Question

A media startup runs an on-demand video transcoding service. Users upload source files, and the system must transcode them into multiple resolutions. Jobs are stateless, each takes 2-15 minutes, and demand is highly unpredictable—sometimes zero jobs for hours, then hundreds arriving in minutes. The CTO wants to minimize idle cost while ensuring jobs start processing within seconds of arrival and complete reliably even if the platform reclaims capacity. Which compute design best meets these business and technical requirements?

Accepted Answer

Run transcoding on a managed instance group of Spot VMs with autoscaling driven by a Pub/Sub queue depth metric, and design workers to checkpoint and re-enqueue interrupted jobs.

Answer

Deploy transcoding as containers on a GKE cluster with a fixed pool of on-demand nodes sized for peak load, using the Horizontal Pod Autoscaler.

Answer

Use a single large Compute Engine instance kept always-on to process the queue sequentially as jobs arrive.

Answer

Trigger a Cloud Function per uploaded file to perform the transcoding synchronously within the request.