gke-ai-troubleshooting-tpu-dynamic-slices-monitoring

Files (4)

• SKILL.md
• attribution.md · references/
• failure_signatures.md · references/
• validate_queries.sh · scripts/

SKILL.md (first lines)

# GKE TPU Dynamic Slices Monitoring & Management

Monitors the status of TPU Slice custom resources, troubleshoots provisioning
failures, validates workload manifests on dynamic slices, and performs cleanups.

## Attribution

**MUST READ**: Before executing any `gcloud`, REST API, or MCP command, you
MUST read [Usage Attribution](references/attribution.md). For commands executed
directly in the terminal (never in generated scripts, Terraform, or user code),
prefix `gcloud` with
`CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:gke-ai-troubleshooting-tpu-dynamic-slices-monitoring)"`
and append `agent-skills/1.0 (skill:gke-ai-troubleshooting-tpu-dynamic-slices-monitoring)` to `User-Agent` headers.

## Prerequisites

-   Cloud Logging enabled for the project.
-   `kubectl` and `gcloud` CLIs configured to access the GKE cluster.

## Diagnostic Workflow

### Step 0: Context Acquisition & Time Window Definition

Gather project, cluster, and slice context using cluster tools or the following
parameters:

-   **Project ID**: `{project_id}` (e.g., `my-gcp-project`)
-   **Cluster Name**: `{cluster_name}` (e.g., `tpu-cluster`)
-   **Region/Zone**: `{location}` (e.g., `us-central1-a`)
-   **Slice Name**: `{slice_name}` (e.g., `test-slice`)
-   **Issue Time**: `{timestamp}` (Optional; default to the last 30 minutes
    window `[T - 30m]` to `[T + 30m]`)

--------------------------------------------------------------------------------

### Step 1: Describe the Slice Custom Resource [Low Risk]

When asked to inspect, troubleshoot, or check a slice status, immediately execute `kubectl describe slice {slice_name}` using available cluster tools to perform the inspection. Parse the resulting `Status.Conditions` output against the condition table below to diagnose the exact state and provide concrete recommendations.

-   **Command**:

    ```bash
    kubectl describe slice {slice_name}
    ```

#### State & Reason Analysis

Analyze the `Status.Conditions` (especially `Type: Ready` and its `Reason` and
`Status`):

| Lifecycle State / Reason | Meaning | Recommended Action |
| :--- | :--- | :--- |
| **`SliceNotCreated`** | GKE Slice Controller is initializing the slice and performing resource checks. | Wait a few minutes and re-check slice status. |
| **`SliceCreationFailed`** | Prerequisites validation failed (e.g., selected nodes don't exist, nodes are already used by another slice, or the topology doesn't match the number of partitions). | Verify selected nodes exist, are unallocated, and topology matches partition count. |
| **`ACTIVATING`** | GKE is actively forming and provisioning the TPU slice. | Monitor node provisioning. |
| **`ACTIVE`** | The TPU slice is successfully formed and ready to host workloads. | Proceed to deploy or check workloads. |
| **`ACTIVE_DEGRADED`** | The slice is usable, but one or more sub-blocks are degraded. | Monitor workload logs for interconnect or device errors. Check faulty node VMs. |
| **`FAILED`** | GKE failed to form the TPU slice (e.g., selected nodes are not part of the same reservation block). | Ensure all selected nodes belong to the same reservation block. |
| **`DEACTIVATING`** | The slice is dismantling (triggered by user deletion or a critical systemic failure). | Wait for dismantling to finish, or patch finalizers if stuck. |
| **`INCOMPLETE`** | The terminal phase before the Slice CR is deleted from the cluster. | No action required; the resource will be removed shortly. |

#### Provisioning Failure Troubleshooting Checklist

When investigating slice creation or provisioning failures (`SliceCreationFailed` or `FAILED`), perform the following verification steps:

1. **Node Existence & Allocation Check**: Verify that the selected TPU nodes exist in the cluster and are not already allocated to another slice (`kubectl get nodes -l cloud.google.com/gke-tpu-slice`, `kubectl get slice -A`).
2. **Topology Alignment**: Confirm that the partition count matches the requested topology dimensions (e.g. topology `2x2` requires 4 nodes).
3. **Reservation Block Alignment Check**: Confirm that all selected TPU nodes belong to the same reservation and reservation block.

--------------------------------------------------------------------------------

### Step 2: Verify Workload Specification [Low Risk]

Ensure workload manifests are configured correctly to target the dynamic slice.

#### 1. Single-Slice Workload Requirements

Check that the Pod template contains the following annotations and selectors:

-   **Annotations**:
    -   `cloud.google.com/gke-tpu-slice-topology: "{topology}"` (e.g.,
        `"4x4x4"`)
-   **NodeSelector**:
    -   `cloud.google.com/gke-tpu-topology: "{topology}"` (e.g., `"4x4x4"`)
    -   `cloud.google.com/gke-tpu-accelerator: "{accelerator_type}"` (e.g.,
        `"tpu7x"`)
    -   `cloud.google.com/gke-tpu-slice: "{slice_name}"` (e.g., `"test-slice"`)

#### 2. Multi-Slice (JobSet) Workload Requirements

If deploying a multi-slice JobSet, verify:

-   **JobSet Annotation**:
    -   `alpha.jobset.sigs.k8s.io/exclusive-topology:
        cloud.google.com/gke-tpu-slice`
-   **Pod Template Annotations**:
    -   `cloud.google.com/gke-tpu-slice-topology: "{topology}"`
-   **Pod Template NodeSelector**:
    -   `cloud.google.com/gke-tpu-topology: "{topology}"`
    -   `cloud.google.com/gke-tpu-accelerator: "{accelerator_type}"`
    -   *Note: Do NOT manually specify `cloud.google.com/gke-tpu-slice` in the
        nodeSelector; JobSet handles slice assignment automatically.*

--------------------------------------------------------------------------------

## Resolution & Management Workflow

### Resolution 1: Force Delete a Stuck Slice [High Risk]

If a slice is stuck in `DEACTIVATING` or deletion hangs indefinitely due to stuck finalizers:

1. **Identify Cause**: Explain that finalizers on the slice resource (`metadata.finalizers`) are preventing Kubernetes from completing resource deletion.
2. **Propose Resolution**: Propose removing finalizers from the metadata path (`/metadata/finalizers`) using a JSON patch operation:

    ```bash
    kubectl patch slice {slice_name} --type json -p='[{"op": "remove", "path": "/metadata/finalizers"}]'
    ```

3. **Provide Warning**: Explicitly warn the user that removing finalizers bypasses standard controller dismantling and may leave underlying VM, network, or accelerator resources uncleaned or orphaned.

From google/skills, Apache-2.0.

More from google/skills