Amazon SageMaker HyperPod has added Data Capture, a feature that records live inference requests and responses to Amazon S3. The goal is to help teams find drift, recurring failures and new inputs that were absent from the original training set, then use selected records to improve a model.
AWS provides three capture tiers. Teams can store full payloads for monitoring, collect traffic metadata without content, or retain payloads intended for later analysis and training. The tiers can be enabled independently, although turning on both full-payload options duplicates each record.
A published walkthrough uses a synthetic appliance-support router. It captures traffic, queries the records with Amazon Athena, identifies a new-product blind spot, labels examples with a SQL policy, fine-tunes the model on a second GPU and deploys the updated version for comparison. AWS makes the scripts, manifests and expected output available, but the demonstration uses synthetic data rather than a measured customer deployment.
Production capture creates a separate risk: prompts and responses may contain names, addresses, order numbers or confidential material. AWS recommends sampling, encryption with customer-managed keys, lifecycle rules and personal-data redaction. Teams must also configure the correct storage and service-account permissions; an endpoint can remain healthy even if a permissions error silently prevents captures. The useful feedback loop depends as much on data governance and validation as on the capture mechanism itself.