The largest item in the recorded cost review was D1 reads. The important question was not just how often a page opened, but how much data the server examined behind it. Containers, AI calls and storage were separate workloads.
An initial report included scheduled cancellations. A later email-restoration report explicitly retained Workers Paid and superseded that earlier conclusion. I preserve that sequence instead of describing a historical plan as today's account state.
The cost explanation I want to give
Alongside a lower bill, I want to explain which functions still run. A stopped content update or narrower search is a meaningful change. I document retained and paused features together.
Read usage as work
Separate database queries, collection jobs, AI generation, storage and status polling. Identify who starts each task and how frequently it runs.
D1 exposes rows read separately from returned results. Worker CPU time is also different from the total time a visitor waits. Confusing these measures sends optimization toward the wrong bottleneck.
My measurement procedure
- Map each billed item to a feature. Do not present one account's old bill as a universal price.
- Separate ordinary-traffic observation from audit-generated traffic.
- Record dataset size, cache state, invocation count and rows examined.
- Classify automation as essential, on-demand or unnecessary.
- Describe any reduced search scope, result limit or disabled automatic feature.
What a sample cannot prove
The September record specified a clean observation window without an extra full audit. That is a useful measurement condition, not a guarantee about every future day.
A starting workload estimate is ordinary requests times work per request, plus scheduled executions times work per execution. Check bursts and cache misses separately, and recalculate when the workload changes.
Explain the trade-off another operator inherits
Record which features remain and which stop. Service subscription, domain renewal and data retention are different decisions. A cancellation request in an old report is not proof of today's account state.
A valuable low-cost guide explains the feature boundary and observation conditions that made the result possible.