# Volcano 1.16 adds NamespaceQueue hierarchies, fairshare scheduling, and Unschedulable Job Cache

> **Key takeaway:** Volcano 1.16 introduces namespace-local queue hierarchies, Slurm-inspired fairshare scheduling, and an Unschedulable Job Cache that skips redundant scheduling checks to improve throughput under large pending backlogs.

**Published:** 2026-10-28T20:45:08+00:00 
**Source:** Cloud Native Computing Foundation (CNCF) 
**Category:** cloud-native 
**Canonical URL:** https://fosswire.org/news/volcano-116-adds-namespacequeue-hierarchies-fairshare-schedu.html 

## Summary
Volcano 1.16.0, the CNCF-backed Kubernetes batch scheduler, expands multi-tenant resource management and scheduling efficiency for high-performance computing and AI workloads. The release introduces NamespaceQueue, a namespaced resource that lets tenants build hierarchical queues and quotas within their own namespaces while administrators retain control over cluster-level resource pools. A new fairshare plugin brings historical resource consumption with usage decay into scheduling decisions, improving fairness when multiple namespaces share a queue.

On the performance side, the Unschedulable Job Cache records why Jobs cannot be scheduled and skips repeated checks until relevant conditions change, reducing redundant work in clusters with persistent backlogs. The HyperNode ecosystem also broadens: other schedulers can now deploy the HyperNode controller independently for topology discovery and supernode management. Additional changes include PodGroup topology anti-affinity for inference services, Kubernetes 1.36 support, DRA scoring, and bulk Pod dispatch load estimation.

## Scheduling architecture and multi-tenant queue changes in Volcano 1.16
- **Unschedulable Job Cache:** Records per-Job rejection causes and skips enqueue checks, whole-Job evaluation, or task-level checks across scheduling sessions; plugins provide event hints for wakeups on node label changes, resource releases, queue changes, storage events, and DRA events; bounded by a configurable watchdog defaulting to five minutes.
- **NamespaceQueue:** A namespaced CRD with capability, deserved, guarantee, and reclaimable fields; tenants manage queue hierarchies within authorized namespaces while cluster administrators control parent Queue authorization via spec.allowedNamespaces; avoids queue name collisions across namespaces.
- **fairshare plugin:** Applies Slurm-inspired usage decay to historical resource consumption when scheduling among namespaces that share a queue; improves long-term fairness for tenant-isolated workloads.
- **HyperNode controller decoupling:** The HyperNode controller can run standalone for topology discovery and supernode hierarchy management; exposes a HyperNode API for integration with non-Volcano schedulers.
- **PodGroup topology anti-affinity:** Required and preferred rules spread inference service instances across HyperNode domains to reduce shared failure domain exposure; complements existing gang scheduling semantics.

## Why it matters
For Kubernetes operators running AI training, inference, or HPC workloads, Volcano 1.16 reduces scheduling overhead in backlogged clusters while giving platform teams finer-grained multi-tenant controls. The standalone HyperNode controller opens topology-aware scheduling to broader scheduler ecosystems, and the fairshare plugin addresses a longstanding gap in proportional fairness for namespace-isolated tenants. Cloud providers and enterprise platform engineering teams building shared GPU or accelerator clusters will find the NamespaceQueue and performance improvements directly reduce operational complexity.