GKE Batch & HPC Workloads
Configure and run queued, parallel, and high-performance workloads on GKE.
The skill names the Kubernetes operations it expects and includes useful safeguards such as resource requests, quotas, limits, and checkpointing. However, it exposes high-impact actions such as deletion and installation of remote latest/raw manifests without user confirmation, least-privilege guidance, sensitive-data flow disclosure, dependency verification, or rollback procedures, so substantial points are deducted.
The batch, JobSet, Kueue, MPI, and scaling material is broadly coherent and includes some retry and checkpoint guidance. It provides no tests, compatibility matrix, abnormal-input handling, or diagnosable failure feedback, and the availability and versions of remote CRDs are unverified; static calibration therefore limits the score.
The intended workloads and the boundary against standard web deployments are reasonably clear. Inputs, expected outputs, semantic trigger conditions, and non-fit cases remain under-specified, with no evidence of Chinese-language support or mainland-China network reachability, so points are deducted.
The document has a readable structure covering batch, HPC, cost, and production guidance, with installation snippets and examples. It lacks skill-level versioning, changelog, named maintenance responsibility, dependency versions, troubleshooting, FAQs, and explicit limitations; repository-level license and support information only partially offsets these gaps.
The Job, JobSet, Kueue, MPI, and node-pool snippets can provide useful starting material for GKE practitioners. However, images, clusters, and regions are placeholders, and the skill omits complete deployment, validation, permissions, networking, and cleanup workflows; no directly verified representative output is supplied, so only limited credit is warranted.
The source contains auditable commands, YAML, and configuration recommendations, while the README, license, and publisher provenance provide limited traceability. There is no committed test suite, CI execution evidence, result sample, or independent corroboration, so the static evidence supports only a low score.
- The commands use remote latest or unpinned YAML; review provenance, versions, permissions, and changes before execution.
- The listed delete_k8s_resource operation may be irreversible; require explicit confirmation, tightly scoped targets, and a backup or recovery plan.
- Kueue, JobSet, and MPI Operator CRDs are assumed but installation order, compatible versions, namespaces, IAM, and networking requirements are not documented.
- Spot capacity, maintenance exclusions, and autoscaling can affect cost, availability, and upgrade safety; validate checkpointing and recovery before production use.
- No evidence is provided for Chinese-language support or mainland-China network reachability.
What does this skill do, and when should you use it?
This skill is for running batch processing and high-performance computing workloads on Google Kubernetes Engine. It covers Kubernetes Jobs, JobSet, Kueue, compact node placement, and MPIJob. It also documents production guidance for Spot VMs, scale-to-zero node pools, resource quotas, maintenance exclusions, and checkpointing. It is explicitly not intended for standard web application deployments; those should use gke-app-onboarding.
It provides Kubernetes manifests and commands for Jobs, JobSets, ClusterQueues, LocalQueues, MPIJobs, and GKE HPC node pools. It shows how to install Kueue and the MPI Operator with kubectl, and how to create a compact-placement, autoscaling node pool with gcloud. It references apply_k8s_manifest, get_k8s_resource, describe_k8s_resource, get_k8s_logs, delete_k8s_resource, and list_k8s_events for Kubernetes resource and log operations.
- A data engineer running parallel batch-processing pipelines on GKE.
- A research or engineering team running CFD, molecular-dynamics, or financial-modeling simulations.
- An ML team submitting distributed training jobs on GKE.
- A platform team using Kueue to allocate CPU and memory quotas across tenants.
- An engineering team orchestrating distributed MPI applications with the MPI Operator.
What are this skill's strengths and limitations?
- Covers common GKE batch and HPC patterns from Kubernetes Jobs through JobSet, Kueue, and MPIJob.
- Includes copyable YAML, kubectl, and gcloud examples.
- Includes cost and resilience guidance for Spot VMs, scale-to-zero, quotas, and checkpointing.
- Examples contain placeholders such as `<IMAGE>`, `<CLUSTER_NAME>`, and `<REGION>` and are not production-ready as written.
- The source provides no test suite, version-compatibility matrix, or complete troubleshooting procedure.
- Some installation commands fetch manifests from external GitHub URLs, so execution depends on network access and cluster permissions.
How do you install this skill?
This skill is part of the google/skills collection. Run npx skills add google/skills and select skills/cloud/gke-batch-hpc during installation. The source does not document a standalone installation flow or additional local files.
How do you use this skill?
In an Agent Skills-compatible client with the skill installed, submit a concrete request such as: Configure a GKE batch Job with Kueue and parallelism 10, or Configure compact placement for a GKE HPC cluster and run an MPIJob. Execution also requires an available Kubernetes/GKE environment and the referenced tools.
How does this skill compare with similar options?
For standard web application deployments, the SKILL.md explicitly directs users to the sibling gke-app-onboarding skill instead.