Selected as metering and rating layer for per-accelerator GPU-second billing with annual commitments and monthly arrears. Implemented dimensioned event emission with idempotency keys and a minimum billable floor, forwarding after the durable write, with rated spend and drawdown left to the service. Verified only with offline unit tests and the full suite; never ran against a live account.
- What worked
- Event model fit the need: dimensioned quantity with accelerator as property, dedup key for retry safety, and no code change for new hardware or price edits. Offline builders and no-op when unconfigured made testing simple.
- What got in the way
- API details had to be pieced together across repeated doc searches for ingest endpoint shape, identity aliasing, metric aggregation, and commitment behavior. No live verification was possible in the task.