Amazon SageMaker training and processing jobs can now lazy-load container images instead of waiting for the entire image to download. The service uses Seekable OCI, or SOCI, indexes to fetch files needed for startup first and stream the remaining layers in the background.
AWS says its tests reduced GPU idle time during image pulls by as much as 90 percent. Across eight compressed images ranging from 1.8 GB to 35.4 GB, indexed containers became ready in roughly three to 11 seconds. A 13.2 GB image fell from 3 minutes 42 seconds to about six seconds, while a 35.4 GB image fell from 7 minutes 17 seconds to the same single-digit range.
A SOCI index is a separate artifact stored beside the image in Amazon ECR. It maps files within each layer without changing the original image digest or invalidating an existing signature. SageMaker checks for the index automatically. If none exists, it downloads normally; if lazy loading fails, the job falls back to the standard path rather than stopping.
Teams can select pre-indexed AWS Deep Learning Containers or add indexes to their own images without changing the job definition, entry point or data channels. The largest gains will come from big images and short or frequently repeated jobs. Actual savings will vary with image composition, network conditions and which files an application touches immediately after startup.