ec2-rightsizing-optimizer
Right-sizes EC2 instances for cost optimization using a structured utilization-driven decision framework — gathers 14-30 day CloudWatch utilization data (CPUUtilization, NetworkIn/Out, DiskReadOps/WriteOps, and the load-bearing MemoryUtilization that requires the CloudWatch Agent), interprets Compute Optimizer findings (Optimized, Underprovisioned, Overprovisioned, Not-enough-data) with confidence gating, applies a rightsizing decision matrix (CPU < 30% + Memory < 50% → downsize; CPU > 70% sustained OR Memory > 80% → upsize; network-bound → Enhanced Networking or larger instance; burstable t-family → CPU credit balance check), selects instance families by workload shape (general-purpose m7i/m6i, compute c7i/c7a, memory r7i/r7a/x2idn, storage i4i/im4gn, GPU g5/p5), evaluates Graviton (ARM) migration opportunities with up to 40% better price-performance, and recommends pricing-model optimizations (On-Demand → Reserved Instance 1/3yr for steady-state with 30-72% savings, Compute/Instance Savings Plans for...
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-24
- Stars
- ★ 4
- Forks
- 0
01 前言 Front Matter
- name
- ec2-rightsizing-optimizer
- license
- Apache-2.0
- metadata
- { "tags": "ec2, compute, cost-optimization, finops, right-sizing, graviton, compute-optimizer", "phase": "3", "author": "Jacky Chan — AWS Community Builder", "domain": "aws-cloudops", "family": "Compute", "version": "0.1.0", "keywords": "EC2, right-sizing, cost optimization, Compute Optimizer, CloudWatch Agent, MemoryUtilization, CPUUtilization, NetworkIn, Graviton, ARM, m7i, c7i, r7i, i4i, g5, p5, Reserved Instance, Savings Plans, Spot Instances, FinOps, instance family selection, EBS-optimized, Enhanced Networking, burstable, t-family, CPU credit balance, price-performance", "task_type": "optimize", "complexity": "high", "entry_point": "false", "skill_class": "capability", "when_to_use": "Right-sizing EC2 instances for cost optimization, planning a FinOps rightsizing batch, interpreting Compute Optimizer EC2 findings, evaluating Graviton (ARM) migration opportunities, selecting an instance family for a new workload, comparing pricing models (On-Demand / RI / Savings Plans / Spot), or building a monthly savings estimate for an optimization initiative.", "requires_llm": "true", "verdict_shape": "OPTIMIZED | OPPORTUNITY_FOUND | ALREADY_OPTIMAL", "when_not_to_use": "Auditing Compute Optimizer finding confidence on a single resource (use compute-optimizer-findings-auditor), EBS volume right-sizing (use compute-optimizer-findings-auditor for EBS findings), Lambda memory tuning (use compute-optimizer-findings-auditor), Fargate task sizing (use compute-optimizer-findings-auditor for ECS), or performance troubleshooting for a slow instance (use rds-connectivity-troubleshooter for database reachability, or Performance Insights / CloudWatch metrics directly for EC2). This skill focuses on cost-driven EC2 rightsizing decisions, not single-finding confidence audits or non-EC2 resources.", "invocation_schema": "Input: either (a) an EC2 instance identifier + live-account context, (b) a Compute Optimizer EC2 finding document, OR (c) CloudWatch utilization metrics (CPU, Memory, Network, Disk) for one or more instances with at least 14 days of observation. Output: a deterministic TARGET/VERDICT/REASON/RECOMMENDATION/ESTIMATED_SAVINGS/ MIGRATION_STEPS block per instance, where VERDICT ∈ {OPTIMIZED, OPPORTUNITY_FOUND, ALREADY_OPTIMAL} and RECOMMENDATION includes the target instance type, family, pricing model, and Graviton flag.", "supports_pipeline": "true", "invocation_example": "# Minimal valid input (offline finding classification):\nInstanceId: i-0abc123\nCurrent instance type: m5.2xlarge\nRegion: us-east-1\nCompute Optimizer finding: Overprovisioned\nFinding reasons: [\"CPUOverprovisioned\", \"MemoryOverprovisioned\"]\nUtilization metrics (last 30 days):\n - CPUUtilization: avg=8%, max=15%, source=CloudWatch\n - MemoryUtilization: avg=22%, max=30%, source=CloudWatchAgent\n - NetworkIn: avg=0.05 MB/s, max=0.1 MB/s\n - DiskReadOps+DiskWriteOps: avg=10/s, max=20/s\nRecommendation options:\n - rank: 1, instanceType: t3.large, performanceRisk: 1,\n savingsOpportunity: {savingsPercentage: 75, estimatedMonthlySavings: 150}\nPricing: On-Demand, no RI/Savings Plan commitment.\nEmit the standard optimization block (TARGET, VERDICT, REASON,\nRECOMMENDATION, ESTIMATED_SAVINGS, MIGRATION_STEPS).", "activation_triggers": "right-size EC2 instances, EC2 cost optimization, Compute Optimizer EC2 recommendations, EC2 underutilized, EC2 overprovisioned, FinOps EC2 savings, Graviton migration opportunity, EC2 Reserved Instance, Savings Plan recommendation, EC2 instance family selection, EC2 monthly savings estimate, EC2 fleet rightsizing, t-family CPU credit exhaustion, m5 to m7i migration, EC2 pricing model optimization, spot instance opportunity" }
- description
- Right-sizes EC2 instances for cost optimization using a structured utilization-driven decision framework — gathers 14-30 day CloudWatch utilization data (CPUUtilization, NetworkIn/Out, DiskReadOps/WriteOps, and the load-bearing MemoryUtilization that requires the CloudWatch Agent), interprets Compute Optimizer findings (Optimized, Underprovisioned, Overprovisioned, Not-enough-data) with confidence gating, applies a rightsizing decision matrix (CPU < 30% + Memory < 50% → downsize; CPU > 70% sustained OR Memory > 80% → upsize; network-bound → Enhanced Networking or larger instance; burstable t-family → CPU credit balance check), selects instance families by workload shape (general-purpose m7i/m6i, compute c7i/c7a, memory r7i/r7a/x2idn, storage i4i/im4gn, GPU g5/p5), evaluates Graviton (ARM) migration opportunities with up to 40% better price-performance, and recommends pricing-model optimizations (On-Demand → Reserved Instance 1/3yr for steady-state with 30-72% savings, Compute/Instance Savings Plans for...
- compatibility
- Agent runtime that reads SKILL.md (Claude Code, Cursor, Windsurf, Codex, Gemini). Offline recommendation-document classification works from pasted Compute Optimizer findings and CloudWatch metrics. Live-account optimization uses aws ec2 describe-instances, aws cloudwatch get-metric-statistics, aws compute-optimizer get-ec2-instance-recommendations, aws ce get-cost-and-usage, and aws ec2 describe-instance-types (AWS CLI v2, SSO or key-based credentials).
02 资源文件 20
- 参考文档eval/test-cases.yaml8.5 KB
- 参考文档evals/baselines/already-optimal-with-savings-plan.md445 B
- 参考文档evals/baselines/memory-data-missing-blocked.md428 B
- 参考文档evals/baselines/overprovisioned-downsize-graviton.md798 B
- 参考文档evals/baselines/t-family-credit-exhaustion.md454 B
- 参考文档evals/baselines/underprovisioned-upsize-memory.md411 B
- 参考文档evals/eval_config.yaml1.8 KB
- 参考文档evals/evals.json4.4 KB
- 参考文档evals/prompts/already-optimal-with-savings-plan.md900 B
- 参考文档evals/prompts/memory-data-missing-blocked.md845 B
- 参考文档evals/prompts/overprovisioned-downsize-graviton.md1.1 KB
- 参考文档evals/prompts/t-family-credit-exhaustion.md997 B
- 参考文档evals/prompts/underprovisioned-upsize-memory.md1.0 KB
- 参考文档examples/README.md8.0 KB
- 参考文档references/advanced-patterns.md17.1 KB
- 参考文档references/diagnostic-commands.md5.0 KB
- 参考文档references/error-handling.md6.7 KB
- 参考文档references/instance-selection-reference.md16.3 KB
- 参考文档references/worked-examples.md3.3 KB
- 参考文档skill.usf.yaml4.6 KB
03 SKILL.md 正文
EC2 Rightsizing Optimizer
Quick start
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Mindset
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Philosophy
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Quick reference — verdict thresholds
| Observation (14-30 day window) | Verdict | Recommendation |
|---|---|---|
| CPU < 30% avg AND Memory < 50% avg AND Network < 30% limit AND Disk low | OPPORTUNITY_FOUND (downsize) | Step 4 — downsize 1-2 sizes, prefer Graviton equivalent |
| CPU > 70% sustained OR Memory > 80% OR Network > 50% limit | OPPORTUNITY_FOUND (upsize — performance risk) | Step 5 — upsize to next class that resolves the bottleneck |
| Burstable t-family with CPUCreditBalance chronically near 0 | OPPORTUNITY_FOUND (architecture change) | Step 6 — migrate to m-family or enable Unlimited |
| Graviton-compatible workload on x86 with stable utilization | OPPORTUNITY_FOUND (price-performance) | Step 7 — Graviton migration during next AMI refresh |
| On-Demand pricing on steady-state workload | OPPORTUNITY_FOUND (pricing model) | Step 8 — RI or Compute Savings Plan |
| MemoryUtilization metric absent | NEED_MORE_INFO (BLOCKED — not ALREADY_OPTIMAL) | Install CloudWatch Agent, wait 14-30 days, re-evaluate |
Compute Optimizer finding Optimized + metrics present + already on RI/SP |
ALREADY_OPTIMAL | None — continue monitoring |
Compute Optimizer finding NotOptimized (insufficient data) |
NEED_MORE_INFO | Wait for additional observation |
| All thresholds pass for current type AND pricing already optimized | OPTIMIZED | None required |
See the ordered steps for edge cases (cross-family migration, EBS-bound workloads, Spot-eligible workloads, multi-AZ fleets).
Pre-flight: data gate (run before any right-sizing decision)
Right-sizing decisions are only as good as the underlying data. Several data-quality conditions short-circuit the optimization — misclassifying them produces recommendations that fail at runtime.
Required data sources
Moved verbatim to references/diagnostic-commands.md - load on demand (see References below).
Data-quality short-circuits
| Condition | Effect on optimization |
|---|---|
MemoryUtilization metric absent (CWAgent not installed) |
NEED_MORE_INFO: Install CWAgent, wait 14-30 days, re-evaluate. Do NOT recommend a downsize. |
| Observation window < 14 days | NEED_MORE_INFO: workload may reflect atypical load (deploy week, incident response, seasonal dip). Minimum 14 days; 30 days preferred. |
| Instance stopped for > 50% of window | Metrics are skewed. Re-pull after a stable 14-day running window. |
Compute Optimizer enrollment Inactive |
No Compute Optimizer findings to cross-reference. Rely on CloudWatch metrics directly. |
Compute Optimizer lastRefreshTimestamp > 30 days old |
Stale finding — workload may have changed. Re-run get-ec2-instance-recommendations. |
Instance in Stopped state at evaluation time |
Right-sizing is moot until the instance restarts. Note and skip. |
Conflicting-data arbitration
When CloudWatch metrics and Compute Optimizer findings disagree (e.g., CloudWatch shows CPU 60% but Compute Optimizer says Overprovisioned), trust CloudWatch. Compute Optimizer averages over 30 days and may smooth out recent workload spikes. The freshest signal wins.
Process — Optimization logic (apply in order)
Step 0: Non-obvious behaviours that change the recommendation
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Step 1: Validate input and data sufficiency
If MemoryUtilization is absent from the input metrics (CWAgent not
installed), emit NEED_MORE_INFO with a specific installation instruction:
Moved verbatim to references/worked-examples.md - load on demand (see References below).
If the observation window is < 14 days, emit NEED_MORE_INFO: "Observation window is N days; minimum 14 days required for representative data." Proceed only if both conditions are met.
Step 2: Compute Optimizer reconciliation
If a Compute Optimizer finding is provided, reconcile with CloudWatch metrics. Compute Optimizer's 30-day analysis is a sanity check; the fresh CloudWatch signal wins on disagreement.
Moved verbatim to references/diagnostic-commands.md - load on demand (see References below).
| Field to verify | What it tells you | Action if missing/stale |
|---|---|---|
finding |
High-level classification (Optimized/Under/Over) | If absent, finding is NotOptimized due to insufficient data → emit NEED_MORE_INFO. |
recommendationOptions[].performanceRisk |
Risk that the recommended type cannot handle the workload (1=low ... 5=high) | Always prefer options with performanceRisk ≤ 2 for production. If only risk ≥ 4 options exist, emit NEED_MORE_INFO and investigate manually. |
lastRefreshTimestamp |
When Compute Optimizer last re-analyzed | If > 30 days old, treat as stale; re-run get-ec2-instance-recommendations or rely on CloudWatch. |
utilizationMetrics[] |
Source data Compute Optimizer used | Cross-check against your CloudWatch pull. If they disagree, freshest CloudWatch wins. |
savingsOpportunity.estimatedMonthlySavings |
Dollar savings estimate | Sanity-check against your own hourly math (Step 9). Compute Optimizer uses list price; your actual RI/SP rate may differ. |
| Compute Optimizer finding | CloudWatch agreement | Action |
|---|---|---|
Overprovisioned + HIGH confidence (Memory metric present, performanceRisk ≤ 2) |
CPU < 30% AND Memory < 50% | Proceed to Step 4 (downsize). |
Overprovisioned + LOW confidence (Memory inferred, performanceRisk ≥ 4) |
— | Treat as NEED_MORE_INFO per the compute-optimizer-findings-auditor. Install CWAgent first. |
Underprovisioned |
CPU > 70% OR Memory > 80% | Proceed to Step 5 (upsize). |
Optimized |
All thresholds pass for current type | Proceed to Step 8 (pricing model check). |
NotOptimized (insufficient data) |
— | NEED_MORE_INFO: wait for additional observation. |
Step 3: Workload classification
Classify the workload shape to guide family selection (Step 4-5):
| Workload shape | Indicators | Recommended family |
|---|---|---|
| Balanced (web server, app server, dev) | CPU ~ Memory ~ Network balanced | General purpose: m7i, m7a, m7g (Graviton), m6i, m6a |
| Compute-bound (batch, HPC, CI runners, video encoding) | CPU > Memory utilization | Compute optimized: c7i, c7a, c7g, c6i, c6a |
| Memory-bound (in-memory cache, big-data analytics, relational DB) | Memory > CPU utilization, large working set | Memory optimized: r7i, r7a, r7g, r6i, x2idn, x2iedn |
| Storage-bound (NoSQL, data warehousing, OLAP) | High disk ops, large local storage need | Storage optimized: i4i, im4gn, i4g |
| GPU (ML training/inference, rendering) | GPU utilization, CUDA workloads | GPU: g5, g6, p5, p4d |
| Network-bound (real-time streaming, HFT) | NetworkIn + NetworkOut > 50% of limit | Network optimized: c6n, m6n, r6n; or Enhanced Networking (ENA) |
| Bursty (dev/test, low-traffic web) | t-family with healthy CPUCreditBalance | Burstable: t3, t3a, t4g |
Step 4: Downsize decision (CPU/Memory/Network/Disk all low)
Apply the downsize matrix when CPU < 30% avg AND Memory < 50% avg AND Network < 30% of limit AND Disk < 30% of provisioned IOPS.
Downsize magnitude:
- If CPU < 10% AND Memory < 30%: downsize 2 sizes (e.g., m5.4xlarge → m5.large).
- If CPU < 30% AND Memory < 50%: downsize 1 size (e.g., m5.2xlarge → m5.large).
- Conservative default: downsize 1 size, observe 7 days, downsize again if utilization remains low. This staged approach catches edge cases that a 2-size jump misses.
Family selection for downsize:
- Prefer staying within the same family (m5.2xlarge → m5.large) for lowest migration risk.
- If the workload is bursty with healthy credit balance, consider migrating to t-family (t3/xlarge saves 60%+ vs m5.xlarge).
- If Graviton-compatible, prefer the g-suffix equivalent (m7g.large vs m7i.large: ~20% additional savings).
Burstable family (t3/t3a/t4g) check:
- Verify the workload's burst pattern: average CPU < 20% baseline (below credit-earning rate) and burst duration < 1 hour/day.
- If the workload sustains > 20% CPU, do NOT migrate to t-family — use Unlimited mode (charges for overage) or stay on m-family.
Step 5: Upsize decision (CPU or Memory high)
Apply the upsize matrix when CPU > 70% sustained OR Memory > 80% OR Network > 50% of limit.
Upsize magnitude:
- Identify the bottleneck dimension (CPU, Memory, Network, or Disk).
- Upsize to the next instance class that addresses the bottleneck:
- CPU-bound: same family, more vCPUs (m5.large → m5.xlarge).
- Memory-bound: switch to memory-optimized family if Memory > 80% persistently (m5.xlarge → r6i.xlarge, doubling memory per vCPU).
- Network-bound: switch to network-optimized family (m5.xlarge → c5n.large or enable Enhanced Networking).
- Disk-bound: migrate to storage-optimized family OR move the workload to EBS-optimized io2 volumes.
Severity grading for upsizing:
- HIGH: CPU sustained > 90% OR Memory > 95% — active performance degradation; upsize immediately.
- MEDIUM: CPU 70-90% OR Memory 80-95% — performance risk; upsize at next maintenance window.
- LOW: peaks above 70% but average < 50% — investigate via Performance Insights / CloudWatch before upsizing (might be a query optimization issue, not capacity).
Step 6: Burstable t-family architecture change
If the current instance is t-family AND CPUCreditBalance chronically drops below ~50 credits (visible in CloudWatch), the workload is not burst-compatible. Two remediation paths:
- Enable Unlimited mode (cheapest, preserves instance):
aws ec2 modify-instance-credit-specification --instance-id <id> --cpu-credits unlimited. Charges for overage credits but maintains baseline performance. Recommended for workloads with occasional sustained-CPU bursts. - Migrate to m-family (most robust, higher baseline cost): migrate from t3.large to m6i.large (or m7g.large for Graviton). Eliminates the credit-balance cliff entirely. Recommended for workloads with sustained CPU > 20%.
Step 7: Graviton (ARM) migration evaluation
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Step 8: Pricing model optimization
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Step 9: Impact estimation
Compute the monthly savings for each recommendation:
Monthly savings = (current_hourly - new_hourly) * 730 hours
Include EBS implications (instance-type change may unlock/include EBS-optimization costs; current-gen instances have EBS-optimized built in, legacy may not).
Include data transfer implications (instance-type change may have different network performance characteristics — usually a non-issue for same-family right-sizes, material for cross-family).
For pricing-model changes (RI/Savings Plan), compute savings separately:
RI savings = On-Demand hourly * (1 - RI discount) * hours
Total savings = rightsize savings + pricing-model savings.
Step 10: Final verdict
The verdict is the worst-case (most-actionable) finding across all dimensions:
- If any dimension recommends a change (downsize, upsize, architecture, Graviton, pricing model), verdict is OPPORTUNITY_FOUND.
- If all dimensions pass AND pricing is already optimized (RI/Savings Plan in place), verdict is OPTIMIZED.
- If all dimensions pass for current type but pricing model could be improved, verdict is OPPORTUNITY_FOUND (pricing dimension).
- If data is insufficient (memory absent, window < 14 days), verdict is NEED_MORE_INFO.
Output format
TARGET: <instance-id or fleet-description>
VERDICT: OPTIMIZED | OPPORTUNITY_FOUND | ALREADY_OPTIMAL
REASON: <1-2 sentences naming the recommendation and the supporting data>
RECOMMENDATION:
Current: <instance-type> at <pricing-model> in <region>
Proposed: <instance-type> at <pricing-model> in <region>
Graviton: <yes/no/N/A>
Family change: <yes/no>
Confidence: <HIGH/MEDIUM/LOW> — <one-line rationale>
ESTIMATED_SAVINGS:
Monthly (rightsize): $<amount>
Monthly (pricing model): $<amount>
Annual total: $<amount>
Assumptions: <list (730h/month, On-Demand baseline, etc.)>
MIGRATION_STEPS:
1. <specific action with CLI command>
2. <verification step>
CONFIRM: Before executing any state-changing CLI, emit and await operator
approval: "CONFIRM: About to <action> on <instance-id> in <region>.
Proceed? (yes/no)"
Worked example — overprovisioned, high-confidence downsize + Graviton
TARGET: i-0abc123
VERDICT: OPPORTUNITY_FOUND
REASON: m5.2xlarge at 8% CPU / 22% Memory / low Network/Disk over 30 days
is a clear downsize candidate. Graviton-compatible workload (Java 17);
m7g.large saves 75% on the rightsize plus 20% on Graviton price-
performance. Adding a 3-year Compute Savings Plan captures additional
40% on the residual spend (Step 4 + Step 7 + Step 8).
RECOMMENDATION:
Current: m5.2xlarge at On-Demand in us-east-1
Proposed: m7g.large at 3-year Compute Savings Plan in us-east-1
Graviton: yes (Java 17, multi-arch AMI available)
Family change: yes (m5 → m7g)
Confidence: HIGH — Memory metric present (CWAgent), CPU+Memory both
well below thresholds, Compute Optimizer cross-check agrees
(Overprovisioned, performanceRisk 1).
ESTIMATED_SAVINGS:
Monthly (rightsize): $232.32 (m5.2xlarge $0.384/h → m7g.large $0.0644/h;
730h × $0.3196 delta)
Monthly (pricing model): $35.96 (40% off m7g.large On-Demand via
3-yr Compute Savings Plan)
Annual total: $3,219.84
Assumptions: 730h/month, us-east-1 On-Demand pricing as of 2026-08-07,
workload steady-state, no significant growth expected.
MIGRATION_STEPS:
1. Provision the new instance:
aws ec2 run-instances --image-id <ami-arm64> --instance-type m7g.large
--key-name <key> --security-group-ids <sg> --subnet-id <subnet>
--tag-specifications "ResourceType=instance,Tags=[{Key=Name,
Value=i-0abc123-m7g}]"
2. Migrate application data / EBS volumes:
- Create an AMI from i-0abc123 (for rollback).
- Or attach the existing EBS volume to the new instance (requires
stop/detach/attach).
3. Validate application functionality on m7g.large for 24-48 hours
(focus on memory leaks, JIT behaviour, GC pauses under load).
4. Cutover DNS / load balancer to the new instance.
5. Purchase a 3-year Compute Savings Plan for the m7g.large commit:
aws savingsplans create-savings-plan --savings-plan-offering-id <id>
--commitment <amount>
6. Decommission i-0abc123 once the new instance is verified.
CONFIRM: Before provisioning the new instance, emit and await:
"CONFIRM: About to run-instances m7g.large in us-east-1 for rightsize
of i-0abc123. This will incur ~$47/month On-Demand until the Savings
Plan is in place. Proceed? (yes/no)"
Do NOT run the CLI until the operator confirms.
Worked example — already optimal
Moved verbatim to references/worked-examples.md - load on demand (see References below).
Worked example — NEED_MORE_INFO (memory data missing)
Moved verbatim to references/worked-examples.md - load on demand (see References below).
Verdict semantics — reconciling the verdict_shape
The verdict_shape metadata declares the three primary verdicts
(OPTIMIZED | OPPORTUNITY_FOUND | ALREADY_OPTIMAL). Two additional
data-gating verdicts (NEED_MORE_INFO, BLOCKED) appear in the
workflow when the data required for a confident decision is missing.
Treat them as pre-decision guards, not as alternatives to the primary
three:
| Verdict | When to emit | Position in workflow |
|---|---|---|
OPPORTUNITY_FOUND |
At least one dimension (size, family, Graviton, pricing) has a concrete, savings-bearing recommendation. | Primary — terminal for actionable findings. |
OPTIMIZED |
A change was applied and verified this session; metrics confirm the new size lands within healthy bands. | Primary — only emitted post-remediation. |
ALREADY_OPTIMAL |
All dimensions pass for the current type AND the pricing model is already committed (RI/Savings Plan covering ≥ 95% of steady-state spend). | Primary — terminal for healthy findings. |
NEED_MORE_INFO |
Data gate failed: MemoryUtilization absent, observation window < 14 days, Compute Optimizer NotOptimized due to insufficient data, or freshest CloudWatch signal disagrees with an old Compute Optimizer finding. |
Pre-decision — emit before any sizing recommendation; the next action is data collection, not remediation. |
BLOCKED |
A hard precondition prevents evaluation: Compute Optimizer enrollment Inactive AND no CloudWatch metrics retrievable, instance is in Stopped/Terminated for > 50% of the window, or IAM denies cloudwatch:GetMetricStatistics. |
Pre-decision — emit when no reliable signal exists at all. |
Rule: never emit OPPORTUNITY_FOUND without first discharging every
NEED_MORE_INFO/BLOCKED gate in Step 1. A downsize recommendation
that hides a missing MemoryUtilization signal is the single highest-
risk misclassification the skill can make.
Network-limit calculation (concrete formula)
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
Error handling — CLI and data-source failures
Moved verbatim to references/error-handling.md - load on demand (see References below).
Anti-Patterns — NEVER
-
NEVER recommend a downsize without MemoryUtilization data. CPU is the visible signal; memory is the load-bearing constraint. A downsize based on CPU alone may crash a memory-bound workload.
-
NEVER recommend Spot Instances for stateful workloads. Databases, message queues, singleton services, and any instance holding local state will lose data on Spot interruption. Spot is for stateless, fault-tolerant workloads only.
-
NEVER assume Graviton compatibility without verification. Most JVM/ Python/Go workloads work; C/C++ and any code with platform-specific binaries need recompilation. Always check
describe-instance-typesand the application runtime support matrix. -
NEVER recommend migrating to a t-family (t3/t3a/t4g) without checking the workload's burst pattern. Sustained CPU > 20% baseline will exhaust CPUCreditBalance and trigger the 20% baseline performance cliff. Use Unlimited mode or stay on m-family for sustained workloads.
-
NEVER recommend a Standard RI for a workload with uncertain growth. Standard RIs cannot be exchanged across families. Use Convertible RI or Compute Savings Plan for flexibility insurance.
-
NEVER assume a 30-day observation window is sufficient for seasonal workloads. E-commerce (holiday peaks), tax software (April), and education (semester starts) all have seasonal patterns invisible in a generic 30-day window. Pull data from peak and off-peak periods before right-sizing.
-
NEVER right-size based on a single metric (CPU only, Memory only, Network only). Require at minimum CPU + Memory + one more dimension for HIGH-confidence downsize recommendations.
-
NEVER skip the rollback plan for production right-sizes. Always create an AMI before stopping/modifying; always keep the original instance for 24-48 hours post-cutover; always document the rollback procedure.
-
NEVER trust memorized instance specs. AWS releases new generations frequently; always query
aws ec2 describe-instance-typesfor current vCPU, memory, network, and storage specs. -
NEVER recommend a Zonal RI for a workload that may need to migrate AZs. Zonal RIs are tied to a specific AZ; Regional RIs and Compute Savings Plans apply across AZs in the region. Default to Regional unless there's a specific capacity-reservation need.
-
NEVER stack a right-size and a pricing-model change in a single step. Right-size first (utilization change), verify for 7 days, then adjust the pricing model commitment. Stacking both at once obscures which change produced the savings.
-
NEVER recommend Compute Savings Plan commitments beyond the steady- state baseline. The flexible nature of Savings Plans means the commitment applies regardless of instance mix, but committing beyond steady-state leaves you paying for unused capacity. Right-size first, identify the steady-state, then commit.
-
NEVER ignore EBS-optimization costs in legacy migrations. m4.10xlarge and earlier (some m4/c4 generations) charge hourly for EBS-optimized status; current-gen (m5+) includes it free. A migration from m4 to m6i unlocks "free" EBS optimization in addition to other savings.
-
NEVER treat Compute Optimizer's recommendation options at rank 1 as the safest. Rank 1 is the cheapest (highest savings); it may carry the highest performanceRisk. Always check
performanceRiskon each option and prefer risk ≤ 2 for production workloads. -
NEVER right-size without considering cross-AZ transfer costs. A right-size that moves an instance to a different AZ (via replacement) may shift cross-AZ data-transfer patterns. Same-AZ right-sizes have no transfer impact.
-
NEVER recommend Unlimited mode as a permanent fix for a t-family workload with sustained CPU > 50%. Unlimited charges for overage credits indefinitely; migrating to m-family is cheaper above ~40% sustained utilization.
-
NEVER recommend Spot without checking the Spot Placement Score (SPS) for the target instance type and AZ. SPS predicts the likelihood of Spot capacity being available; a low score (≤ 10) means Spot requests in that AZ/type are frequently unfulfilled or interrupted. Without SPS, a "90% savings" Spot recommendation may be unlaunchable in practice, or may interrupt so frequently that the workload fails. Always pull
aws ec2 get-spot-placement-scores --instance-types <type> --target-capacity 1 --region-name <region>and surface the SPS in the recommendation. Prefer types with SPS ≥ 50 for production- adjacent Spot workloads. -
NEVER recommend a Standard RI when the workload may need to change instance families within the commitment term. Why Standard RIs cannot be exchanged: AWS models Standard RIs as a fixed reservation of a specific instance family + size + AZ (Zonal) or family + size (Regional). The discount is bound to that exact configuration for the full 1- or 3-year term because AWS uses the commitment to provision underlying capacity. Convertible RIs allow exchanges (family, size, OS, tenancy) but charge a slightly lower discount in exchange for the optionality. If a right-size may trigger a family migration (e.g., m5 → r6i for memory-bound drift), default to Convertible RI or Compute Savings Plan. Standard RI is only appropriate when the workload is locked to a single family for the full term (e.g., a vendor-certified appliance that pins to c6i).
-
NEVER interpret Compute Optimizer
Optimizedas ALREADY_OPTIMAL without checking the pricing model separately. Compute Optimizer evaluates utilization only — it does not flag On-Demand spend that could move to an RI or Savings Plan. An "Optimized" instance may still be 100% On-Demand, missing 30-72% in commitment savings. Always run Step 8 (pricing model) before emitting ALREADY_OPTIMAL.
Pre-flight safety checks (run before any remediation CLI)
Moved verbatim to references/diagnostic-commands.md - load on demand (see References below).
Remediation guidance
Moved verbatim to references/error-handling.md - load on demand (see References below).
Deep reference: EC2 instance selection
Moved verbatim to references/instance-selection-reference.md - load on demand (see References below).
Recent AWS features (2024-2026)
Moved verbatim to references/advanced-patterns.md - load on demand (see References below).
References (load on demand)
- references/advanced-patterns.md — Quick start rules, Mindset, Philosophy, Step 0 expert behaviours, Step 7 Graviton and Step 8 pricing deep-dives, network-limit formula, and 2024-2026 features, moved verbatim from SKILL.md
- references/worked-examples.md — secondary worked examples (already optimal, NEED_MORE_INFO memory-missing) and the Step 1 NEED_MORE_INFO emit template, moved verbatim from SKILL.md
- references/error-handling.md — CloudWatch/Compute Optimizer/EC2 API failure tables and per-verdict remediation guidance, moved verbatim from SKILL.md
- references/diagnostic-commands.md — pre-flight data-source CLI, Compute Optimizer parsing, and pre-flight safety checks, moved verbatim from SKILL.md
- references/instance-selection-reference.md — extended with the instance family taxonomy, Graviton generations, pricing comparison, credit math, and network tiers, moved verbatim from SKILL.md
Domain
AWS CloudOps / EC2 Compute Cost Optimization, FinOps & Right-Sizing.
AWS documentation
- Amazon EC2 User Guide — Instance types — https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-types.html
- AWS Compute Optimizer User Guide — https://docs.aws.amazon.com/compute-optimizer/latest/ug/what-is.html
- AWS CloudWatch Agent User Guide — https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/Install-CloudWatch-Agent.html
- AWS Savings Plans User Guide — https://docs.aws.amazon.com/savingsplans/latest/userguide/
- Amazon EC2 pricing — https://aws.amazon.com/ec2/pricing/
- AWS Graviton — https://aws.amazon.com/ec2/graviton/
- AWS CLI EC2 reference — https://docs.aws.amazon.com/cli/latest/reference/ec2/
- AWS Well-Architected Framework — Cost Optimization — https://docs.aws.amazon.com/wellarchitected/latest/cost-optimization-pillar/welcome.html
04 同仓库更多 Skills
xray-tracing-deployer
Deploys AWS X-Ray distributed tracing with production-grade configuration: X-Ray daemon deployment (EC2 systemd, ECS sidecar, EKS DaemonSet, Lambda built-in), SDK instrumentation (Java, Python, Node.js, Go, .NET) with annotations and metadata, sampling rules (default reservoir + rate vs custom rules with URL / method / service-name predicates), service map generation, trace query and retrieval, groups and insights, CloudWatch ServiceLens integration, IAM permissions (PutTraceSegments, GetSamplingRules), and latest features (Lambda Powertools tracing, OpenTelemetry / ADOT Collector support, W3C trace context). Emits a READY_TO_DEPLOY checklist with every configuration item verified. Use when enabling X-Ray on a new application, adding tracing to ECS / EKS / Lambda, validating a sampling configuration, or generating daemon deployment CLI and IaC templates. Triggers: X-Ray, distributed tracing, X-Ray daemon, sampling rules, service map, X-Ray SDK, OpenTelemetry, ADOT, Lambda Powertools, CloudWatch ServiceLens.
wellarchitected-workload-auditor
Audits AWS Well-Architected Tool workloads for review staleness, per-pillar high-risk issue counts, milestone tracking gaps, and remediation plan completeness. Emits a deterministic verdict (STALE_REVIEW | HIGH_RISK | CONFIG_GAP | OK) per workload with per-pillar risk breakdown, milestone trend analysis, and specific remediation. Use when reviewing Well-Architected workloads, checking WA review freshness, auditing high-risk issues per pillar, validating milestone coverage, or tracking remediation plans.
wellarchitected-review-operator
Operates AWS Well-Architected Tool reviews end-to-end — creates workloads, runs pillar reviews (operational excellence, security, reliability, performance, cost optimization, sustainability, Prosperity), answers questions with risk-tier rationale, generates reports and improvement plans, creates milestones, and integrates with Trusted Advisor and Well-Architected Labs. Runs deterministic pre-checks (workload existence, lens availability, pillar coverage), emits the exact update-answer / get-consolidated-report CLI behind a CONFIRM gate, and verifies risk-tier transitions. Emits READY | BLOCKED | COMPLETED. Use when creating a workload, answering pillar questions, generating an improvement plan, creating a milestone, or integrating with Trusted Advisor. Triggers: well-architected tool, well architected review, workload review, pillar, operational excellence, security, reliability, performance, cost optimization, sustainability, prosperity, improvement plan, milestone, trusted advisor.
wafv2-web-acl-deployer
Provisions AWS WAFv2 Web ACLs with correct production defaults: scope selection (CLOUDFRONT vs REGIONAL), managed rule groups (AWSManagedRulesCommonRuleSet, SQLiRuleSet, LinuxRuleSet, WindowsRuleSet, IPReputationList, BotControlRuleSet, AmazonIpReputationList, ATP), custom rules (byte match, IP set, regex pattern, geo match, size constraint, rate-based), rule priority ordering, logging destinations (Kinesis Firehose, CloudWatch Logs, S3), resource association (ALB, API Gateway, CloudFront), rate-based rules with aggregate keys (IP, forwarded IP), Challenge and CAPTCHA actions, and account takeover prevention. Emits a READY_TO_DEPLOY checklist. Use when creating a Web ACL, attaching managed rules, configuring rate limiting, enabling WAF logging, associating a Web ACL with an ALB/API Gateway/CloudFront, or configuring CAPTCHA/Challenge actions. Triggers: create WAF Web ACL, WAFv2, managed rule groups, WAF rate limiting, WAF logging, WAF CAPTCHA, WAF Challenge action, bot control, ATP managed rules.
wafv2-web-acl-auditor
Audits AWS WAFv2 Web ACL configurations to determine whether the ACL provides effective protection: default-action posture (Allow vs Block), managed-rule-group coverage gaps, rule effectiveness (BLOCK vs COUNT, shadow/bypass rules, stale exclusions), rate-based rule correctness, logging and visibility configuration, and text-transformation bypass vectors. Emits a deterministic verdict per Web ACL. Use when reviewing a WAFv2 Web ACL for security posture, checking managed-rule coverage, validating rate-limit configuration, auditing WAF logging, or hardening a WAF before production deployment. Triggers: WAF, WAFv2, Web ACL, managed rule group, AWSManagedRules, rate-based rule, CAPTCHA, Challenge, OverrideAction, Count mode, forwarded IP, Bot Control, ATP, text transformation, WAF logging, WAF visibility, OWASP, SQLi, XSS, RFI.
waf-rule-deployer
Provisions AWS WAFv2 rule sets with production defaults: Web ACL creation (CloudFront scope in us-east-1 vs Regional scope), rule groups (managed vs custom) with capacity planning in WCU (Web ACL Capacity Units, 1500 default per ACL), managed rule groups (AWSManagedRulesCommonRuleSet, AWSManagedRulesSQLiRuleSet, AWSManagedRulesAmazonIpReputationList, AWSManagedRulesLinuxRuleSet, AWSManagedRulesWindowsRuleSet, AWSManagedRulesUnixRuleSet), custom rules (byte-match, regex pattern set, geo-match, rate-based), IP set and regex pattern set resources, rule priority ordering (lowest number evaluated first), action types (allow, block, count, CAPTCHA, Challenge), bot control managed rule group, account takeover. Triggers: create WAF rule, WAF managed rule group, WAF custom rule, WAF byte match, WAF geo match, WAF rate based rule, WAF IP set, WAF regex pattern set, WAF WCU budget, WAF CloudFront ACL, WAF logging Kinesis Firehose, WAF bot control, WAF ATP, WAF label matching, WAF Challenge action, WAF CAPTCHA.