Voda Scheduler
github.com
- API
- Self-hosted
What Voda Scheduler is
Voda Scheduler is a free, Apache-2.0-licensed GPU scheduler for elastic deep learning workloads in Kubernetes. Built around Kubernetes, Kubeflow Training Operator, and Horovod, it can scale distributed training jobs up or down while they run. Scheduling choices include FIFO, Elastic-FIFO as the default, SRJF, Elastic-SRJF, Tiresias, Elastic-Tiresias, FfDL Optimizer, and AFS-L. Topology-aware scheduling and worker migration can consolidate resources, while the project aims to use idle capacity efficiently during heavy demand. The scheduler tracks node additions and removals, works with an existing autoscaler, and is designed for failing nodes and spot instances that come and go. Deployment is described for cloud or on-premise Kubernetes clusters able to schedule GPUs. Documentation links to API endpoints and Prometheus metrics. The README says it was tested with Kubernetes v1.20 and warns that the project is experimental and may change significantly.
Who it suits
Voda Scheduler suits teams running elastic or distributed deep learning workloads on Kubernetes GPU clusters. It is intended for self-hosted deployments in cloud or on-premise environments.
Good to know, and what to watch
- Supports runtime scaling of distributed training
- Offers several scheduling algorithms
- Handles changing and failing nodes
- Exposes Prometheus GPU metrics
- Project is described as experimental
- README testing references Kubernetes v1.20
- Requires a Kubernetes GPU cluster
The verdict
Voda Scheduler targets GPU allocation for changing deep learning workloads, with multiple algorithms, topology awareness, and node-failure handling. Its experimental status and Kubernetes prerequisite matter when evaluating production adoption.
What the maker publishes
- API
- The project documentation links to API endpoints as part of its getting-started steps. source
- Changing nodes
- The scheduler is aware of node additions and deletions, works with an existing autoscaler, and is designed to tolerate failing nodes and spot instances that may come and go. source
- Compatibility
- The project says it was tested with Kubernetes v1.20. source
- Deployment
- The project says Voda Scheduler is designed to deploy in any Kubernetes cluster, on-cloud or on-premise, that can schedule GPUs. source
- Elastic training
- It supports dynamically scaling distributed training jobs up or down at runtime without interrupting training. source
- Fault tolerance
- The project lists fault tolerance as a feature. source
- Intended users
- The stated use case is elastic or distributed deep learning workloads running in a Kubernetes cluster. source
- License
- The GitHub repository identifies an Apache-2.0 license. source
- License and maturity
- The repository lists an Apache-2.0 license and warns that everything is experimental and may change significantly at any time. source
- Maturity
- The README warns that everything is experimental and may change significantly at any time. source
- Monitoring
- The README links to documentation for Prometheus metrics exposed by the scheduler. source
- Node changes
- It is aware of node additions and deletions, works with an existing autoscaler, and tolerates failing nodes. source
- Prerequisite
- The README says the scheduler was tested with Kubernetes v1.20. source
- Purpose
- Voda Scheduler is a GPU scheduler for elastic deep learning workloads built on Kubernetes, Kubeflow Training Operator, and Horovod. source
- Related tools
- The README names Kubeflow Training Operator, Kubeflow MPI Operator, Horovod, and a Prometheus NVIDIA SMI exporter among related projects. source
- Resource allocation
- The scheduler aims to use idle resources and allocate resources efficiently when a cluster is heavily loaded. source
- Scheduling algorithms
- The listed algorithms include FIFO, Elastic-FIFO (the default), SRJF, Elastic-SRJF, Tiresias, Elastic-Tiresias, FfDL Optimizer, and AFS-L. source
- Topology awareness
- Topology-aware scheduling and worker migration can consolidate resources to maximize cluster throughput. source
Checked against the maker's pages on 8 Oct 2026.