Volcano

A Kubernetes-native batch scheduler that adds gang scheduling, queues, and fair-share to clusters running AI, ML, and big-data workloads.

• Category: Orchestration & Scheduling
• CNCF maturity: Incubating
• Language: Go
• License: Apache-2.0
• Repository: volcano-sh/volcano
• Documented at commit: 7110813 (master, 2026-06-24)

What it is

Volcano is a batch scheduling system for Kubernetes. It runs as a separate scheduler process alongside the default kube-scheduler and claims pods through their schedulerName, then makes its own placement and bind decisions. The scheduler is built on top of the Kubernetes SIG-Scheduling kube-batch project.

The core problem it solves is that the default scheduler places one pod at a time. That is wrong for distributed training and big-data jobs, where a group of pods must all start together or none should start at all. Volcano adds gang scheduling (all-or-nothing placement), queues with fair-share and capacity policies, and a pluggable set of scheduling algorithms. It ships a Job CRD that manages the full lifecycle of a multi-task batch workload.

Beyond the scheduler, Volcano runs a controller manager that reconciles its CRDs, an admission webhook manager, and an optional node agent for colocation and QoS of mixed online and offline workloads.

When to use it

• You run distributed training (PyTorch, TensorFlow, MPI/Horovod) or Spark/Flink jobs where partial pod placement wastes resources or deadlocks the job.
• You need queues with fair-share, capacity, or hierarchical quota across teams sharing one cluster.
• You want topology-aware or NUMA-aware placement, or scheduling for GPU/NPU and other scalar devices.
• It is a poor fit when you only run long-lived services and Deployments. The default kube-scheduler already covers that, and Volcano adds operational overhead.

In this deep-dive

• History: origin at Huawei, the kube-batch lineage, and the path through CNCF.
• Architecture: the four processes and how one scheduling cycle flows.
• Adoption & Ecosystem: who runs it and the alternatives.
• Internals: the session and statement transaction model, read from source.
• Getting Started: install and run a first VolcanoJob.

Sources

1. volcano-sh/volcano (README, LICENSE, source): https://github.com/volcano-sh/volcano
2. Pinned commit 7110813: https://github.com/volcano-sh/volcano/commit/7110813b198e99d0282170ef022f51ceb43d9403
3. Cloud Native Batch System Volcano moves to the CNCF Incubator: https://www.cncf.io/blog/2022/04/07/cloud-native-batch-system-volcano-moves-to-the-cncf-incubator/
4. Volcano project page (CNCF): https://www.cncf.io/projects/volcano/
5. Why Spark chooses Volcano as built-in batch scheduler on Kubernetes: https://www.cncf.io/blog/2022/06/30/why-spark-chooses-volcano-as-built-in-batch-scheduler-on-kubernetes/
6. Volcano adopters list: https://github.com/volcano-sh/community/blob/master/adopters.md