Databricks Connects AWS HealthLake Data Through OpenSharing
Databricks released a customer-deployed reference integration that presents AWS HealthLake FHIR tables as read-only shared catalogs, with Databricks compute reading the underlying files directly from Amazon S3.
Databricks released a reference integration connecting AWS HealthLake-managed FHIR data to its platform through OpenSharing. The customer-deployed implementation presents HealthLake tables as read-only shared catalogs for Unity Catalog governance, SQL and notebook queries, dashboards, and Genie exploration.
This is not a built-in HealthLake or Unity Catalog connector. Customers deploy a Python OpenSharing server as a Databricks App with a Databricks Asset Bundle, then configure a provider and catalog. The design exposes data held in HealthLake without first copying the clinical files into Databricks.
AWS HealthLake converts imported, nested FHIR JSON into Apache Iceberg tables in Amazon S3, with a separate table for each FHIR resource type. For each table request, the app obtains temporary Lake Formation credentials, resolves the current Iceberg snapshot, creates presigned URLs for its Parquet files, and returns an OpenSharing response. Databricks compute downloads the files directly from S3, keeping them out of the app server’s data path.
The identity path spans both platforms. Unity Catalog calls the app with OAuth client credentials, with calls restricted to a configured recipient service principal. The app uses its Databricks identity and OIDC federation to assume an AWS IAM role. AWS documents that Lake Formation can issue temporary credentials to third-party query engines through its application-integration authorization path.
That architecture also sets the security boundary. The app’s AWS role has SELECT access to every table it serves, while HealthLake’s Lake Formation grants are full-table in this path. The reference repository says customers must apply grants, views, row filters, and column masks in Unity Catalog for protected health information controls. Because presigned URLs bypass Unity Catalog grants, the implementation limits direct app access to the configured recipient service principal.
The implementation does not support time travel, change data feed, predicate pushdown, or streaming reads. Without predicate pushdown, each query reads all files in the requested table. HealthLake analytics tables are append-only: updates and deletions add rows, so matching the current FHIR API state requires selecting the newest row for each resource ID and excluding delete markers.
Databricks says the integration can make analysis easier for less-technical users and help teams reach clinical insights faster. The opened evidence does not independently measure those outcomes. The company also says new FHIR writes appeared once HealthLake’s analytics pipeline committed a new Iceberg snapshot, a process that took four to six minutes in its tests. AWS publishes no target for that interval, so the result is not a service-level commitment.
The repository uses Databricks’ DB license, which limits use and redistribution to work connected with Databricks services and provides the implementation as-is.
More news

AWS documents SageMaker AI data preparation workflows

Snowflake opens live Iceberg data to external REST Catalog clients

Google DeepMind opens PARDIG and SatWeave weather datasets
