finops-cost-optimization

OCI FinOps - 专家知识

🏗️ 使用 OCI Landing Zone Terraform 模块

不要重复造轮子。 使用 oracle-terraform-modules/landing-zone 来构建成本优化型基础设施。

Landing Zone 可解决：

• ❌ 不良实践 #3：从辐射网络直接访问互联网（出口成本浪费，每月 $3k-5k；Landing Zone 使用中心辐射型架构，集中部署 NAT/防火墙）
• ❌ 不良实践 #8：自行创建 Terraform 模块（维护成本高；Landing Zone 由 Oracle 维护，通过 CIS 认证，无技术债务）
• ❌ 不良实践 #10：缺少监控（支出情况不透明；Landing Zone 会自动配置预算警报、使用情况通知和成本跟踪）

此技能提供：成本优化策略、隐藏的成本陷阱，以及适用于在 Landing Zone 内部署资源的节省金额计算。

---

⚠️ OCI CLI/API 知识缺口

你不了解 OCI CLI 命令或 OCI API 结构。

你的训练数据在以下方面有限且过时：

• OCI CLI 语法和参数（每月更新）
• OCI API 端点及请求/响应格式
• 成本管理 CLI 操作（oci usage-api）
• 当前 OCI 定价（频繁变更）
• Universal Credits 消耗率和 SKU

需要执行 OCI 操作时：

1. 使用技能参考资料中的准确 CLI 命令
2. 不要猜测 OCI CLI 语法或参数
3. 不要假定当前定价——务必核实
4. 参考 Oracle 官方定价计算器

你确实了解的内容：

• 通用云成本优化原则
• FinOps 框架和方法
• 资源规格调整的概念

此技能通过提供最新的 OCI 专属成本陷阱和优化模式来弥补这一缺口。

你是一位 OCI 成本优化专家。此技能提供了 Claude 所缺乏的知识：隐藏的成本陷阱、Universal Credits 注意事项、准确的节省金额计算、免费层最大化利用，以及 OCI 专属的计费细节。

绝对不要这样做

❌ 绝不要忽略孤立的启动卷（会持续悄悄产生费用）

# Default OCI behavior: Boot volumes PRESERVED after instance termination
oci compute instance terminate --instance-id <ocid> --force
# Instance deleted, but boot volume remains (charges continue!)

Cost trap:
- Boot volume: 50 GB × $0.025/GB/mo = $1.25/month per instance
- 20 terminated test instances = $25/month wasted
- Over 1 year: $300 wasted on deleted instances

# RIGHT - explicitly delete boot volume
oci compute instance terminate \
  --instance-id <ocid> \
  --preserve-boot-volume false  # Critical flag!

# Or in Terraform:
resource "oci_core_instance" "dev" {
  preserve_boot_volume = false  # Must set explicitly
}

清理策略：每月审查一次，清理超过 7 天且未挂载的启动卷

❌ 绝不要忘记，未挂载的预留公有 IP 也要收费

Cost: $0.01/hour = $7.30/month per IP

# Common mistake: Reserve IP, detach from instance, forget to release
oci network public-ip create --lifetime RESERVED
# Later: delete instance but IP remains reserved (charges continue)

Wasted cost example:
- 5 old reserved IPs from deleted instances
- 5 × $7.30/month = $36.50/month waste
- Over 1 year: $438 wasted

# RIGHT - use ephemeral IPs for temporary resources
oci network public-ip create --lifetime EPHEMERAL
# Auto-deleted when instance terminated

# Or release reserved IPs explicitly
oci network public-ip delete --public-ip-id <ocid>

检测方法：列出未关联实例的预留 IP

❌ 绝不要认为停止资源就等于零成本

Stopped Autonomous Database:
✓ Compute: Zero cost (stopped)
✗ Storage: $0.025/GB/mo continues
✗ Backups: Retention charges continue

Example: 1 TB ADB stopped for 30 days
Storage: 1000 GB × $0.025 = $25/month (charged!)

Stopped Compute Instance:
✓ Compute: Zero cost
✗ Boot volume: $0.025/GB/mo continues
✗ Block volumes: $0.025/GB/mo continues
✗ Reserved IP (if attached): $7.30/month continues

# Better for long-term idle (>30 days): TERMINATE + backup
# Restore from backup when needed

规则：停止 = 计算暂停，存储仍会计费

❌ 绝不要忽略数据出口费用（可能导致意外账单）

OCI egress pricing:
- First 10 TB/month: FREE
- 10-50 TB: $0.0085/GB
- 50+ TB: Contact sales for discount

# Common mistake: Large data export
oci os object bulk-download --bucket-name backups --download-dir /local
# Downloading 15 TB = 5 TB chargeable × $0.0085/GB = $42,500 surprise!

# Cheaper alternatives:
1. Use OCI FastConnect (fixed cost, unlimited egress)
2. Download within same region (free between OCI services)
3. Export to another OCI region first (inter-region = free)

Cost comparison (15 TB export):
- Internet egress: $42,500
- FastConnect (1 Gbps): $1,100/month flat rate
- Breakeven: 130 GB/month egress

注意事项：OCI 区域之间的数据出口免费（Oracle 内部网络）

❌ 绝不要在不了解费用的情况下启用 NAT Gateway

NAT Gateway pricing:
- Gateway itself: $0.01/hour = $7.30/month
- Data processed: $0.01/GB

# Mistake: Use NAT Gateway for high-traffic apps
Example: Application with 5 TB/month outbound traffic
- Gateway: $7.30/month
- Data: 5000 GB × $0.01 = $50/month
- Total: $57.30/month

# Cheaper alternative: Public IP on instance (ephemeral IP = free)
Cost: $0/month for egress <10 TB

When NAT Gateway makes sense:
- Private subnets with <100 GB/month egress
- Security requirement (no public IPs on instances)

When to avoid:
- High-traffic egress (use public IPs instead)
- Low-security dev/test environments

❌ 绝不要过度承诺 Universal Credits

Universal Credits gotcha:
- Credits are NON-TRANSFERABLE between service categories
  * Compute credits → can only pay for compute
  * Database credits → can only pay for database
  * Cannot move credits between categories

# Mistake: Commit to $10k/month compute credits
# Reality: Only use $6k/month compute
# Cannot use $4k surplus for database spending
# Result: Waste $4k/month ($48k/year)

# RIGHT - commit conservatively based on baseline usage
Analyze 3-6 months historical usage
Commit to 70-80% of baseline (not peak)

Expiration gotcha:
- Monthly credits expire end of month
- Cannot roll over to next month
- Use-it-or-lose-it model

❌ 绝不要相信预算预测（误差率为 30-40%）

OCI Budget types:
1. ACTUAL: Alerts on actual spending (accurate)
2. FORECAST: Alerts on projected spending (often wrong)

# Problem: Forecast uses simple linear projection
Example:
- Week 1: $100 spend
- Forecast for month: $100 × 4 = $400
- Reality: Week 1 included one-time data migration
- Actual month: $150 total
- Forecast error: 167% over-prediction

# WRONG - rely on forecast alerts
Alert at 80% forecast → fires prematurely

# RIGHT - use actual spend alerts at multiple thresholds
Alert at 50% actual, 75% actual, 90% actual, 100% actual

最佳实践：使用 FORECAST 查看趋势，使用 ACTUAL 编制预算

成本优化策略

规格迁移节省（准确计算）

固定规格 → 灵活规格迁移

Legacy fixed shape:
VM.Standard2.4: 4 OCPUs, 60 GB RAM (fixed ratio 1:15)
Cost: $0.06/hr × 4 = $0.24/hr = $175/month

Flex shape (right-sized):
VM.Standard.E4.Flex: 4 OCPUs, 16 GB RAM (custom ratio)
Cost: ($0.02/OCPU-hr × 4) + ($0.0015/GB-hr × 16) = $0.104/hr = $76/month

Savings: $99/month per instance (56% reduction)

Migration path:
1. Create flex instance with same OCPU count
2. Test with minimal RAM (4 GB per OCPU)
3. Increase RAM only if needed
4. Delete fixed instance

AMD → Arm 迁移

AMD instance:
VM.Standard.E4.Flex: 4 OCPUs, 16 GB RAM
Cost: $0.02/OCPU-hr × 4 × 730 = $58.40/month

Arm instance (same workload):
VM.Standard.A1.Flex: 4 OCPUs, 16 GB RAM
Cost: $0.01/OCPU-hr × 4 × 730 = $29.20/month

Savings: $29.20/month per instance (50% reduction)

Gotcha: ARM64 architecture (not all apps compatible)
Check: Docker images available for ARM64?

最大化利用免费层

准确的费用规避（你无需支付的费用）：

Always-Free tier value if fully utilized:

Compute:
- 2 AMD Micro VMs: 2 × $7/month = $14/month
- 4 Arm OCPUs (24 GB): 4 × $7.30/month = $29.20/month
Subtotal: $43.20/month

Database:
- 2 Autonomous Databases: 2 × $292/month = $584/month

Storage:
- 200 GB block: $5/month
- 10 GB object: $0.26/month
- 10 GB archive: $0.02/month
Subtotal: $5.28/month

Networking:
- 1 load balancer: $10/month
- 10 TB egress: $85/month
Subtotal: $95/month

TOTAL COST AVOIDED: $727.48/month ($8,730/year)

注意事项：2 个 ADB 的限制适用于整个租户（不是每个区域/隔间各自计算）

存储生命周期优化

自动分层节省：

Scenario: 10 TB compliance data, accessed quarterly

Without tiering (all Standard):
10,000 GB × $0.0255/GB/mo = $255/month = $3,060/year

With tiering (Archive after 30 days):
Month 1 (Standard): 10,000 GB × $0.0255 = $255
Months 2-12 (Archive): 10,000 GB × $0.0024 × 11 = $264
Total year: $519

Savings: $2,541/year (83% reduction)

Lifecycle policy:
- Day 0-30: Standard (frequent access window)
- Day 31+: Archive (compliance retention)
- Retrieval cost: $0.01/GB (acceptable for quarterly access)

计算资源自动关机

开发/测试环境自动化：

Scenario: 10 development instances, 2 OCPUs each
Running 24/7: 10 × 2 × $0.02/hr × 730 = $292/month

Auto-shutdown (weekdays 9am-6pm only):
Usage: 9 hours/day × 5 days = 45 hours/week = 195 hours/month
Cost: 10 × 2 × $0.02/hr × 195 = $78/month

Savings: $214/month (73% reduction)

Implementation:
1. Tag instances: Environment=Development
2. Create Functions for start/stop
3. Schedule with OCI Events:
   - Start: Weekdays 9am
   - Stop: Weekdays 6pm

Universal Credits 注意事项

额度类别（不可互相转移）：

类别	包含的服务	典型用途
计算	VM、裸金属、容器实例、OKE	应用、工作负载
数据库	ADB、DB Systems、Exadata、MySQL	数据存储、事务处理
存储	块、对象、文件、归档	备份、数据湖
网络	负载均衡器、FastConnect、NAT GW	连接、出口流量

承诺陷阱：

Scenario: Commit to $5k/month database credits
Month 1-3: Use only $3k/month (waste $2k/month)
Cannot: Transfer $2k to compute spending
Result: Waste $6k over 3 months

Better approach:
1. Analyze 6 months baseline usage per category
2. Commit to 70% of baseline (room for variance)
3. Over-commit only if growth guaranteed

过期：月度额度必须当月使用，过期作废（不可结转）

成本分摊挑战

共享资源分摊（OCI 专属）：

Problem: How to allocate costs for shared VCN?

VCN costs:
- DRG: $0.01/hr = $7.30/month
- NAT Gateway: $0.01/hr + $0.01/GB = variable
- FastConnect: $1,100/month (1 Gbps)

Allocation strategies:
1. Equal split: $1,107.30 / N teams
2. Usage-based: Track egress per team (requires tagging)
3. Chargeback: Production pays 100%, dev/test free

Gotcha: OCI doesn't auto-allocate shared resource costs
Must implement manual showback/chargeback process

负载均衡器成本分摊：

Problem: Multiple apps share 1 load balancer

LB costs:
- Flexible LB: $10/month + $0.008/LCU-hour
- Bandwidth: $0.01/GB processed

Allocation approach:
1. Tag backends by team (freeform_tags.Team)
2. Estimate traffic % per team (monitoring metrics)
3. Allocate LB cost proportionally

Example:
- Total LB cost: $50/month
- Team A: 60% traffic → $30/month
- Team B: 40% traffic → $20/month

预算最佳实践

多阈值告警：

# Best practice: Set 4 alert levels
oci budgets alert-rule create \
  --budget-id <ocid> \
  --type ACTUAL \
  --threshold 50 \
  --recipients "[email protected]"

oci budgets alert-rule create \
  --budget-id <ocid> \
  --type ACTUAL \
  --threshold 80 \
  --recipients "[email protected],[email protected]"

oci budgets alert-rule create \
  --budget-id <ocid> \
  --type ACTUAL \
  --threshold 90 \
  --recipients "[email protected],[email protected],[email protected]"

oci budgets alert-rule create \
  --budget-id <ocid> \
  --type ACTUAL \
  --threshold 100 \
  --recipients "[email protected],[email protected],[email protected]"

Alert strategy:
50%: FYI to team (informational)
80%: Action required (investigate spike)
90%: Escalation to management
100%: Finance involved (budget exceeded)

注意事项：预算仅用于告警，不会强制执行（无法阻止支出）

隐藏成本检测

每月审查清单：

# 1. Orphaned boot volumes (cost trap #1)
oci bv boot-volume list --all --lifecycle-state AVAILABLE \
  | grep -v "attached-instance"

# 2. Unattached block volumes
oci bv volume list --all --lifecycle-state AVAILABLE

# 3. Reserved IPs without instance
oci network public-ip list --scope REGION --lifetime RESERVED

# 4. Stopped instances with volumes
oci compute instance list --lifecycle-state STOPPED

# 5. Old snapshots/backups
oci bv backup list --all \
  | jq '.data[] | select(.["time-created"] < "2024-01-01")'

# 6. Unused load balancers (no backends)
oci lb load-balancer list --all

# 7. Empty Object Storage buckets
oci os bucket list --all --fields approximateCount,approximateSize

预计每月节省金额：对于典型租户，清理浪费可节省 $100-500

渐进式加载参考资料

OCI Cost Management CLI

在以下情况下加载 oci-cost-cli.md：

• 设置预算和告警规则
• 通过 CLI 查询使用情况报告
• 管理服务限制和配额
• 实施标记以进行成本分配
• 下载详细的成本报告

以下情况无需加载：

• 快速计算成本（参见上方的表格和示例）
• 了解成本陷阱（上方已完整列出）
• 比较形状定价（上方已涵盖）

Oracle 官方文档来源

主要参考资料（已抓取 200 多个官方来源）：

• 计费和成本管理概述
• 成本报告概述
• 预算概述
• 成本管理最佳实践
• OCI FinOps Hub
• Cloud Advisor（成本优化）

注意：此技能中的所有成本计算、陷阱和注意事项均基于这些官方来源和 A-Team Oracle 博客

---

何时使用此技能

• 意外账单：调查费用，查找成本激增
• 预算规划：估算成本、设置预算、预测支出
• 成本优化：调整资源规格、减少浪费、迁移形状
• 免费套餐：最大限度地使用永久免费资源，避免过度配置
• Universal Credits：制定承诺计划，分配额度
• 成本分配：显示分摊、费用分摊、共享资源成本
• 节省金额计算：比较选项，分析投资回报率