kaxil commented on PR #70184: URL: https://github.com/apache/airflow/pull/70184#issuecomment-5037269704
Thanks for the work here, and for turning the review comments around quickly. I'm going to close this along with #70093. The short version: the triggerer is the wrong home for an LLM/agent call. A trigger's `run()` has to be idempotent and side-effect-free because it re-runs on any triggerer restart or redistribution, and a billed, non-deterministic `agent.run()` with tools executing in-process is neither. `LLMTrigger.run()` here re-issues the whole call from scratch on every relaunch, so a restart mid-flight re-bills it, which is the opposite of the "no re-billed calls" property the issue is after. I've left the full reasoning, and the primitives that do fit (3.2 async tasks for slot-free concurrency, `AgentOperator(durable=True)` for replay-safe long runs), on #70093. Appreciate the effort, and sorry to send it back. This one's an architecture-direction call, not a comment on the code itself. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
