Design the pipeline that produces still images for uploaded videos: a poster frame, a set of thumbnail resolutions for catalog and player UI, and a scrub-preview sprite sheet used when a viewer hovers or drags the seek bar. It runs whenever a new video is available and must be reliable, idempotent, and cheap at scale.

Operating context. Hundreds of thousands of videos land each day, from short clips to multi-hour films. For each, the pipeline extracts frames at a fixed sampling policy, generates several image sizes plus a sprite sheet with a WebVTT index, stores them, and exposes stable URLs. Triggers can be delivered more than once, videos can be re-encoded and re-processed, and workers can fail; none of that may create duplicate images or leave a video stuck.

Out of scope. The video transcode itself, ML-based ranking to pick the most engaging thumbnail (assume a fixed sampling policy), the image CDN's cache configuration, and the upload service. Assume separate teams own those.

What to produce. A high-level architecture covering: how the pipeline is triggered, frame extraction and the sampling policy, generation of multiple resolutions and the sprite sheet, how work is made idempotent and de-duplicated, storage and serving of the images, and how retries, dead-lettering, and per-video status are tracked. Sketch the components and the flow for one video; we will probe specifics during checkpoints.