gke-ai-troubleshooting-tpu-reservation-stockout

Troubleshoot GKE TPU reservation stockouts caused by leaked node pool capacity

Use this skill to diagnose and resolve GKE TPU node pool creation or scale-up failures that fail with GCE_STOCKOUT, ZONE_RESOURCE_POOL_EXHAUSTED, or ZONE_RESOURCE_POOL_EXHAUSTED_WITH_DETAILS (or queue due to lack of capacity) even though they target a specific Compute Engine TPU reservation.

If node auto-repair is disabled on an existing TPU node pool, GKE doesn't repair its nodes when they become NotReady. While their VMs are still running, they keep consuming slots in the reservation (inUseCount is the number of running instances that consume the reservation). Per Ensure reservation availability, the reservation must have sufficient available TPU chips at the time of node pool creation, and per Multi-host TPU slice node pools, if GKE is unable to deploy one node in a slice, no nodes in the TPU slice node pool are deployed. When the free slots in the reservation (specificReservation.count - specificReservation.inUseCount; see Consumable instances in a reservation) are fewer than the number of nodes that the new TPU node pool needs, creation fails with GCE_STOCKOUT or ZONE_RESOURCE_POOL_EXHAUSTED (or queues due to lack of capacity).

Prerequisites

• Required IAM Roles:
  • Logs Viewer (roles/logging.viewer): Query Cloud Audit Logs in Cloud Logging.
  • Monitoring Viewer (roles/monitoring.viewer): Query GKE node readiness metrics in Cloud Monitoring.
  • Kubernetes Engine Viewer (roles/container.viewer): Inspect GKE node pools (management.autoRepair), cluster operations, and node conditions.
  • Compute Viewer (roles/compute.viewer): Inspect Compute Engine reservation utilization (specificReservation.inUseCount vs specificReservation.count), consuming instances, and VM labels (including in the reservation's owner project if using a shared reservation).
  • For remediation ([High Risk] steps): Kubernetes Engine Cluster Admin (roles/container.clusterAdmin)
• API Enablement & Billing:
  • Ensure the target project has an active Google Cloud billing account attached: gcloud billing projects describe {project_id}.
  • Enable required APIs: container.googleapis.com, compute.googleapis.com, logging.googleapis.com, and monitoring.googleapis.com.
• CLI & Authentication Setup:
  • Install Google Cloud SDK (gcloud) and kubectl: https://cloud.google.com/sdk/docs/install.
• Cluster mode: This skill applies to Standard clusters. Autopilot clusters always auto-repair nodes, and you can't turn this off (see Settings for Autopilot and Standard).
• Reference Documentation:
  • Auto-repair nodes (Sections: "Settings for Autopilot and Standard", "Repair criteria", "Node repair process", "Verify node auto-repair is enabled for a Standard node pool", "Node auto repair in TPU slice nodes", "Enable auto-repair for an existing Standard node pool", "Get information about recent automated repair events")
  • Troubleshoot GKE Standard node pools (Sections: "Issue: Node pool creation fails due to insufficient resources")
  • Troubleshoot TPUs in GKE (Sections: "How GKE handles capacity issues", "Insufficient TPU resources to satisfy the TPU request")
  • About TPUs in GKE (Sections: "Multi-host TPU slice node pools")
  • Consuming reserved zonal resources (Sections: "Consuming a specific shared reservation")
  • About reservations (Sections: "How reservations work")
  • View reservations or future reservation requests (Sections: "Limitations", "View reservations", "Consumable instances in a reservation", "View a list of instances that are consuming a reservation")
  • Organize resources using labels (Sections: "View labels")
  • Create and manage cluster and node pool labels (Sections: "Automatically applied labels")
  • Deploy TPU workloads in GKE Standard (Sections: "Ensure reservation availability", "Monitor health metrics for TPU nodes and node pools", "Configure auto repair for TPU slice nodes")
  • Add and manage node pools (Sections: "Delete a node pool")

Read-only rule: Run read-only diagnostic commands only. Never drain, delete, or re-create nodes, or run any other command that changes the cluster. Give the user any fix to apply themselves.

When you recommend a fix, link the doc section that describes it.

---

Diagnostic workflow

Step 0: Collect context and set the investigation window

Identify the following environment parameters from the user prompt or cluster context:

• {project_id}, {cluster_name}, {location}, {failing_nodepool_name}, {reservation_name}, {reservation_project_id} (the project that owns {reservation_name}, which defaults to {project_id} for a single-project reservation or is the owner project for a shared reservation; see Consuming a specific shared reservation), {zone} (the zone of the reservation), and {issue_time}.
• Time window: By default, query a 60-minute window [{issue_time} - 30m, {issue_time} + 30m] ({start_time} to {end_time}).

---

Step 1: Confirm the stockout error and reservation utilization [Low Risk]

1. Check Cloud Audit Logs and GKE operations for stockout or capacity errors:
   • Run the query in Issue: Node pool creation fails due to insufficient resources across [{start_time}, {end_time}] and look for the documented error messages, such as ZONE_RESOURCE_POOL_EXHAUSTED or ZONE_RESOURCE_POOL_EXHAUSTED_WITH_DETAILS.
   • Per Insufficient TPU resources to satisfy the TPU request, TPU errors can also contain GCE_STOCKOUT.
   • Per How GKE handles capacity issues, GKE sometimes queues a TPU node pool creation request instead of failing it immediately, surfacing a ... cannot be created ... due to lack of capacity message instead of GCE_STOCKOUT. List the node pool creation operation with gcloud container operations list (--filter="operationType=CREATE_NODE_POOL").
2. Check Compute Engine reservation utilization:
   • Per Limitations, query shared reservations in the owner project where they were created ({reservation_project_id}).
   • Follow the View reservations section to run gcloud compute reservations describe {reservation_name} --zone={zone} --project={reservation_project_id} or gcloud compute reservations list --project={reservation_project_id}, and compare specificReservation.inUseCount against specificReservation.count.

Interpret the results

• Don't suggest deleting any node pool until you have identified which existing VMs and node pools hold the reserved capacity (Step 2).
• Stockout confirmed: Per Ensure reservation availability and Multi-host TPU slice node pools, if GKE can't deploy one node in a multi-host slice, no nodes in the slice are deployed. If a stockout or lack-of-capacity message is present and the free slots (specificReservation.count - specificReservation.inUseCount; see Consumable instances in a reservation) are fewer than the number of nodes that {failing_nodepool_name} needs, proceed to Step 2.

---

Step 2: Find which VMs and node pools hold the reservation [Low Risk]

1. Check which VMs are consuming {reservation_name}:
   • Per How reservations work, a specifically targeted reservation can be used by VMs in the reservation's project ({reservation_project_id}) or, for a shared reservation, in any project that it's shared with.
   • Follow View a list of instances that are consuming a reservation to list the VMs that use {reservation_name}.
   • Follow View labels to view each VM's labels, and check the GKE labels in Automatically applied labels (goog-k8s-cluster-name, goog-k8s-cluster-location, and goog-k8s-node-pool-name) to find each VM's cluster and node pool.
2. Check node readiness in {cluster_name}:
   • Query the Cloud Monitoring metric kubernetes_io:node_status_condition (monitored_resource="k8s_node", cluster_name="{cluster_name}", condition="Ready") described in Monitor health metrics for TPU nodes and node pools. The status label value is True, False, or Unknown (capitalized); use status!="True" to find nodes whose Ready condition is False or Unknown.
   • List TPU nodes and their node pool labels with kubectl get nodes -l cloud.google.com/gke-tpu-accelerator -L cloud.google.com/gke-nodepool.
3. Exclude the failing node pool: Leave {failing_nodepool_name} out of the unhealthy list. It's the node pool that couldn't get capacity, not the one holding it.

Interpret the results

• VMs outside {cluster_name}: If some of the VMs consuming {reservation_name} aren't nodes in {cluster_name} (for example, their labels show another GKE cluster or they are Compute Engine VMs without GKE labels), tell the user which VMs or other clusters hold those reservation slots.
• Fewer listed VMs than specificReservation.inUseCount: Per View a list of instances that are consuming a reservation, the list only includes VMs in one project. If the list has fewer VMs than specificReservation.inUseCount, tell the user that VMs in other projects that share {reservation_name} are consuming the remaining slots.
• Unhealthy node pool in {cluster_name} ({leaking_nodepool_name}): If the VMs consuming {reservation_name} belong to an existing NotReady node pool in {cluster_name} ({leaking_nodepool_name}), proceed to Step 3.
• Healthy reservation (rule out a leak): If the reserved slots are held by healthy, in-use node pools, there is no leak. Explain that healthy node pools hold the reserved slots, so too few slots are free for {failing_nodepool_name}. Tell the user not to delete or restart those healthy node pools, and point the user to the consumption options in Insufficient TPU resources to satisfy the TPU request.

---

Step 3: Verify node auto-repair on the leaking node pool [Low Risk]

Follow Verify node auto-repair is enabled for a Standard node pool for {leaking_nodepool_name}:

• Auto-repair is disabled (management.autoRepair: false): GKE doesn't repair the NotReady nodes in {leaking_nodepool_name}. Their running VMs keep counting toward specificReservation.inUseCount in {reservation_name}, so {failing_nodepool_name} can't get reserved capacity. Proceed to Step 4.
• Auto-repair is enabled (management.autoRepair: true), but nodes are still NotReady:
  • Check whether the nodes have been NotReady for less than the repair time threshold (approximately 10 minutes), per Repair criteria.
  • Check the repair status and the failure reason in the operation history, per Configure auto repair for TPU slice nodes and Get information about recent automated repair events.

---

Step 4: Recover or release the reservation capacity [High Risk]

Guardrail: Don't modify GKE-managed instance groups or VMs directly through Compute Engine. Manage node pools through GKE. Never delete healthy node pools that run active jobs.

1. If the user wants to repair {leaking_nodepool_name} in place, enable node auto-repair (doesn't free slots for {failing_nodepool_name}):
   • Per Node repair process, GKE drains and re-creates an unhealthy node.
   • Per Node auto repair in TPU slice nodes and Configure auto repair for TPU slice nodes, GKE re-creates the entire multi-host TPU slice node pool when a node in it needs repair. The re-created VMs still consume {reservation_name}, so specificReservation.inUseCount doesn't drop: enabling auto-repair makes the nodes in {leaking_nodepool_name} healthy again, but {failing_nodepool_name} still can't get their slots.
   • Link the user to Enable auto-repair for an existing Standard node pool to turn on auto-repair for {leaking_nodepool_name}.
2. If {leaking_nodepool_name} is stale or the user wants to free its slots for {failing_nodepool_name}, delete it through GKE: Link the user to Delete a node pool to delete {leaking_nodepool_name}, which frees the reserved slots its VMs use.
3. After the user deletes {leaking_nodepool_name}, verify that enough reservation slots are free: Follow View reservations to run gcloud compute reservations describe {reservation_name} --zone={zone} --project={reservation_project_id} and confirm that specificReservation.inUseCount drops so the free slots (specificReservation.count - specificReservation.inUseCount) are at least the number of nodes that {failing_nodepool_name} needs before retrying {failing_nodepool_name}.