AWS Details ZS’s Internet-Free SageMaker Platform Across 200+ Domains

ZS Associates has turned Amazon SageMaker into its primary ad-hoc analytics environment for most application teams while operating it without direct internet access by default. The newly detailed design spans more than 200 SageMaker domains and tries to resolve a persistent enterprise tradeoff: give analysts usable, self-service tools while retaining controls over code, data access, audit trails and compute spending.

The platform uses a separate SageMaker domain for each tenant. Those domains have isolated Elastic File System volumes, distinct Identity and Access Management roles, and controllable network settings, giving ZS separate boundaries for access control and cost tracking across teams.

Rather than allowing outbound internet access, the deployment relies on Amazon VPC endpoints for controlled connections to required AWS services. Package installation is routed through JFrog Artifactory for pip, Debian and Maven packages; ZS says Artifactory scans packages before they enter the environment. The platform also applies AWS KMS encryption by default to EFS, S3, ECR and CodeCommit resources.

The controls extend beyond the network. CrowdStrike provides operating-system threat detection, Splunk aggregates logs, and CloudTrail records API activity for audit purposes. ZS also uses a three-tier IAM role structure: domain execution roles provide defaults, user roles can override them for finer-grained control, and space execution roles govern shared workspaces.

ZS built Streamlit applications within SageMaker Spaces for a defined set of operations: starting and stopping Redshift clusters, querying S3 data through DuckDB, and provisioning transient EMR clusters from pre-approved configurations. That shifts routine work from the AWS Management Console into governed interfaces, with SageMaker execution roles enforcing access controls.

The company also addressed a durability gap with automated Space backups. Each domain gets a dedicated S3 bucket when created, and the system synchronizes Space data to it every 30 minutes. For interactive compute, lifecycle configurations and IAM policies limit users to smaller approved instance types; higher-end instances require allowlisting, while auto-shutdown releases idle resources.

AWS’s account gives two different daily-user figures. Its introduction and conclusion describe more than 1,000 daily active users, while an operational section cites more than 500 daily users across 200-plus domains. Both figures indicate a production deployment beyond a small pilot, but the article does not reconcile the difference.

ZS is exploring SageMaker Feature Store, Model Registry, Pipelines and Model Monitor as potential next additions. Those services would move the same governance ambition into feature management, model lifecycle controls, workflow orchestration and production oversight; they remain under evaluation rather than deployed capabilities in this account.