1.5 Cost Optimization & Visibility
Official task mapping
SAP-C02 task 1.5: Billing tools and purchasing knowledge; allocation and cost/performance trade-offs. This lesson applies the official task to an original design exercise; it is not an exhaustive exam syllabus.
Prerequisites
Review cost learning path, allocation tags and Cost Explorer. Be able to explain the relevant failure boundary and access/cost model before choosing a service.
Visibility architecture
Use linked accounts and activated tags/Cost Categories for ownership. Resource tags are not automatically active billing dimensions. Historical backfill requires historical tag presence. Shared charges need an explicit allocation rule.
Cost Explorer is delayed billing analysis: standard history is the current month plus 13 prior months, with separately configured extended history and forecasts up to 12 months. Budgets notify on objectives; Anomaly Detection identifies unexpected patterns. Neither is a real-time hard spending cap.
Compute Optimizer supports multiple current resource types with resource-specific requirements. EC2 recommendations need at least 30 hours of recent data within 14 days, not a universal two-week wait. Guest memory needs a supported metric integration; detailed EC2 monitoring does not supply it. Savings estimation mode can incorporate account-specific discounts.
Visibility and purchasing revision tables
| Need | Tool/view | Limitation |
|---|---|---|
| Historical trends | Cost Explorer | Data latency and selected metric |
| Detailed analytics | Data Exports/CUR + Athena | Schema, query costs and revisions |
| Fixed target | Budgets | Delayed notification, supported actions only |
| Unexpected pattern | CAD | History/model/threshold dependent |
| Resource sizing | Compute Optimizer and metrics | Resource-specific evidence; recommendations need validation |
| Storage overview | S3 Storage Lens | Aggregated insight, not every object’s access log |
| Purchase | Flexibility | Reason to reject |
|---|---|---|
| Standard RI | Eligible modification/size rules | Incompatible future family/platform |
| Convertible RI | Permitted value-constrained exchanges | Not Marketplace-sellable |
| Compute SP | Eligible compute family/Region/service mix | Unused hourly floor or capacity requirement |
| EC2 Instance SP | Selected family/Region | Planned cross-family migration |
| Database SP | Eligible database usage, one-year plan | Ineligible generations/features |
| Spot | Interruptible available pools | Unrecoverable state or firm capacity requirement |
No discount ceiling proves a selected offer is cheaper. For a hypothetical us-east-1 study on 2026-10-12, compare 720 eligible hours at an assumed $1/hour with an assumed $0.70/hour commitment: $720 versus $504 if fully used. If only 400 hours run and no other usage consumes the benefit, On-Demand is $400 while the fixed commitment remains $504. The purchase decision changes even though its nominal hourly rate is lower.
Collect memory through a supported agent/integration, include metric cost and validate peaks. Link purchase recommendations to the owner’s future workload plan rather than treating historical utilization as complete future knowledge.
Scenario and rationale
A platform bill rises 25%, but product owners disagree about who caused it. First split by account, service and usage type; distinguish an upfront reservation purchase from recurring usage using the appropriate metric. Activate stable business tags for future allocation and assess eligible backfill. Examine traffic and workload metrics before downsizing.
Reject a blanket 25% fleet reduction: it can breach SLOs without explaining the charge. Reject immediately buying more commitments: demand may disappear after rightsizing or migration. A monthly average cannot reveal unused nightly hourly commitments.
Practice drill
Use the synthetic billing extract below, or an authorized equivalent, to identify the largest changed dimensions and reconcile unblended versus amortized totals. Propose an allocation rule for shared networking and a budget/anomaly routing scheme. Produce one rightsizing hypothesis with peak-load evidence, acceptance criteria and rollback; do not buy or change capacity in this exercise.
Self-study billing fixture
These hypothetical us-east-1 monthly amounts are study inputs dated 2026-10-12, not tariffs or a CUR schema. The new reservation is a 12-month, all-upfront $1,200 purchase starting on the first day of the current month; this simplified fixture amortizes $100/month. Compute rows are the remaining usage charges, not a second copy of covered reservation usage. Taxes, credits, support and refunds are excluded.
| Owner/service | Prior unblended | Current unblended | Prior amortized | Current amortized |
|---|---|---|---|---|
| Product A / compute | $800 | $850 | $800 | $850 |
| Product B / compute | $500 | $550 | $500 | $550 |
| Shared / networking | $200 | $300 | $200 | $300 |
| Product A / reservation fee | $0 | $1,200 | $0 | $100 |
| Total | $1,500 | $2,900 | $1,500 | $1,800 |
Expected checks: unblended rises $1,400 (93.33%), dominated by the upfront purchase; amortized rises $300 (20%), comprising $100 compute, $100 networking and $100 amortized reservation. Neither percentage alone demonstrates a traffic spike. Allocate networking by measured usage, not by the upfront payment: assumed A:B network bytes of 2:1 assign $200/$100 of the current shared charge. Current amortized owner totals then become A $1,150 and B $650, still $1,800 overall.
For the rightsizing hypothesis, assume A’s four nodes have peak CPU 35%, memory 85% and p95 latency 120 ms against a 150-ms objective; B’s peak CPU is 90% and p95 145 ms. Reject a blanket CPU-based reduction. Request A’s per-node memory/cache and burst evidence, then propose a reversible canary only if headroom is demonstrated. Alert the product owners for their allocations and the platform owner for shared networking; keep purchase review separate from usage-anomaly triage.
References
- Official Domain 1 task guide
- Cost Explorer history and forecast
- Compute Optimizer requirements
- Tag backfill
Return to Domain 1. The drill is proposed study work, not a claim of a completed deployment.
Across the wiki
- GCP Cost Optimization, Committed Use Discounts (CUDs), and FinOps — cost management (GCP)
- Cost Optimization — cost management (Kubernetes)