← LeaderboardThe Skill Leaderboard

gke-kueue-troubleshooting

Diagnose a GKE batch or AI workload (Job, JobSet, LeaderWorkerSet, RayJob, Kubeflow training jobs, or plain Pods) that is pending, suspended, stuck, evicted or misconfigured when Kueue manages it - from a live GKE cluster, a support bundle or exported controller/audit logs. Load it as soon as the investigation shows a Kueue footprint, even if the user never said "Kueue" - anything under kueue.x-k8s.io (labels, Pod scheduling gates, Workload/LocalQueue/ClusterQueue objects and their conditions), kueue-controller-manager events or logs, or a Job with spec.suspend=true together with a Workload ob

Published by google ↗google/skills

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

name: gke-kueue-troubleshootingdescription: Diagnose a GKE batch or AI workload (Job, JobSet, LeaderWorkerSet, RayJob, Kubeflow training jobs, or plain Pods) that is pending, suspended, stuck, evicted or misconfigured when Kueue manages it - from a live GKE cluster, a support bundle or exported controller/audit logs. Load it as soon as the investigation shows a Kueue footprint, even if the user never said "Kueue" - anything under kueue.x-k8s.io (labels, Pod scheduling gates, Workload/LocalQueue/ClusterQueue objects and their conditions), kueue-controller-manager events or logs, or a Job with spec.suspend=true together with a Workload object in its namespace. Scripts compact cluster state and logs into verifiable facts and validate the report, while the agent classifies the cause as Kueue configuration, Kueue defect, or outside Kueue. Don't use for workloads with no Kueue footprint (use gke-node-notready, gke-workload-scaling-troubleshooting, gke-ai-troubleshooting-tpu-reservation-stockout, or gke-ai-troubleshooting-jobset-interruption).metadata:   version: "1.1.0"  category: Containers

GKE Kueue troubleshooting

Read-only rule: Run read-only diagnostic commands only. Never drain, delete, or re-create nodes, or run kubectl patch/apply/edit/delete or any other command that changes the cluster. When exporting logs for trace_logs.py, run only read-only log queries scoped to the target cluster, kueue-system, Kubernetes Event logs, Admin Activity audit logs, and the incident time window as described in Log sources and View GKE logs. Give the user any fix (remediation YAML, commands, or linked documentation steps) to apply themselves in the final ### Diagnostic Summary.

Usage Attribution

MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read and follow the instructions in Usage Attribution.

Prerequisites

  • Tools and cluster access: Python 3.10+ with PyYAML (all diagnostic scripts live in scripts/ next to this file and require no cloud SDKs). In live mode, follow "Before you begin" in Deploy a batch system using Kueue to initialize gcloud and select the target project, and follow Install kubectl and configure cluster access to configure cluster credentials (offline_data and logs_archive read local files or stdin only).
  • Cloud Billing and APIs: For live cluster inspection or Cloud Logging exports, ensure the project has an active Cloud Billing account and the Google Kubernetes Engine API (container.googleapis.com) and Cloud Logging API (logging.googleapis.com) are enabled.
  • Required IAM roles: Per Kubernetes Engine IAM roles and permissions, live cluster resource reads require Kubernetes Engine Viewer (roles/container.viewer), and reading controller pod logs via kubectl logs (container.pods.getLogs) or applying Kueue/cluster fixes requires Kubernetes Engine Developer (roles/container.developer) or Kubernetes Engine Admin (roles/container.admin). Exporting Cloud Logging and Admin Activity audit logs requires Logs Viewer (roles/logging.viewer).
  • Scope and related skills: Upstream Kueue ships experimental interactive skills (kueue-lineage and kueue-who-preempted, documented in Troubleshooting Agent Skills) for quick object-tree and preemption queries on a live cluster. This skill adds offline support-bundle and Cloud Logging compaction, GKE infrastructure boundary classification (Outside Kueue), and report validation.

1. Three-step protocol

The scripts detect the Kueue release and API version (active_api_schema_reference._detected_kueue_version, e.g. v0.10.x/v1beta1 vs v0.19.x/v1beta2) from controller deployments, CRDs, object apiVersions or startup/audit logs, and load the CRD and KueueConfiguration schema for that version. If detection_sources is only default:v1beta2, the version was assumed — pass --kueue-version <tag> or confirm the served CRD version before writing remediation YAML.

Where the schema comes from (schema_source / config_schema_source): in live the cluster's served CRDs are used (live_cluster_crds, no downloads); in offline_data bundled CRD manifests are used if present (bundled_crds); otherwise release CRDs are read from GKE_KUEUE_TROUBLESHOOTING_KUEUE_REPO or fetched once from GitHub (github_release:<tag>) and cached under $XDG_CACHE_HOME/gke-kueue-troubleshooting (override with GKE_KUEUE_TROUBLESHOOTING_CACHE_DIR; --no-remote-schema or GKE_KUEUE_TROUBLESHOOTING_NO_REMOTE_SCHEMA=1 disables downloads). KueueConfiguration is not a CRD; its field tree is parsed from the KueueConfiguration v1beta2 reference or KueueConfiguration v1beta1 reference matching config_api_version (falling back to builtin_defaults offline).

If schema_source is unavailable (_schema_notice present): CRD schemas are never approximated; only KueueConfiguration may fall back to builtin_defaults. Note in one sentence at the top of the report that the release-exact schema was unavailable, name the remedy from _schema_notice (cluster access, bundled CRDs, GitHub access, or --kueue-version), and continue the diagnosis using the fields, conditions, and logs in the snapshot (see LocalQueue and Topology Aware Scheduling). Remediation YAML still uses the detected apiVersion.

Choose the data mode (--mode or GKE_KUEUE_TROUBLESHOOTING_MODE, default live):

  • live: when kubectl reaches the cluster.
  • offline_data: directory of YAML/JSON manifests (optionally with logs/); pass --data-path <dir>.
  • logs_archive: exported logs only (no cluster or manifests); pass --data-path <file-or-dir> or pipe a read-only export to --data-path - (scoped per Log sources; never cat or page raw log exports into context).

Pass the same --ledger-path <ledger> to every script call; Step 3 verifies quotes and schema against it. Do not read or parse the ledger.

Output contract. snapshot.py and trace_logs.py print a digest ordered most-important-first (detected_kueue_version, derived facts, bulk_object_groups with counts/names/representative, individual objects/log excerpts, schema reference last) capped at --max-output-bytes (48 KB default for snapshot, 24 KB for logs). The complete output is written to full_report_file (next to the ledger). If command output is truncated to a saved file, read that saved file once instead of re-running --action get or grep. If the digest ends with truncated:, all tagged and coverage facts are already above it; to inspect one dropped object, use --action get --kind <kind> --name <name> (at most 4 get calls, never the same kind/name twice) or read a ≤ 200-line slice of full_report_file / raw bundle files (at most twice). Never read the skill's own Python scripts.

  1. Step 1 — one-shot snapshot or log timeline (schema attached)

    • live / offline_data: run scripts/snapshot.py with --mode <mode> --action snapshot --ledger-path <ledger> (add --namespace <ns> if known; --kueue-namespace if not kueue-system; --ignore-namespaces ns1,ns2 for extra add-on namespaces). If no active Workload/Job is present or controller/audit history is needed, also run scripts/trace_logs.py --mode <mode> --ledger-path <ledger> once (at most 2 log commands total).
    • logs_archive: run scripts/trace_logs.py --mode logs_archive --ledger-path <ledger> with --data-path <file-or-dir> or --data-path - first (add --workload <name> to focus). Never call kubectl in this mode.
    • Schema lookup (optional): scripts/snapshot.py --action explain with --kind <Kind[.field.path]> --ledger-path <ledger>.
    • If a script fails: read the error once, follow next_step, and continue with available files without retrying in a loop.
  2. Step 2 — reason over the facts (the scripts do not reason for you)

    • Script outputs contain facts, not conclusions (active_api_schema_reference, derived_graph_diagnostics, derived_log_diagnostics with mechanism_hint, key_errors_and_events, discovered_entities, audit_resource_actions). Nothing is pre-ranked as a root cause.
    • Check the whole admission and execution pipeline (stages 1–6) — do not stop at the first symptom if another independent blocker exists. For every Workload in the request or workload_timeline / discovered_entities, state its terminal state (Pending, Admitted, Evicted: <reason>, Requeued <n>/<limit>, Deactivated, Finished, Deleted by <owner cascade | TTL | retention policy>) and explain each terminal event. A Workload deleted via its Job's ownerReferences (user delete, ttlSecondsAfterFinished, or parent JobSet) is the Job's own lifecycle, not a Kueue retention defect — report it as why the object is gone, not as an RC<n>. Consult the Documentation catalog for the full per-stage upstream Kueue and GKE remediation links and section titles:
      1. Object chain: Job/JobSet/LWS → Workload → spec.queueName → LocalQueue (exists, Active) → ClusterQueue (exists, Active) → Cohort (see Troubleshooting Jobs and Troubleshooting Queues).
      2. Quota and cohort sharing: requests per pod set × count vs nominalQuota, borrowing (lendingLimit/borrowingLimit, cohort quota layout, reclaimWithinCohort / fairSharing), queueingStrategy ordering, and GPU/TPU reservation alignment (see ClusterQueue, Cohort, and Preemption).
      3. Flavor assignment per pod set: ResourceFlavor nodeLabels/nodeTaints/tolerations, topologyName vs the pod set's topology request and across resourceGroups, and resources.transformations. Per "Implicit TAS and mixed-flavor ClusterQueues" in Topology Aware Scheduling, in a ClusterQueue combining TAS and non-TAS flavors, implicit TAS is disabled (unannotated pod sets can use only non-TAS flavors and TAS-annotated pod sets can use only TAS flavors), and a single pod set cannot mix TAS and non-TAS flavors across its requested resources.
      4. Admission checks: AdmissionCheck Active and attached to the ClusterQueue, DWS ProvisioningRequest / retry configuration, reservation/DWS fallback, and external controllers (see ProvisioningRequest AdmissionCheck and Run a large-scale workload with flex-start with queued provisioning).
      5. After admission: scheduling gates (kueue.x-k8s.io/topology or external gates; see Troubleshooting Pods), waitForPodsReady (PodsReadyTimeout, RequeuingLimitExceeded), preemption, container/Job failure policies (backoffLimit, backoffLimitPerIndex, maxFailedIndexes, podFailurePolicy), Cluster Autoscaler noScaleUp / stockouts, TPU slice provisioning, and GKE node NotReady (NetworkPluginNotReady / CNI not initialized — follow "Troubleshoot network connectivity" in Troubleshoot nodes with the NotReady status in GKE or hand off to gke-node-notready).
      6. Controller health and configuration: KueueConfiguration (kueue-manager-config ConfigMap) fields (integrations.frameworks, manageJobsWithoutQueueName, featureGates, resources.transformations, objectRetentionPolicies), client-side rate limiting vs kube-apiserver overload (HTTP 429/503, webhook timeouts), controller restarts, and audit-logged create/delete/patch of any object in the chain.
    • Synthesize one unified Root Cause (RC<n>) per underlying defect — do not fragment symptoms:
      • If multiple workloads fail because of the same underlying configuration defect on a ClusterQueue or ResourceFlavor, report that single underlying defect as one RC1 rather than splitting each workload's symptom into separate root causes with conflicting fixes.
      • If an earlier event in the same layer is the upstream precursor of the current blocker, explain the causal chain together rather than inventing redundant competing root causes for the same failure chain.
      • Conversely, when two genuinely independent defects both affect the workload (e.g. a Kueue configuration defect and an independent external control-plane/capacity failure, or two independent Kueue configuration defects across different stages such as flavor/topology wiring and AdmissionCheck attachment), report each as its own RC<n> rather than dropping one or misclassifying a missing Kueue configuration field as a Kueue defect.
    • Defaults and uninvolved settings are not root causes — configured mechanisms with terminal events are: a field left at its Kueue default (such as waitForPodsReady omitted, waitForPodsReady.requeuingStrategy.backoffLimitCount unset/null, or default leaderElection timings) is not a root cause (RC<n>). When a workload repeatedly triggers a Kueue mechanism (such as PodsReadyTimeout requeuing) because its own pods or Job failure policy never reach completion or terminal failure, the single root cause (RC1) is the workload's failure/completion configuration (Outside Kueue: workload spec), while Kueue's default retry/requeue setting is background context. Do not propose waitForPodsReady, leaderElection, or integrations.frameworks ConfigMap changes as "hardening" on quota, cohort, preemption, TAS, workload-spec, or infrastructure incidents. Conversely, a mechanism that is explicitly configured (waitForPodsReady.timeout, requeuingStrategy.backoffLimitCount, objectRetentionPolicies, preemption policies) and has a terminal event in the evidence (Evicted with reason PodsReadyTimeout, Deactivated / RequeuingLimitExceeded, Preempted, or deletion of deactivated Workload objects by objectRetentionPolicies.workloads, which per "Notes" in Setup object retention policy may cascade to objects not created by Kueue such as a JobSet) is part of the causal chain and must be reported — as its own RC<n> when the configured value causes the failure, or as an explicit stage of the chain otherwise. "Working as designed" never justifies leaving such an event unexplained.
    • Every root cause must be backed by a verbatim quote (from status.conditions, spec, Events, or controller/audit logs) reproduced in #### Observations. A quota or capacity claim additionally needs a rejection fact for that same resource — insufficient quota / insufficient unused quota in a Workload condition or controller log, NotTriggerScaleUp, or a Pending / FailedScheduling reason naming the resource. A nominalQuota smaller or larger than node allocatable, nominalQuota: 0 on a queue that borrows, blockAdmission: false, or a short ttlSecondsAfterFinished is an observation, not a root cause, unless such a rejection fact ties it to the failure.
    • Scope boundary — classify the causal layer (#### Scope) before writing Root Cause or Remediation: Declare one of Kueue configuration | Kueue defect | Outside Kueue: <capacity | node | control plane | workload spec> | Mixed: <layers> | Undetermined: <what is missing>. Attribute each blocking fact by which component produced the signal and which object owns the field, not by how the symptom looks:
      1. Locate the blocking stage(s) in the pipeline (stages 1–6):
        • Stages 1–4 (queue chain, quota/cohort layout, flavor/TAS assignment, admission check wiring) are decided by Kueue from Kueue objects (Kueue configuration, or Kueue defect if Kueue contradicts its documented behavior for the detected version).
        • Stage 5 (after Admitted=True) is decided by kube-scheduler, kubelet, node provisioner / autoscaler, admission-check controllers, and the workload's own pods. A blocker there is Outside Kueue unless a Kueue configuration caused it (e.g. ResourceFlavor.spec.nodeLabels pointing at non-existent labels, a stranded kueue.x-k8s.io/topology scheduling gate, or cohort nominalQuota mismatched to cluster capacity).
        • Cohort nominalQuota, borrowing, and Fair Sharing: Per ClusterQueue and Cohort, queues in a cohort can borrow unused nominalQuota subject to lendingLimit (which caps how much unused nominalQuota a member lends) and borrowingLimit. Compare cohort nominalQuota declarations, audit_resource_actions, and Fair Sharing / preemption settings against observed admission, preemption, or scheduling events. A queue with nominalQuota: 0 is a valid borrow-only design, not a defect on its own, and #### Remediation must modify only the resources and fields implicated by the rejection facts. Per "ClusterQueue share value" in Preemption, Kueue summarizes borrowed usage in .status.fairSharing.weightedShare weighted by .spec.fairSharing.weight (higher weight lowers share value for the same borrowed usage; Kueue admits lowest share first and preempts highest share first; when cohort members configure different .spec.fairSharing.weight values, report each member's weight and how Fair Sharing weights their borrowed usage in #### Observations).
        • Stage 6 (controller health) splits by signal provenance: facts tagged source side: kueue-client reflect client-side rate limiting controlled by clientConnection.qps and clientConnection.burst in the "ClientConnection" section of the KueueConfiguration reference; facts tagged source side: apiserver (HTTP 429/503, API Priority and Fairness, webhook/etcd timeouts) belong to Outside Kueue: control plane (see Monitor GKE control plane metrics).
      2. Independent layers → Mixed. If both a Kueue-side blocker and an external-layer blocker independently affect the workload(s), set #### Scope to Mixed: <layers>, give each layer its own RC<n>, and include the literal phrase Outside Kueue in the external RC<n> header line (e.g. - RC2 (Outside Kueue — control plane): ...). Never write Outside Kueue in a Kueue-configuration RC header or its remediation.
      3. Never patch across layers. An Outside Kueue RC is remediated on the component that owns it (node pool, image, control plane, cloud capacity, the Job's own spec). A Kueue configuration RC is remediated on the Kueue object that owns the field.
    • When evidence is incomplete, use #### Open Questions (OQ<n>): BUNDLE_COVERAGE and LOG_COVERAGE state which object kinds and log sources are present and absent. If the runtime status objects and logs needed to confirm why a stage failed are absent from both (e.g. a static manifest-only bundle or a logs-only archive lacking manifests), do not assert an unverified guess as a definite RC<n>: record OQ1 under #### Open Questions stating what is missing, the candidate explanations, the discriminating observation, the read-only Collect: command, and the conditional remediation for each candidate.
    • Commit when the evidence decides — do not hedge a decided finding. When runtime evidence (conditions, events, or logs) settles the cause, write a definite RC<n> under #### Root Cause with one concrete fix under #### Remediation, without listing "or possibly ..." alternatives next to it. Reserve OQ<n> for questions the available evidence genuinely cannot answer and that name one read-only command to answer them. If the first-listed candidate is the only one with a rejection or terminal fact in the evidence, it is the RC<n>, not an OQ<n>. Explicitly configured mechanisms with terminal events belong in the causal chain, not under #### Ruled Out.
  3. Step 3 — evidence and schema validation

    • Number confirmed root causes RC1, RC2, … and give a matching remediation entry for each RC<n> under #### Remediation (conditional fixes for unresolved candidates go with their OQ<n>):
      • For a Kueue configuration RC: link the official Kueue or GKE documentation section from Documentation catalog alongside the exact kubectl command or YAML manifest in the detected API version, preserving all active_api_schema_reference invariants and covering every resource requested by the workload's PodSets. KueueConfiguration is not an API resource: tell the user to follow "Install a custom-configured released version" and "Change the feature gates configuration" in Installation (or the GKE sections in Documentation catalog) to edit data.controller_manager_config.yaml (kind: Configuration) in the kueue-manager-config ConfigMap and restart the kueue-controller-manager Deployment, showing the complete controller_manager_config.yaml document rather than a partial patch.
      • For a Kueue defect RC: do not invent unsupported workarounds that break workload semantics. Tell the user to collect snapshot and controller log evidence, check or file an upstream issue at https://github.com/kubernetes-sigs/kueue/issues or contact Cloud Customer Care per Getting support for GKE, and state the confirmation signal (Confirm by checking that ...).
      • For an Outside Kueue RC: never propose a ClusterQueue, ResourceFlavor, or KueueConfiguration patch to mask an external infrastructure failure. State the external-layer remediation (linking the relevant GKE guide from Documentation catalog or handing off to gke-node-notready, gke-workload-scaling-troubleshooting, gke-ai-troubleshooting-tpu-reservation-stockout, or gke-ai-troubleshooting-jobset-interruption), state the confirmation signal (Confirm by checking that <object condition> transitions to <expected state>), and for persistent control-plane or hardware issues advise contacting Cloud Customer Care per Getting support for GKE.
    • In #### Observations, wrap a multi-word phrase in backticks only when it is a verbatim substring of tool output (synthesised summaries go in plain text).
    • Save the draft and run scripts/validate_report.py with --draft-file <draft.md> --ledger-path <ledger>. It checks required headers, fabricated_quotes, remediation_gaps, config_remediation_errors, unknown_kubectl_resources, and schema_warnings. If it reports an error, fix it and re-run once. unsupported_root_causes is advisory: tie any listed RC<n> to a rejection/terminal signal in #### Observations or move it to #### Open Questions; do not re-run the validator for it.
    • Finish with:
### Diagnostic Summary
#### Scope
<Kueue configuration | Kueue defect | Outside Kueue: <layer> | Mixed>
#### Observations
- Verbatim evidence from `status.conditions`, `spec`, `Events`, or
  controller/audit logs, with timestamps for historical facts.
#### Root Cause
- RC1: <precise root cause and causal mechanism>, tied to the evidence above.
- RC2 (Outside Kueue — <layer>): <if Mixed / external layer, state the external
  root cause>.
#### Open Questions
- none (or `OQ1` with missing data, candidates, discriminator, read-only
  `Collect:` command, and conditional remediation when data is absent).
#### Ruled Out
- Brief list of secondary observations, resolved/precursor events, or
  non-blocking settings ruled out as root causes.
#### Remediation
- RC1: link the official documentation section describing the fix (plus the
  exact `kubectl` command or YAML manifest with that link for Kueue RCs), or
  external-layer remediation + confirmation signal (for external RCs).
- RC2: ...