gke-compute-classes

Name: gke-compute-classes
Author: google

1.3kinstalls1.1kforks

How to install gke-compute-classes?

google/skills/gke-compute-classes

$npx -y skills add google/skills --skill gke-compute-classes

Installs into the current project.

Summary

Configures, optimizes, and troubleshoots GKE ComputeClasses. Use when configuring Spot VMs with on-demand fallback, targeting specific accelerators (GPUs/TPUs) or machine families, restricting ComputeClass access, or debugging pending pods related to node pool auto-creation. Do n

Files · 1

View on GitHub

SKILL.md

1	<!-- disableFinding(LINE_OVER_80) -->
2
3	# GKE ComputeClasses
4
5	Guidance on configuring, optimizing, and troubleshooting GKE ComputeClasses.
6
7	## When to Use
8
9	- Cost optimization: Spot VMs with on-demand fallback.
10	- GPU/TPU workloads: Target specific accelerators (e.g., L4, H100, v5p).
11	- Performance tuning: Select specific machine families (c3, c4, n4).
12	- Zone targeting: Colocate workloads with zonal resources.
13
14	--------------------------------------------------------------------------------
15
16	## Engagement Rules: Generalized First, Refine Later
17
18	ComputeClasses depend on zone availability, CUDs, and workload constraints. **Do
19	not block the user's initial request.** If asked for YAML/recommendations:
20
21	1. Provide Generalized Answer Immediately: Fulfill request using best
22	practices and placeholders (`<YOUR-ZONE-HERE>`).
23	* CRITICAL CUD RULE: You MUST state that the provided machine families
24	(e.g., N4, C4) are generic best-practice examples. You MUST explicitly
25	state that the final choice of machine family should be aligned with the
26	user's existing Committed Use Discounts (CUDs) or Reservations.
27	* YAML REQUIREMENT: Any generated YAML template MUST include a comment
28	near the `machineFamily` field: `# IMPORTANT: Align machineFamily with
29	your existing CUDs/Reservations`.
30	* MUST label initial YAML as `EXAMPLE TEMPLATE - DO NOT DEPLOY`.
31	* STRICT SCHEMA RULE: NEVER hallucinate fields. Do NOT use
32	`spec.description`, `gvnic`, `transparentHugepageEnabled`, or
33	`shutdownGracePeriodSeconds`. Use `bootDiskSize` (NOT `bootDiskSizeGb`).
34	* YAML FORMATTING RULE: NEVER quote integer or boolean values (e.g.,
35	use `bootDiskSize: 50`, not `bootDiskSize: "50"`). `imageType` MUST be
36	lowercase.
37	* CRITICAL AI/ML RULE: DO NOT recommend Spot instances as the primary
38	priority for AI/ML Inference, even if the workload is stateless.
39	Accelerator node startup latency is severe. The correct priority is:
40	`Reservations -> On-Demand -> DWS FlexStart -> Spot`.
41	* CRITICAL PROVISIONING RULE: Do NOT confuse node pool auto-creation
42	with cluster-level Node Auto Provisioning. Starting with GKE
43	`1.33.3-gke.1136000`, `nodePoolAutoCreation.enabled: true` in the
44	ComputeClass achieves automatic node pools scoped directly to the
45	ComputeClass. **It does NOT require turning on Node Auto Provisioning at
46	the cluster level.**
47	* CRITICAL TAINT RULE: The ONLY redundant taint is re-adding
48	`cloud.google.com/compute-class` on auto-created pools — node pool
49	auto-creation already applies AND auto-tolerates that key, so
50	duplicating it breaks scheduling → REMOVE it (don't add a toleration).
51	This is NOT "never add taints": an intentional dedication/isolation
52	taint (e.g. `dedicated=ml:NoSchedule`) in `nodePoolConfig.taints` is
53	valid — it keeps other workloads off, and the intended workloads need a
54	matching toleration (normal K8s contract). Judge intent before deleting;
55	only the compute-class key is redundant. **Manual pools STILL require
56	`cloud.google.com/compute-class=<NAME>` as label AND taint to bind to
57	the ComputeClass — never remove that. Schema limit:** a
58	`nodePoolConfig.taints` key may NOT contain the reserved `kubernetes.io`
59	substring (GKE Warden rejects it) — so the Cluster-Autoscaler-ignored
60	prefixes
61	(`startup-taint.`/`status-taint.cluster-autoscaler.kubernetes.io/`)
62	cannot be set via a ComputeClass; those are node-pool-level taints.
63	* CRITICAL GPU-TAINT RULE: GKE auto-taints GPU nodes
64	`nvidia.com/gpu:NoSchedule` — this is separate from the
65	`cloud.google.com/compute-class` auto-toleration and is NOT covered by
66	it. A GPU Pod stuck `Pending` / `noScaleUp` is almost always missing the
67	toleration. Add to the PodSpec: `tolerations: [{key: nvidia.com/gpu,
68	operator: Exists}]`.
69	* CRITICAL SPOT-TAINT RULE: GKE auto-taints Spot nodes with
70	`cloud.google.com/gke-spot=true:NoSchedule`. Pods targeting a Spot
71	priority tier must tolerate this taint, or they will stay `Pending` /
72	`noScaleUp` with a scheduling block. Tell the user to add the matching
73	toleration to their PodSpec: `tolerations: [{key:
74	cloud.google.com/gke-spot, operator: Equal, value: "true", effe

People who liked this also viewed

6 picks

Type

gke-compute-classes

1.3kinstalls1.1kforks

View on GitHub

How to install gke-compute-classes?

google/skills/gke-compute-classes

$npx -y skills add google/skills --skill gke-compute-classes

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	<!-- disableFinding(LINE_OVER_80) -->
2
3	# GKE ComputeClasses
4
5	Guidance on configuring, optimizing, and troubleshooting GKE ComputeClasses.
6
7	## When to Use
8
9	- Cost optimization: Spot VMs with on-demand fallback.
10	- GPU/TPU workloads: Target specific accelerators (e.g., L4, H100, v5p).
11	- Performance tuning: Select specific machine families (c3, c4, n4).
12	- Zone targeting: Colocate workloads with zonal resources.
13
14	--------------------------------------------------------------------------------
15
16	## Engagement Rules: Generalized First, Refine Later
17
18	ComputeClasses depend on zone availability, CUDs, and workload constraints. **Do
19	not block the user's initial request.** If asked for YAML/recommendations:
20
21	1. Provide Generalized Answer Immediately: Fulfill request using best
22	practices and placeholders (`<YOUR-ZONE-HERE>`).
23	* CRITICAL CUD RULE: You MUST state that the provided machine families
24	(e.g., N4, C4) are generic best-practice examples. You MUST explicitly
25	state that the final choice of machine family should be aligned with the
26	user's existing Committed Use Discounts (CUDs) or Reservations.
27	* YAML REQUIREMENT: Any generated YAML template MUST include a comment
28	near the `machineFamily` field: `# IMPORTANT: Align machineFamily with
29	your existing CUDs/Reservations`.
30	* MUST label initial YAML as `EXAMPLE TEMPLATE - DO NOT DEPLOY`.
31	* STRICT SCHEMA RULE: NEVER hallucinate fields. Do NOT use
32	`spec.description`, `gvnic`, `transparentHugepageEnabled`, or
33	`shutdownGracePeriodSeconds`. Use `bootDiskSize` (NOT `bootDiskSizeGb`).
34	* YAML FORMATTING RULE: NEVER quote integer or boolean values (e.g.,
35	use `bootDiskSize: 50`, not `bootDiskSize: "50"`). `imageType` MUST be
36	lowercase.
37	* CRITICAL AI/ML RULE: DO NOT recommend Spot instances as the primary
38	priority for AI/ML Inference, even if the workload is stateless.
39	Accelerator node startup latency is severe. The correct priority is:
40	`Reservations -> On-Demand -> DWS FlexStart -> Spot`.
41	* CRITICAL PROVISIONING RULE: Do NOT confuse node pool auto-creation
42	with cluster-level Node Auto Provisioning. Starting with GKE
43	`1.33.3-gke.1136000`, `nodePoolAutoCreation.enabled: true` in the
44	ComputeClass achieves automatic node pools scoped directly to the
45	ComputeClass. **It does NOT require turning on Node Auto Provisioning at
46	the cluster level.**
47	* CRITICAL TAINT RULE: The ONLY redundant taint is re-adding
48	`cloud.google.com/compute-class` on auto-created pools — node pool
49	auto-creation already applies AND auto-tolerates that key, so
50	duplicating it breaks scheduling → REMOVE it (don't add a toleration).
51	This is NOT "never add taints": an intentional dedication/isolation
52	taint (e.g. `dedicated=ml:NoSchedule`) in `nodePoolConfig.taints` is
53	valid — it keeps other workloads off, and the intended workloads need a
54	matching toleration (normal K8s contract). Judge intent before deleting;
55	only the compute-class key is redundant. **Manual pools STILL require
56	`cloud.google.com/compute-class=<NAME>` as label AND taint to bind to
57	the ComputeClass — never remove that. Schema limit:** a
58	`nodePoolConfig.taints` key may NOT contain the reserved `kubernetes.io`
59	substring (GKE Warden rejects it) — so the Cluster-Autoscaler-ignored
60	prefixes
61	(`startup-taint.`/`status-taint.cluster-autoscaler.kubernetes.io/`)
62	cannot be set via a ComputeClass; those are node-pool-level taints.
63	* CRITICAL GPU-TAINT RULE: GKE auto-taints GPU nodes
64	`nvidia.com/gpu:NoSchedule` — this is separate from the
65	`cloud.google.com/compute-class` auto-toleration and is NOT covered by
66	it. A GPU Pod stuck `Pending` / `noScaleUp` is almost always missing the
67	toleration. Add to the PodSpec: `tolerations: [{key: nvidia.com/gpu,
68	operator: Exists}]`.
69	* CRITICAL SPOT-TAINT RULE: GKE auto-taints Spot nodes with
70	`cloud.google.com/gke-spot=true:NoSchedule`. Pods targeting a Spot
71	priority tier must tolerate this taint, or they will stay `Pending` /
72	`noScaleUp` with a scheduling block. Tell the user to add the matching
73	toleration to their PodSpec: `tolerations: [{key:
74	cloud.google.com/gke-spot, operator: Equal, value: "true", effe

People who liked this also viewed

6 picks

Type