yunfengzhou-hub commented on PR #1199: URL: https://github.com/apache/flink-agents/pull/1199#issuecomment-6413263209
> Hi @yunfengzhou-hub, thanks for the heads-up. > > After rebasing, I see that #1138 landed with a different internal-call model, and #1220 then moved internal sub-agent execution onto `SubgraphInvocation` (removing `InternalSubagentCallStatus`). So this PR can't be rebased mechanically: the admission and queueing logic needs to be rebuilt on the new invocation model. > > Instead of one ~4.9k-line PR, I'd like to split #1140 into smaller PRs: > > 1. Bounded admission for internal sub-agent invocations in the Java runtime: a concurrency option plus queueing/backpressure on top of `SubgraphInvocation`, with unit tests. > 2. Durable state for queued-but-not-started calls, and recovery of an in-flight batch. > 3. Python runtime parity. > 4. E2E tests and the Sub-Agents guide. > > Two questions before I start: > > * Should the concurrency bound apply only to internal sub-agent invocations, or to all `SubgraphInvocation`s, including the event-backed chat calls from [[api][runtime][java][python] Add event-backed chat calls from actions #1220](https://github.com/apache/flink-agents/pull/1220)? > * Should the bound be per action execution or per operator subtask? > > I'll convert this PR to a draft and keep it for reference until the first split PR is up. Hi @dangzitou, Flink Agents already has configurations that present an upper bound to the concurrency of sub-agent invocations, like num-async-threads or async.batch.parallelism. The "bounded parallelism" requirement of #1140 can be satisfied by these existing limits for now, without introducing a new upper bound — which also means neither of the two questions (scope and granularity) needs to be answered at this point. What do you think? And if we could set aside the code related to this upper bound, will this PR still be so big that it needs to be divided into multiple ones? -- 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]
