eks-workload-best-practice-assessment（eks 工作负载最佳实践评估）

EKS Workload Best Practice Assessment

Assess Kubernetes workloads on Amazon EKS against best practices from K8s official documentation and the EKS Best Practices Guide. Covers 8 dimensions: workload configuration, security, observability, networking, storage, EKS platform integration, CI/CD, and image security.

Prerequisites

This skill requires:

• aws knowledge mcp server tools:

- aws___search_documentation — search AWS documentation - aws___read_documentation — read full documentation pages - aws___recommend — get related documentation

• context7 MCP tools:

- context7_resolve-library-id — resolve K8s library ID - context7_query-docs — query K8s documentation

• AWS CLI (aws) — configured with read access to the target EKS cluster and ECR
• kubectl — configured to access the target EKS cluster
• jq — for parsing JSON output from AWS CLI and kubectl commands

Scope Boundary

This skill focuses on workload-level checks — items that require kubectl or in-cluster inspection. It complements aws-best-practice-research which covers the infrastructure layer (control plane, node groups, addons, etc.).

Workflow

Step 1: Confirm Assessment Scope

Determine from user input:

• Cluster name and AWS Region
• Assessment scope:

- Full cluster — assess all namespaces (excluding kube-system, kube-public, kube-node-lease by default) - Specific namespaces — user-specified list - Specific workloads — user-specified Deployments/StatefulSets

• Include infrastructure layer? — whether to also invoke aws-best-practice-research for the EKS infrastructure layer and merge results (default: yes)

If the user provides only a cluster name, default to full cluster assessment.

Step 2: Environment Detection & Version Awareness

Run the following commands to detect the environment:

# Cluster info via AWS CLI
aws eks describe-cluster --name {CLUSTER} --region {REGION}

# K8s version
kubectl version --output=json

# Node distribution
kubectl get nodes -o wide --no-headers

Record:

• K8s server version (e.g., 1.30) — used for version-aware filtering
• EKS platform version (e.g., eks.15)
• Node count and AZ distribution
• Node instance types

Version-aware filtering rules (apply in Step 3):

• K8s >= 1.25: Check Pod Security Admission (PSA), skip PodSecurityPolicy (PSP)
• K8s < 1.25: Check PSP, note PSA as upgrade recommendation
• K8s >= 1.20: Check Startup Probes
• K8s >= 1.19: Check Topology Spread Constraints
• K8s >= 1.29 + VPC CNI >= 1.21.1: Check Admin Network Policies
• EKS with Pod Identity available: Prefer Pod Identity over IRSA

Step 3: Dynamic Best Practice Research

Research the latest best practices using context7 and aws-knowledge-mcp-server. Run all queries sequentially (one at a time) to avoid rate limiting.

For each of the 8 assessment dimensions, execute the search queries defined in references/search-queries.md. The general flow per dimension is:

1. Query context7 (/websites/kubernetes_io) for K8s official best practices
2. Query aws-knowledge-mcp-server for EKS-specific best practices
3. Read key documentation pages from search results (max 2-3 pages per dimension)
4. Extract check items with specific thresholds and conditions

After all research is complete, merge results with the baseline framework in references/check-dimensions.md to ensure no critical dimension is missed.

Apply version-aware filtering from Step 2 to remove inapplicable items and add version-specific recommendations.

Rate limit protection: If any MCP request returns "Too many requests", wait 5 seconds and retry once. If it fails again, skip and continue. Sequential execution is mandatory.

Step 4: Infrastructure Layer Assessment (Optional)

If infrastructure layer assessment is included (default: yes):

1. Invoke the aws-best-practice-research skill for the EKS cluster
2. Store the infrastructure-layer checklist and assessment results
3. These will be merged into the final report in Step 7

If the user opts out, skip this step.

Step 5: Workload Data Collection

Collect workload configurations using kubectl. Independent commands can run in parallel (they are not subject to MCP rate limits).

See references/kubectl-assessment-commands.md for the complete command list. Key data to collect:

# Core workloads
kubectl get deployments,statefulsets,daemonsets,jobs,cronjobs --all-namespaces -o json

# Pod specifications (within workloads above)
# Already included in the -o json output

# Disruption and scaling
kubectl get pdb,hpa --all-namespaces -o json

# Networking
kubectl get networkpolicies,services,ingresses --all-namespaces -o json

# Security
kubectl get serviceaccounts --all-namespaces -o json
kubectl get clusterrolebindings,rolebindings -o json

# Storage
kubectl get pvc,storageclass -o json

# Namespace labels (for PSA)
kubectl get namespaces -o json

# Events (recent issues)
kubectl get events --all-namespaces --sort-by='.lastTimestamp' -o json

For ECR image scanning (if images are from ECR):

# For each unique ECR image found in workloads
aws ecr describe-image-scan-findings --repository-name {REPO} --image-id imageTag={TAG}
aws ecr describe-repositories --repository-names {REPO}
aws ecr get-lifecycle-policy --repository-name {REPO}

Filter collected data to the assessment scope (namespaces/workloads from Step 1).

Step 6: Per-Dimension Assessment

For each check item from the research phase (Step 3), evaluate every in-scope workload:

For each finding, record:

• Check item ID and name
• Status (PASS/FAIL/WARN/N/A)
• Actual value observed (not just "not configured")
• The specific workload(s) affected
• Version relevance notes (if any)

Step 7: Generate Report and Save to Local File

Generate a single comprehensive report using the template in references/output-template.md and write it directly to a local markdown file.

IMPORTANT — File Writing Rules:

• Use the Write/file tool (not bash heredoc/echo/cat) to create the report file
• If the report is too large for a single write, split into sections: write the file with the first half, then use an append/edit operation to add the remaining sections
• Do NOT output the full report content to the terminal

Use the following file naming convention:

TIMESTAMP=$(TZ=Asia/Shanghai date +%Y-%m-%d-%H-%M-%S)
CLUSTER_SLUG=$(echo "{CLUSTER_NAME}" | tr '[:upper:]' '[:lower:]' | tr ' :/' '-')

Assessment Report — see references/output-template.md

• Full cluster overview
• Compliance scorecard with rating scale, top 3 priorities, and quick stats
• Dimension-by-dimension assessment tables
• Per-workload detail section
• Critical issues and prioritized remediation
• Data sources and reference links
• Save to: ${TIMESTAMP}-${CLUSTER_SLUG}-assessment-report.md

If infrastructure layer results exist from Step 4, merge them into the report.

After saving, print a brief summary to the terminal listing only:

• The file path of the generated report
• Overall compliance score
• Number of PASS / FAIL / WARN findings

Step 8: Remediation Guidance & Next Steps

After saving the reports, offer:

• "I can help fix specific FAIL items — which ones would you like to address?"
• "I can re-run the assessment after remediation to verify improvements."

For Critical Issues (FAIL + High priority), provide:

• Specific remediation commands or manifest changes
• Whether the fix requires workload restart or is in-place
• Impact assessment of the change

Important Guidelines

• Be comprehensive: The value of this skill is thoroughness. Better to include a check and mark it N/A than to miss it.
• Always cite sources: Every check item must reference its source (EKS Best Practices Guide, K8s official docs, etc.).
• Sequential MCP queries: All context7 and aws-knowledge-mcp requests must be sequential. kubectl commands can be parallel.
• Rate limit protection: Wait 5s and retry once on "Too many requests". Skip on second failure.
• Version awareness: Always filter checks by detected K8s/EKS version. Never recommend features unavailable in the cluster's version.
• Actual values in findings: Always report what was observed, not just "not configured". Good: "resources.requests.memory: not set — container has no memory request" Bad: "Memory request missing"
• Per-workload granularity: Report findings at the individual Deployment/StatefulSet level, not just cluster-wide summaries.
• Exclude system namespaces by default: Skip kube-system, kube-public, kube-node-lease unless the user explicitly includes them.
• Respect language: Output in the same language as the user's conversation.
• Infrastructure vs workload boundary: Never duplicate checks from aws-best-practice-research. This skill handles ONLY what requires kubectl/in-cluster access.