Performance · Published 26 August 2026
Benchmarking a durable multi-channel publishing scheduler
A reproducible protocol for claim races, jittered retries, idempotency, adapter faults, receipts, and workspace isolation.
Model the job lifecycle
Publish the state machine, lease duration, retry schedule, adapter versions, channel count, payload class, and concurrency. Measure enqueue, claim, adapter call, receipt persistence, and final visibility separately.
One fast happy-path post says nothing about recovery.
Race every transition
Start multiple workers against the same due job, expire a lease mid-call, repeat a provider response, and crash between provider success and receipt persistence.
The outcome must be one logical publication or an explicit reconciliation state, never silent duplication.
Inject provider behaviour
Test rate limits, timeouts, permanent validation errors, credential revocation, partial media upload, and delayed success. Respect retry guidance and cap attempts.
Noisy channels and workspaces must not starve quiet ones.
Verify receipts and budgets
Recompute receipt integrity, trace each external effect to approval and revision, and confirm AI reservations settle exactly once.
Benchmark report and fixture fingerprint should be exportable without production secrets.
Production checklist
- Verify Domain-first ownership against the deployed environment, not a screenshot.
- Verify Receipts for effects against the deployed environment, not a screenshot.
- Verify Transactional AI budget against the deployed environment, not a screenshot.
- Verify Truthful capability state against the deployed environment, not a screenshot.