Layer, an AI creative platform for game studios, shares how they scaled from a single FileUploadWorkflow to 50 Temporal workflow types in production. The post covers their visual DAG pipeline builder (where each node is a Temporal child workflow), a unified RunInferenceWorkflow handling images, video, audio, and 3D with modality-aware timeouts and a webhook+polling hybrid, batch fan-out with custom flow control, task queue isolation to prevent noisy-neighbor problems, observability via a LoggingContextInterceptor plus OpenTelemetry, and a hard-learned lesson about using asyncio.shield() to protect cleanup activities from cancellation. They also discuss their MCP connector enabling LLM agents to invoke Layer workflows autonomously.
Table of contents
The year everything got more interesting #Then video showed up and broke all our assumptions #The last human in the loop #50 workflows and counting #The one where we built a visual pipeline builder on top of Temporal #One workflow, four modalities, zero redesigns #50 ad creatives and one short window #Meet the Sheriff #The good and the "oh yeah, we set alerts for that" #Don't skip asyncio.shield() #What's next #39 Impressions