What it is
Eventstream is the visual stream-processing canvas in Fabric. Drop a source on the left, a destination on the right, transformations in the middle — manage filters, joins, aggregations, and unions without writing code. Under the hood it's a managed stream-processing service that scales with Fabric capacity.
Supported sources
- Azure Event Hubs (the default for app telemetry)
- Apache Kafka — Confluent, MSK, self-managed
- Azure IoT Hub for device data
- Change-data-capture from Azure SQL, Cosmos DB (NoSQL/Mongo), and PostgreSQL
- Custom endpoints (HTTPS / AMQP / MQTT)
- Fabric sample data — useful for testing without real source wiring
Inline transformations
The visual designer covers the operations you'd want for routing and light shaping: filter, join, group-by/aggregate, manage fields, union. For anything heavier (windowed analytics, ML scoring), let the events land and use KQL update policies or a Spark notebook on the materialized table.
Destinations
- Eventhouse / KQL DB for time-series storage and KQL analytics
- Lakehouse for batch-friendly analytics and ML feature stores
- Activator for detect-and-act on streams
- Derived stream — another Eventstream consumes this one (composition)
- Custom endpoint for external systems
Best practices
- Keep transformations minimal. Heavy lifting belongs downstream.
- Use derived streams for branching destinations. One source → many destinations via composition.
- Always include the source timestamp and a Fabric ingest timestamp — they'll disagree under load.
- Monitor the metrics tab. Throttling and lag are your early warnings.
Common pitfalls
!
Doing complex joins in the designer. Some joins look easy in the UI but melt under realistic event rates. Move to Spark structured streaming if you need windowed event-to-event joins.