Ray Event Export: Project Announcement

[can-anyscale]

What is Ray Event Export?

We've heard your feedback about improving observability in Ray to help developers better understand their applications at runtime. We're excited to share that we're working on Ray Event Export, a new observability project we aim to deliver by the end of the year.

Ray Event Export exports structured events (such as task, actor state transitions) from Ray to a user-defined HTTP endpoint periodically. These events can be used to build observability tooling with scalability and performance in mind.

Previously, developers used the State or Export APIs to query this data. Ray Event Export reimagines these systems with a focus on scalability, performance and API stability, offering much better scaling without adding overhead to the Ray runtime.

We’re excited to see what developers will build with this new system and the rich data it provides.

Roadmap

• September: Alpha launch, focusing on Task events and UX+infrastructure.
• October: Add support for Job and Node events.
• November: Add support for Actor events.
• Beyond: Add support for library-specific events (e.g. data, serve, train)

Please leave any comments, questions, or feedback you might have. Thankks!

[can-anyscale]

OneEvent Alpha was shipped with the Ray 2.49 release and includes support for task events. Here is the documentation on how to consume these events: https://docs.ray.io/en/master/ray-observability/user-guides/ray-event-export.html. Thank you.

[can-anyscale]

Related anyscale persistent dashboard built on top of this infra https://www.anyscale.com/blog/ray-task-monitoring-persistent-logs, quoted: "With task persistence on Anyscale, built on Ray's Event Export framework, developers now have a scalable, cost-efficient way to capture and analyze task events via managed dashboards."