---

name: infrastructure-storage description: "Configure and troubleshoot SkyPilot clouds, Kubernetes, Slurm, SSH node pools, resources, GPUs, workspaces, storage, and volumes." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Infrastructure And Storage

Use this sub-skill when the user needs to configure, select, or troubleshoot SkyPilot infrastructure and data surfaces: cloud credentials, Kubernetes contexts, Slurm clusters and partitions, SSH node pools, workspaces, GPU/resource availability, sky check, sky gpus, object-store storage, file mounts, and volumes.

Start Here

1. Read references/infrastructure-reference.md for supported infrastructure forms, resource-selection rules, GPU/catalog commands, workspace behavior, object-store storage, and volumes.
2. Use scripts/resource_checklist.py to print a no-credential preflight checklist for a provider, Kubernetes, Slurm, SSH node pool, GPU, or storage task.
3. Read references/troubleshooting.md for disabled clouds, missing credentials, quota/capacity failures, Kubernetes context issues, Slurm partitions, SSH pools, storage mount errors, volume/PVC failures, and GPU catalog interpretation.

Owns

• Infrastructure selection with resources.infra, --infra, sky check, sky gpus list, sky gpus label, and provider setup triage.
• Cloud/resource concepts: cloud, region, zone, accelerator, CPU, memory, instance type, spot, disk, local disk, ephemeral storage, disk tier, network tier, and labels.
• Kubernetes usage: kubernetes, k8s, k8s/<context>, kube contexts, namespaces, pod config, GPU labels, service accounts, ports, priorities, autoscaler hints, PVC-backed volumes, and multi-context routing.
• Slurm usage: slurm, slurm/<cluster>, slurm/<cluster>/<partition>, partition/GPU/resource mapping, Pyxis/FUSE support, and scheduler-specific symptoms.
• SSH node pools: ssh, ssh/<pool>, node-pool setup expectations, host volumes, and Kubernetes-compatible implementation details.
• Data surfaces: workdir, file_mounts, SkyPilot Storage objects, MOUNT, COPY, MOUNT_CACHED, cloud object-store backends, sky storage, sky volumes, persistent volumes, ephemeral volumes, and mount-mode troubleshooting.
• Safe credential triage that explains which command to run locally without asking the user to paste secrets or credentials.

Route Elsewhere

• Task YAML syntax, resources, file_mounts, volumes, envs, and secrets field construction belongs to ../task-yaml/SKILL.md.
• Interactive cluster launch, status, logs, queue, stop/start/down, autostop, and cleanup belongs to ../cluster-operations/SKILL.md.
• Managed jobs, recovery, queues, pools, and job logs belongs to ../managed-jobs/SKILL.md.
• SkyServe service YAML operations, replicas, readiness, updates, and serving logs belongs to ../serving/SKILL.md.
• Python SDK request IDs, API server login/status, dashboard, and remote server compatibility belongs to ../sdk-api-server/SKILL.md.
• Repository source changes, tests, formatting, and contribution workflow belongs to ../repo-development/SKILL.md.

Safe Preflight

Run the bundled checklist helper to plan what the user should verify locally:

python sub-skills/infrastructure-storage/scripts/resource_checklist.py \
  --cloud k8s --gpu H100:1 --storage volume --include-triage

The helper prints checks and SkyPilot commands only. It does not import SkyPilot, read credentials, inspect kubeconfig, contact cloud APIs, or mutate infrastructure.

Operating Principles

• Prefer requirements over provider pins: ask for accelerator, CPU, memory, disk, and storage intent first; pin --infra, region, zone, instance type, Kubernetes context, Slurm partition, or SSH pool only when the user has a concrete placement or compliance requirement.
• Use sky check <infra> for credential/config readiness, sky gpus list for catalog and availability, and sky launch --dryrun for optimizer planning; do not parse pricing tables manually to choose the cheapest region.
• Separate authentication from capacity: credential and disabled-cloud errors are fixed with provider setup, while quota/capacity/stock errors are fixed by changing resource requirements, failover options, region/zone pins, or cloud choices.
• Treat object-store storage, file_mounts, and volumes as different tools: object stores are cross-region/cloud data surfaces, file mounts sync paths into a cluster, and volumes are high-performance provider/Kubernetes/RunPod-backed storage with lifecycle rules.
• Never ask the user to paste secrets, tokens, kubeconfigs, SSH private keys, or full cloud credential files; ask them to run local checks and share redacted error categories instead.