kubernetes and Slurm, dynamic node scaling, and scheduling Kubernetes workloads via Slurm states.
  3. How To Get Better Workload Balancing | CoreWeaveThis post details how CoreWeave's SUNK (Slurm on Kubernetes) implementation enables unified management of training and inference workloads on a single cluster. It highlights the benefits of reduced boundary between Slurm and Kubernetes for improved resource flexibility, simplified scaling by allowing dynamic resource allocation between training and inference, and enhanced ease-of-use by layering a familiar Slurm interface onto Kubernetes. The post also mentions the integration with CoreWeave Tensorizer for reduced model and checkpoint load times.
  4. Introducing Automated User Provisioning and SUNK User Provisioning | CoreWeave BlogIntroduces Automated User Provisioning (AUP) for CoreWeave IAM and SUNK User Provisioning (SUP) for Slurm on Kubernetes. AUP leverages SCIM for identity federation with enterprise IdPs (Okta, Microsoft Entra), ensuring continuous sync of users and groups. SUP automates the creation of POSIX users, SSH keys, and Slurm users/accounts within SUNK clusters based on CoreWeave IAM data. It utilizes NSSCache for resilient, local authentication, improving performance and reliability for large-scale HPC environments. This eliminates manual provisioning steps, reducing setup time from days to minutes.
  5. Kueue: Kubernetes-Native AI Workload Scheduling with CoreWeaveThis post introduces Kueue as a Kubernetes-native solution for scheduling AI training and batch inference workloads on CoreWeave Kubernetes Service (CKS). It explains the limitations of traditional Kubernetes for GPU-bound, 'all-at-once' workloads and details how Kueue addresses these through 'all or nothing' scheduling, fine-grained resource management using ResourceFlavors, and multi-tenant resource sharing via ClusterQueues and LocalQueues with features like quota borrowing and preemption. It also highlights the out-of-the-box integration of Kueue with CKS, including a Helm chart for installation and Prometheus metrics ingestion into Coreweave Hosted Grafana.
  7. Mission Control Simplifies Cluster Management | CoreWeaveThis post introduces CoreWeave Mission Control as a system that simplifies cluster management for AI enterprises. It highlights three key features: Fleet Lifecycle Controller for continuous hardware testing and validation, enhanced transparency into hardware state via managed bare metal compute, and proactive reliability through Node Lifecycle Controller (NLCC) and an observability platform. The post emphasizes the layered approach to reliability engineering and the role of FleetOps and CloudOps teams in maintaining cluster health and performance.
  8. Serverless Kubernetes: What It Is and How It Works | CoreWeaveThis post introduces the concept of serverless Kubernetes, explaining its benefits (infrastructure abstraction, pay-as-you-go pricing, easy auto-scaling) and contrasting it with traditional and managed Kubernetes. It highlights use cases like inference serving, distributed training, batch simulations, and rendering. The post also details how CoreWeave Cloud's Kubernetes-native platform, with its bare-metal performance, rapid instance spin-up times (as little as 5 seconds) due to container image caching and specialized schedulers, and high-bandwidth networking, enables serverless Kubernetes deployments.
  9. Tarteel Migrates to CoreWeave with Support from ZeetTarteel migrated its AI-powered Quran app infrastructure from AWS to CoreWeave, leveraging Zeet for deployment. This migration resulted in a 22% improvement in latency and approximately a 56% cost reduction. The post highlights Tarteel's use of NVIDIA Riva and NeMo toolkits on CoreWeave for training and deploying Speech-to-Text (STT) models, achieving real-time production performance and reproducible SOTA results. Zeet facilitated a swift migration with minimal downtime, enabling Tarteel to focus on product development.
  10. SUNK: Production-Grade AI Training at ScaleThis post details the production-grade AI training capabilities of SUNK, emphasizing its intelligent, topology-aware scheduling, automated health management and failure mitigation, deep operational visibility, and unified lifecycle management. It highlights metrics such as up to 96% goodput, 10x longer MTTF for thousand-GPU clusters, and ~90 seconds to job restart, demonstrating its effectiveness at scale. The integration with CoreWeave Mission Control for continuous health management and observability is also emphasized.