Why a context engine needs Durable Execution: How Unblocked rebuilt its ingestion backbone on Temporal

This post details Unblocked's migration of their AI context engine's ingestion backbone from a DIY actor model to Temporal's Durable Execution. Key technical contributions include: collapsing pages of orchestration code to lines of Temporal Workflow code using a thin Kotlin shim over the Java SDK, minimal abstraction philosophy exposing Temporal primitives, opinionated retry and exception handling tuned to ingestion patterns, and the adoption of Temporal for various data sources (Asana, CI, Slack). The post highlights the benefits of deployment-free long-running jobs, proper scheduling of background jobs, unified handling of user-triggered and cron flows, and built-in observability via the Temporal UI. It also provides production deployment details: self-hosted on Kubernetes with 12 History and 12 Matching replicas, Aurora Serverless persistence, and KEDA autoscaling per queue. The migration is ongoing, with code ingestion transitioning from Prefect.