Engineers from Affirm's data infrastructure team describe how they integrated Temporal as a durable batch execution layer beneath Apache Airflow, without replacing Airflow as the scheduler. The core problem was that Airflow's Kubernetes pod operator conflated infrastructure and application concerns, making errors opaque and retry semantics poor. Their solution, the ephemeral worker controller, uses a Temporal controller workflow to spin up dedicated worker pods for each batch job, proxy the user workflow to those workers, monitor for infrastructure vs. application errors, and scale workers back down on completion. This approach provides structured error reporting, clean reattachment semantics via Temporal's workflow ID conflict policies, type-safe context propagation, and a standardized execution substrate across all batch workloads including Spark jobs — all while preserving Airflow as the familiar control plane for existing teams.