k10s: A GPU-Aware Kubernetes Terminal Dashboard

k10s is a GPU-aware Kubernetes TUI. See which GPUs are actually doing work, which are burning money idle, and why your training job's ranks are scattered across the cluster. Vim keybindings. Single binary.

k10s.dev

Why k10s?

Most Kubernetes dashboards treat a GPU node like any other node. They have no idea an H100 costs $3/hr and is sitting at 4% utilization. k10s closes that gap.

• GPUs and jobs are the atoms, not pods. The default view is a fleet-level GPU dashboard with per-node utilization bars, memory, temperature, power draw, and workload attribution. Pods are an implementation detail you drill into when needed.
• Idle GPUs are loud, not quiet. Idle nodes sort to the top and glow amber. An unallocated H100 is $3/hr on fire. k10s makes that impossible to miss.
• Training job awareness. Group pods by Job, JobSet, RayJob, PyTorchJob, or MPIJob. See rank status, gang-scheduling state, and restart counts as one logical unit instead of 64 unrelated pods.
• Drill-down, not sprawl. Fleet > node > GPU > workload. Enter drills in, Esc goes back. Dedicated keys for context-filtered jumps: workloads (w), events (e), jobs (g).
• Works without DCGM. GPU count and workload mapping come from the k8s API. Install DCGM exporter for live utilization, memory, temp, and power. k10s degrades gracefully without it.

Roadmap

• Fleet view as default landing screen: per-node GPU count, utilization bars, workload attribution, idle detection from k8s API
• Loud-idle visual treatment: amber highlight for idle nodes, sorted to top by idle duration
• Node detail view: Enter on a node drills into per-GPU breakdown (index, utilization, memory, temp, power, workload, training rank)
• DCGM exporter integration: scrape Prometheus metrics for live GPU utilization, memory, temperature, and power; degrade gracefully without it
• Jobs view: group pods by parent training CRD (Job/JobSet/RayJob/PyTorchJob/MPIJob), show rank counts, status, restarts, Kueue queue
• Context-filtered jumps: w for workloads, e for events, g for jobs, all scoped to the current node
• Kueue queue integration: admission state, queue depth, pending workload visibility
• "Why is this GPU idle?" diagnostic: rule-based explanation of taint mismatches, resource fit, affinity, scheduling failures, PDB blocks

Installation

macOS (Homebrew)

brew tap shvbsle/tap
brew install k10s

Go Install

go install github.com/shvbsle/k10s/cmd/k10s@latest

Then run:

k10s

Usage

Keybindings

Fleet / Table Views

• j / ↓: Move down
• k / ↑: Move up
• h / ← / PgUp: Previous page
• l / → / PgDown: Next page
• g: Jump to top
• G: Jump to bottom
• Enter: Drill down (fleet → node detail → GPU → workload)
• Esc: Go back one level
• w: Workloads view, filtered to current node
• e: Events view, filtered to current node
• :: Enter command mode

Log View

• w: Toggle text wrapping
• t: Toggle timestamps
• s: Toggle autoscroll
• f: Toggle fullscreen
• Esc: Back to previous view

Commands

Press : to enter command mode:

• pods or po: All pods (all namespaces)
• pods <namespace>: Pods in a specific namespace
• nodes or no: All nodes
• namespaces or ns: All namespaces
• services or svc: All services
• jobs: Training jobs view
• quit or q: Exit

Development

Prerequisites

• Access to a Kubernetes cluster (via ~/.kube/config)

make build    # Build
make run      # Run
make test     # Test
make lint     # Lint
make fmt      # Format

Contributing

Contributions are welcome. Check the roadmap for planned work.

Community

Discord: https://discord.gg/rngaJustFD

License

Apache 2.0. See LICENSE for details.