cloudnative wiki

GCP Cost Optimization, Committed Use Discounts (CUDs), and FinOps

9 min read

GCP Cost Optimization, Committed Use Discounts (CUDs), and FinOps 💰📊

Cost optimization in Google Cloud Platform requires deep engineering alignment between infrastructure architecture and financial mechanics. Unlike traditional static infrastructure, GCP provides automated, non-contractual discounts (Sustained Use Discounts - SUDs), contractual reservations (Committed Use Discounts - CUDs across Resource-Based and Flexible Spend-Based vectors), AI-driven rightsizing recommendations via the Active Assist Recommender API, and real-time granular cost attribution through BigQuery Detailed Billing Exports.

1. Architectural Cost Mechanics & Discount Taxonomy

Understanding how Google Cloud discounts stack and apply across projects, folders, and organizations is critical for preventing financial waste.

                           TOTAL RAW ON-DEMAND CONSUMPTION
                                          │
                                          ▼
     ┌────────────────────────────────────────────────────────────────────────┐
     │ 1. SUSTAINED USE DISCOUNTS (SUDs)                                      │
     │    - Automatically applied to N1, N2, N2D, C2, M1 Compute Engine VMs   │
     │    - Progressive tier discount as usage crosses 25%, 50%, 75% of month │
     │    - Up to 30% automatic savings without contractual commitment        │
     └────────────────────────────────────┬───────────────────────────────────┘
                                          │ Uncovered Baseline Workloads
                                          ▼
     ┌────────────────────────────────────────────────────────────────────────┐
     │ 2. COMMITTED USE DISCOUNTS (CUDs)                                      │
     │                                                                        │
     │  ┌─────────────────────────────────┐ ┌──────────────────────────────┐  │
     │  │   RESOURCE-BASED CUDs           │ │   FLEXIBLE SPEND-BASED CUDs  │  │
     │  │   - Pinned to Region & Machine  │ │   - Hourly spend commitment  │  │
     │  │     Family (e.g., N2 in us-cen1)│ │     (e.g., $100/hr on compute│  │
     │  │   - Up to 57% savings (vCPU/RAM)│ │   - Applies across C3, N2,   │  │
     │  │   - Up to 70% for Memory-Optim. │ │     Cloud Run, GKE Autopilot │  │
     │  └─────────────────────────────────┘ └──────────────────────────────┘  │
     └────────────────────────────────────┬───────────────────────────────────┘
                                          │
                                          ▼
     ┌────────────────────────────────────────────────────────────────────────┐
     │ 3. ACTIVE ASSIST & RECOMMENDER ENGINE                                  │
     │    - Machine learning models analyze 8-30 day CPU/RAM/Disk percentiles │
     │    - Recommends machine shape resizing, idle VM purge, and CUD buys    │
     └────────────────────────────────────┬───────────────────────────────────┘
                                          │ Real-Time Streaming Export
                                          ▼
     ┌────────────────────────────────────────────────────────────────────────┐
     │ 4. FINOPS ATTRIBUTION & BIGQUERY DETAILED BILLING                      │
     │    - Line-item cost breakdown with labels and tags                     │
     │    - Amortized cost attribution across departments & cost centers      │
     └────────────────────────────────────────────────────────────────────────┘

Sustained Use Discounts (SUDs)

  • Mechanics: SUDs are automatic discounts applied to eligible Compute Engine instances and Cloud SQL databases that run for more than 25% of a billing month.
  • Incremental Curve:
    • 0% to 25% of the month: 100% of base on-demand rate.
    • 25% to 50% of the month: 20% discount on the incremental usage.
    • 50% to 75% of the month: 40% discount on the incremental usage.
    • 75% to 100% of the month: 60% discount on the incremental usage.
    • An instance running 24/7 for a full month receives an effective 30% net discount compared to hourly on-demand rates.
  • Machine Family Support: Appears automatically on N1, N2, N2D, C2, and M1/M2 families. Note: Newer machine series like C3, C3D, N4, and GKE Autopilot do not receive SUDs; they rely solely on CUDs.

Committed Use Discounts (CUDs): Resource-Based vs Flexible Spend

DimensionResource-Based CUDsFlexible Spend-Based CUDs
Commitment UnitSpecific amount of vCPU, RAM, GPU, or Local SSDDollar spend commitment per hour (e.g., $50.00/hr)
Term Length1 Year or 3 Years1 Year or 3 Years
Region PortabilityRegionally locked (e.g., must run in us-central1)Globally portable across all GCP regions
Family PortabilityPinned to machine family (e.g., N2 only)Cross-family (N1, N2, N2D, C2, C3, E2)
Service ScopeCompute Engine, Cloud SQL, AlloyDB, SpannerCompute Engine, GKE Autopilot, Cloud Run
Maximum SavingsUp to 57% (1-yr ~37%, 3-yr ~57%)Up to 46% (1-yr ~28%, 3-yr ~46%)
Best Used ForPredictable, static database and core compute tiersRapidly evolving microservice clusters & global apps

2. Granular FinOps Governance & Cost Attribution

Billing Sharing & Scope Hierarchy

By default, a CUD purchased within a billing account can be shared across all projects linked to that billing account (Discount Sharing Enabled).

  • If Project A under-utilizes its committed vCPUs, Project B in the same region running N2 machines automatically consumes the unused commitment, preventing financial slippage.
  • Organizations can disable discount sharing or isolate commitments to specific projects using Billing Subaccounts for strict departmental cost boundaries.

BigQuery Detailed Billing Export

To calculate true cost of ownership (TCO) and unit economics:

  • Standard Usage Cost: Hourly line-item usage.
  • Detailed Usage Cost: Includes resource-level metadata, instance IDs, GKE pod labels, and disk serial numbers.
  • Pricing Export: Historical catalog pricing adjustments.
-- Query Top 10 Cost Drivers with Amortized CUD Discounts
SELECT
  project.name AS project_name,
  service.description AS service_description,
  sku.description AS sku_description,
  ROUND(SUM(cost), 2) AS unblended_cost,
  ROUND(SUM(cost + IFNULL((SELECT SUM(c.amount) FROM UNNEST(credits) c), 0)), 2) AS net_cost
FROM
  `billing_export.gcp_billing_export_resource_v1_01AB23_CD45EF_678901`
WHERE
  _PARTITIONDATE >= DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
GROUP BY
  1, 2, 3
ORDER BY
  net_cost DESC
LIMIT 10;

3. Production CLI Operations & Active Assist (gcloud)

1. Query Active Assist Recommendations for Rightsizing & Idle VMs

# List all VM rightsizing recommendations across a production project
gcloud recommender recommendations list \
    --recommender=google.compute.instance.MachineTypeRecommender \
    --location=us-central1-a \
    --project=core-infrastructure-prod \
    --format="table(name.basename(),content.overview.title,content.operationGroups[0].operations[0].resource,primaryImpact.costProjection.cost.units)"
 
# List idle VM instances (candidates for deletion)
gcloud recommender recommendations list \
    --recommender=google.compute.instance.IdleResourceRecommender \
    --location=us-central1-a \
    --project=core-infrastructure-prod \
    --format="yaml"

2. Inspect Committed Use Discount Recommendations

# Query automated recommendations for CUD purchases based on 30-day historical baseline
gcloud recommender recommendations list \
    --recommender=google.compute.commitment.UsageCommitmentRecommender \
    --location=us-central1 \
    --project=billing-admin-prod \
    --format="yaml"

3. Purchase a Resource-Based CUD for Production Infrastructure

# Purchase a 1-year commitment for 64 vCPUs and 256 GB RAM of N2 in us-central1
gcloud compute commitments create prod-n2-us-central1-1yr \
    --region=us-central1 \
    --plan=TWELVE_MONTH \
    --resources=vcpu=64,memory=262144MB \
    --type=COMPUTE_OPTIMIZED \
    --project=core-infrastructure-prod

4. Enforce Budget Alerts with Cloud Pub/Sub Webhooks

Automate notifications and shutdown triggers when project spend approaches monthly budgets:

# Create a budget with programmatic Pub/Sub notification at 50%, 90%, and 100% thresholds
gcloud billing budgets create \
    --billing-account=01AB23-CD45EF-678901 \
    --display-name="Production Infrastructure Monthly Budget" \
    --budget-amount=25000USD \
    --threshold-rule=percent=0.5 \
    --threshold-rule=percent=0.9 \
    --threshold-rule=percent=1.0,basis=forecasted-spend \
    --notifications-topic=projects/secops-kms-prod/topics/billing-budget-alerts

4. Quotas, Performance, and Configuration Limits

Parameter / DimensionDefault LimitOperational Best Practice
Max CUD Purchases per Day20 commitmentsConsolidate commitments into quarterly or monthly reviews
CUD Term ModificationIrrevocable contractCannot cancel or decrease commitment once finalized
CUD Expiration NoticeActive Assist alert at 30dSet up calendar reminders to prevent sudden cliff drop to on-demand
Budgets per Billing Account5,000 budgetsCreate granular budgets per environment / department
Recommender API Lookback8 to 30 days defaultConfigurable lookback prevents rightsizing based on anomalous dips
Billing Export Latency2 to 6 hoursBigQuery billing export streams micro-batches; not instantaneous

5. Official References & Documentation

6. Realistic Pricing Scenarios

Scenario A: Compute Engine Cluster (On-Demand vs 3-Year Resource CUD)

  • Workload:
    • 50 nodes running n2-standard-8 (400 vCPUs, 1,600 GiB RAM) 24/7 in us-central1.
    • Base On-Demand Price:
      • vCPU: $0.031611 / vCPU-hr.
      • RAM: $0.004237 / GB-hr.
      • Hourly cluster cost: (400 \times 0.031611) + (1{,}600 \times 0.004237) = \12.644 + $6.779 = $19.423/\text{hr}$.
      • Unmitigated Monthly Cost: \19.423 \times 730 \text{ hrs} = \mathbf{$14{,}178.79 / month}$.
  • Applying 3-Year Resource CUD (~57% Discount):
    • Net Hourly cluster cost: \19.423 \times (1 - 0.57) = $8.352/\text{hr}$.
    • Discounted Monthly Cost: \8.352 \times 730 \text{ hrs} = \mathbf{$6{,}096.96 / month}$.
  • Net Annual Savings: (\14{,}178.79 - $6{,}096.96) \times 12 = \mathbf{$96{,}981.96 / year}$ saved.

Scenario B: Dynamic Serverless & Microservice Fleet (Flexible Spend CUD)

  • Workload:
    • Dynamic fleet of Cloud Run microservices, GKE Autopilot workloads, and experimental C3 VMs fluctuating between 200/hr total compute spend across worldwide regions.
    • Safe Baseline Spend: $75.00 / hour continuous baseline.
    • Organization purchases a 3-Year Flexible Spend CUD committing to $75.00 / hour.
  • Savings Analysis:
    • 3-Year Flexible Spend discount delivers ~46% savings on committed eligible compute.
    • Effective hourly cost for $75.00 \times (1 - 0.46) = $40.50/\text{hr}$.
    • Hourly savings on commitment: \75.00 - $40.50 = $34.50/\text{hr}$.
    • Any compute above $75.00/hr bills at standard on-demand / SUD rates.
  • Monthly Savings: \34.50/\text{hr} \times 730 \text{ hrs} = \mathbf{$25{,}185.00 / month}302,220/year).

7. Battle-Tested Nuggets & Production Gotchas

  1. The CUD Commitment Trap (Commit to the Valley, Not the Peak): A Committed Use Discount is a legally binding contract to pay for the committed resources for every single hour of the 1-year or 3-year term, whether you use them or not. If traffic drops or architecture migrates to another architecture (e.g., migrating from N2 VMs to Cloud Run), unused resource-based CUDs continue to bill every month. FinOps golden rule: commit only to 70-80% of your historic minimum baseline trough, leaving fluctuating peaks to on-demand or Spot VMs.
  2. Resource CUDs Do Not Cover Cross-Family Migrations: If you purchase 500 vCPUs of a 3-year N2 Resource CUD in us-central1, and next year your team upgrades all workloads to C3 or ARM-based T2A instances, your N2 commitment will not apply to C3 or T2A. You will pay for the empty N2 CUD plus the new C3 instances. If your infrastructure roadmap anticipates machine family modernization, purchase Flexible Spend CUDs instead of Resource CUDs.
  3. SUDs Disappear on Modern Machine Series: Many engineering teams budget assuming Compute Engine provides automatic ~30% Sustained Use Discounts. However, Google intentionally deprecated SUDs on third-generation and newer machine types (C3, C3D, N4, Z3) and GKE Autopilot. If you migrate from N1/N2 to C3 without purchasing CUDs, your monthly bill may jump significantly because the automatic 30% SUD is absent.
  4. Billing Export Labels Are Not Retroactive: BigQuery detailed billing export only records labels and tags that existed on the resources at the moment the usage occurred. If you add an environment: production or cost-center: 4010 label to a fleet of 200 persistent disks today, you cannot back-query or attribute last month’s costs by that label. Enforce resource labeling at birth via Terraform or Google Cloud Policy (require-labels).
  5. Over-Rightsizing Memory Can Trigger OOM Failures: Active Assist Recommender analyzes historical P95 and P99 memory utilization. If an application utilizes JVM or Python memory pools that occasionally spike during end-of-month financial reconciliation runs, an 8-day recommender window will suggest downsizing RAM. Blindly applying rightsizing recommendations via automated scripts can cause fatal Out-Of-Memory (OOMKilled) pod evictions during unexpected traffic bursts.
  6. Cross-Project Discount Sharing Must Be Actively Verified: In large enterprise organizations with multiple billing subaccounts, verify that Commitment Discount Sharing is explicitly turned on in the Cloud Billing Console. If discount sharing is disabled, a project running at 200% capacity in us-central1 cannot consume surplus CUDs purchased by a sister project in the exact same region, resulting in wasted commitments on one project and full on-demand surcharges on the other.

Across the wiki

Graph View

Backlinks