Amazon SageMaker

You can now submit a prioritized list of instance types and counts for training and processing jobs, and SageMaker automatically launches on the first available configuration instead of requiring a single named instance type. This removes the need to build custom retry logic or submit multiple concurrent jobs with different instance types to secure capacity during high-demand periods.

Matters for: Teams training or fine-tuning models on high-demand GPU instances during peak capacity periods.

SageMaker Feature Store now lets data scientists update one or more individual features in a record via a single request, instead of rewriting the entire record, which lowers write latency and cost and removes the need for read-modify-write or merge logic in ingestion pipelines. This also allows multiple pipelines, such as a streaming job and a nightly batch job, to update the same record independently by each writing only the features they compute.

Matters for: Data teams running multiple concurrent pipelines, such as streaming and batch jobs, that write to the same SageMaker Feature Store feature group.

Amazon SageMaker Feature Store now supports the UpdateRecord API, letting applications update one or more feature values in a single atomic call without reading or rewriting the entire record. This eliminates the previous read-modify-write cycle using GetRecord and PutRecord, reducing latency, extra read capacity costs, and race conditions when multiple pipelines update different features of the same record concurrently.

Matters for: Teams running fraud-scoring or other real-time ML pipelines that concurrently update shared feature groups from multiple sources such as clickstream, purchases, and scoring services.