A Media Asset Is a Stateful Object

A media asset is a stateful object, not just a collection of bytes. The bytes may exist, but that does not guarantee the image can be shown, retried, reviewed, or published. Vehicle listings rely heavily on media, but accepting images is simple compared to ensuring the original work is preserved and the publish button remains attached. This post discusses the challenges of maintaining custody across storage, databases, vision workers, derived artifacts, human review, and publication.

The model treats a media asset as a stateful object moving through a bounded vision pipeline. Each expensive computer vision (CV) step must write back to the same asset story, or the product may display an image whose approval came from an earlier step. Intake is proven by an event, not a timestamp. The trigger is a durable row in project_events named capture.media_secured.

Its idempotency key includes the capture session identifier, and the payload contains the project identifier, source asset identifiers, required and missing roles, and a processingEnqueued flag.

Separate endpoints for image work are tempting, but they break when a partial retry occurs. Derived images may exist while quality verdicts are missing, and retries can produce better frames while the listing still points to the old image. A separate status endpoint can check progress by vehicle identifier, but readiness is ultimately determined by the persistent asset graph.

The design uses Postgres to manage six check-constrained stages and states for queued, running, succeeded, failed, cancelled, and declined work. Workers claim rows with a SQL path that uses FOR UPDATE SKIP LOCKED, and stale work can be reclaimed after fifteen minutes. The service role owns the claim path, and the drainer keeps the same shape.

Four lanes - storage, database, worker, and product - must agree for a listing to be ready. Storage proves the existence of signed source bytes and derived media bytes. The database records the capture media secured event and processing job rows. Workers perform matte preflight, studio render, listing generation, and marketing generation. The product displays owner visible readiness, human review queue, and publication gate. The binding between these lanes ensures the listing is only published when all conditions are met.