Skip to content

Conversation

xintongsong
Copy link
Contributor

The tags in the announcement are reminders for myself to resolve before releasing.


Building on Flink's battle-tested streaming engine, Apache Flink Agents inherits distributed, at-scale, fault-tolerant structured data processing and mature state management, and adds first-class abstractions for Agentic AI building blocks and functionalities -- large language models (LLMs), prompts, tools memory, dynamic orchestration, observability, and more.

The key features of Apache Flink Agents include:
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should mention replayability here.

Replayability for Testing and Safety: Supports replay of historical event streams, enabling regression testing, debugging, drift detection, and safe evaluation of agents without impacting live systems.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure about this one.

It depends on what do we mean by "replay the historical event streams".

  • If we are replaying from a Flink's checkpoint, then it can hardly be used for the mentioned testing / debugging / evaluation purposes. Because with durable execution, we will skip all the actions (model invocation, tool use, etc.) that are previously executed and directly reuse their outcomes / results. The models / tools won't be actually called during the replay.
  • If we are not replaying from a Flink's checkpoint, then it's no different from using any other agent framework, as long as it reads from a replayable input like kafka.

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Flink Agents offer this Replay-based Per Action State Consistency feature(implemented by @letaoj) which is related to replayability I think, is this what @thefalc suggested to call out?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@alnzng, that one is mentioned as "Exactly-Once Action Consistency". Here I didn't mention "replay-based", because I want our users to focus on WHAT capability (the consistency) we provide, rather than HOW (replay-based) we deliver that capability. And as I said, that seems to be conflict against the testing / debugging / evaluation purposes that @thefalc mentioned.

@xintongsong
Copy link
Contributor Author

@thefalc, thanks for help refining the words. I addressed most of your comments, except the replayability one.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants