Announcing instance preference lists for Amazon SageMaker AI training jobs
What Changed
Amazon SageMaker now supports instance preference lists for training and processing jobs, allowing teams to submit a single job with up to five ordered GPU instance types. The scheduler evaluates the list in priority order, launching the first type with available capacity and automatically falling back to on-demand if the preferred type is unavailable. This feature eliminates manual retry loops and reduces wait times during peak demand, improving capacity utilization and freeing engineering focus for model development.
Why It Matters
Enterprise architects can expect lower operational overhead and faster model iteration cycles, as jobs start more quickly and reliably. Cost control may improve by better matching reserved capacity plans with on-demand fallback, while governance can be simplified by reducing custom retry scripts and monitoring. Reliability gains come from automated fallback and bounded retry windows, reducing downtime for time‑critical workloads.
The Limitation
The feature supports only up to five instance types and is limited to GPU‑accelerated instances; non‑GPU workloads or larger preference lists are not covered.
What You Can Do
Implement instance preference lists in your SageMaker job definitions, ordering preferred instance types and reserving capacity plans first to optimize cost and availability.