Choose the cost question
Assume the service already has an agreed definition of a completed operation. The next decision is which costs belong to delivering those completions. A comparison of two implementations needs the same completion contract and service requirements. FinOps Foundation distinguishes technical resource-efficiency measures from business unit measures. A lower charge for one request is useful information, but it does not establish the cost of the business work that request supports.
For this exercise, compare daily recurring infrastructure cost per completed job. Include dedicated capacity, metered attempts, and an allocated share of supporting infrastructure. Engineering labor and one-time migration work are outside this narrow calculation and must remain visible as exclusions. If the investment decision depends on those costs, include them in a separate comparison over the intended operating horizon before choosing an implementation.
Keep allocation and billing bases visible
Classify costs once so a shared database or network charge is not counted both directly and through an allocation. FinOps Foundation describes fixed, proportional, and proxy-based approaches to shared costs. Choose the rule for each shared service, record who owns it, and apply the same rule to both options. An unexplained residual is an allocation gap; moving it outside the chart does not make it disappear.
Also choose a consistent cost basis. FOCUS 1.2 distinguishes Billed Cost, which supports invoicing, from Effective Cost, which spreads relevant prepaid purchases across the charges they cover. Keep currency and reporting periods aligned. Do not add an upfront commitment payment again after already allocating its amortized cost. For a comparison using estimates, retain the assumed rates and their scope so later observed charges can replace them without changing the formula.
Work through the full daily ledger
Consider two hypothetical options, A and B. Each starts 100,000 distinct jobs during a closed daily window, completes 90,000 under the same acceptance contract, and ends 10,000 unsuccessfully. There is no work carried across the window boundary. A makes 120,000 attempts in total; B makes 180,000. Every attempt is billed by a dependency at the stated rate, including failed attempts. That charge is separate from the capacity and supporting costs. These are invented inputs, not provider prices or measured workloads.
A costs $180 per day for dedicated capacity, $0.002 for each attempt, and $60 for allocated supporting infrastructure. Its 120,000 attempts cost $240, producing a daily total of $480. Dividing by 90,000 completed jobs gives about $0.00533 per completion, or $5.33 per 1,000 completions. The $480 includes the work spent on the 10,000 jobs that never completed.
B costs $240 per day for dedicated capacity, $0.0015 per attempt, and the same $60 allocation. Its 180,000 attempts cost $270, producing a $570 daily total, or about $6.33 per 1,000 completions. B has a 25 percent lower per-attempt rate and an 18.75 percent higher total cost for the same number of completed jobs. All amounts are assumed US dollars.
- Cost scopeChoose attributable expenses
Identify dedicated, metered, and shared costs with one allocation rule and a consistent billing basis.
- All workInclude every charged attempt
Retain retry and unsuccessful-job expenses in the ledger for the selected period or settled cohort.
- Unit calculationRelate the total to completions
Divide the scoped cost by completed operations; keep exclusions, currency, and service requirements visible.
- Decision checkVary capacity and attempt assumptions
Find which changes would alter the comparison and identify the evidence needed to support them.
Explain the difference without hiding failed work
Because each job has one initial attempt, A has 20,000 additional retry attempts and B has 80,000. At the assumed rates, those retries cost $40 and $120 respectively. B saves $50 on the first 100,000 attempts, then spends $80 more on retries and $60 more on dedicated capacity. The net difference is $90. This decomposition identifies which assumptions would have to change for B to become cheaper.
The Amazon Builders' Library explains that retries consume additional server resources and can amplify load during failures. Their billing consequences still depend on the service contract: the example assumes a charged attempt, while a real service might bill duration, bytes, or another resource. Attribute what was consumed. Keep failed and cancelled work in scope when it generated those charges, even though it produced no completed job. Avoid treating a client timeout as evidence that no resource was used.
Test the capacity floor and the break-even point
The dedicated-capacity expense in this model remains payable for the full day. Suppose demand and attempts halve while A retains its $180 capacity and $60 supporting allocation. With 45,000 completions and 60,000 attempts, A costs $360, or $8 per 1,000 completions. That increase follows from the assumed capacity floor. It is not evidence that a request became more expensive to execute. A forecast must state when capacity can actually shrink and what reliability requirement prevents it.
For B to match A's original $480 total while retaining $300 of daily capacity and supporting costs, its attempts must cost at most $180. At $0.0015 each, that permits 120,000 attempts, including 20,000 retries after the initial 100,000. This is a break-even condition under the same 90,000-completion assumption. It does not show that reducing retries would preserve those completions or the service requirements.
Bring the assumptions to the decision review
Retain the source cost records, allocation rule, attempt totals, and completion evidence for the same window. If work crosses reporting boundaries, declare how its costs and outcomes are matched before comparing periods. Separate an observed recurring-cost baseline from a proposed implementation forecast. For this example, the next useful evidence would establish whether B can reduce charged attempts while keeping accepted completions and required service behavior, and whether its capacity expense can change on the decision horizon.
