← all checklists

Cost at Scale

19 items · source

Copy as markdown

Filter items

Unit economics

• Calculate the infrastructure cost of one request, one active user, and one tenant.
• Verify which costs scale linearly with usage and which scale worse than linearly.
• Identify the query, job or endpoint with the worst cost per invocation.
• Verify a feature's cost is attributable, so an expensive one can be found without guessing.
• Verify the cost of your largest tenant is known, and that the pricing covers it.
• Verify free-tier or trial usage has a ceiling, in resource terms as well as in time.

Where the bill actually comes from

• Verify egress bandwidth is measured; it is frequently larger than compute and rarely modelled.
• Verify log and metric volume is measured — observability spend routinely overtakes the infrastructure it observes.
• Verify per-call third-party costs are metered per user: model tokens, SMS, email, geocoding, image processing.
• Verify storage growth is projected, including backups, versions and anything never deleted.
• Verify idle cost is known separately from peak cost, so autoscaling savings are not imaginary.
• Verify a query that is cheap at a thousand rows has been costed at ten million.
• Verify data transfer between regions and between services is accounted for.

Controls

• Verify a budget alarm exists on rate of change, not only on a monthly total.
• Verify autoscaling has a maximum, so an availability incident cannot become a billing incident.
• Verify per-user and per-tenant quotas exist on anything that costs money per call.
• Verify the cheapest fix has been considered before the expensive one — a limit, a cache or a smaller payload usually beats more capacity.
• Verify committed or reserved capacity has been considered once the baseline is predictable.
• Verify someone owns the bill and looks at it monthly, with the ability to explain a change.