Positron inherits its Athena and S3 access from the SageMaker Space execution role

What replaces the stored key is a role policy. The Space runs under an execution role, the AWS SDKs in the container read those credentials from the environment, and a query issued from an R or Python session in Positron executes with whatever that role's policy allows ^[2]. Two things have to exist before the first row comes back: an execution role with access to Amazon Athena and the AWS Glue Data Catalog, and a configured Athena query-results location alongside the S3 source ^[7].

The adoption cost sits earlier in the chain. An organisation needs a Posit license grant and access to the Posit-published image definition, administrator permissions to manage ECR and configure custom images for the Studio domain, and an ml.t3.xlarge instance or larger for the environment AWS demonstrated ^[9]. Posit Assistant adds one more line to that list: Bedrock model access in the same AWS Region as the Studio domain ^[8]. On data handling, AWS says customer content is encrypted, is not used to improve base models and is not shared with model providers, and that private connectivity can be configured with AWS PrivateLink ^[13].

The session numbers describe one run on one model. Of the 6,657,942 tokens recorded, 6,118,411 were cache reads and 462,905 were cache writes, leaving 76,626 that touched neither side of the cache ^[6][1]. Against the Assistant's own estimate that works out to about $0.95 per million tokens ^[2], with cache reads outnumbering the uncached remainder by roughly 80 to 1 ^[3]. AWS says cache behaviour and pricing depend on the selected model and provider ^[10].

For that estimate to transfer, a team would need the same model on the same provider pricing and a session whose context is already warm. The stated 92.5 percent cache efficiency does not match either ratio the token counts support: cache reads are 91.9 percent of the recorded total, and 93.0 percent of cache reads plus cache writes ^[4]. The post does not show the denominator.

The demonstration is modest in scope. A synthetic 50,000-loan portfolio sits in S3 and is registered in the Glue Data Catalog, Athena queries it, R validates features, Python trains an XGBoost classifier, Shiny for Python invokes the real-time endpoint, and Quarto records the workflow; AWS says the data does not represent a production lending system ^[11].

One Space carried all of it. The project explorer, editor, R and Python sessions, Variables pane, plots, terminal and application preview ran in a single browser-based environment on SageMaker compute under the Space execution role ^[15]. Teams can also run several Spaces at once for independent projects ^[12] and reserve capacity with SageMaker AI training plans for scheduled training ^[14].