pltbkd commented on PR #1138: URL: https://github.com/apache/flink-agents/pull/1138#issuecomment-6040946018
I have a proposal for the recovery consistency model. Given that: 1. Internal sub-agent recovery can provide the same recovery strength as durable execution under deterministic replay. 2. `ActionState` serves an action-level role analogous to reconciliation: one record is initialized before execution, updated as durable calls finish, and marked completed with the action's memory updates and emitted events when the action finishes. 3. The deterministic attributes of `InternalSubagentCallEvent`, including the sub-agent identity, form the event-identity component of the `ActionState` key. Together with the same business key, sequence number, and action identity, replay can therefore locate the original `ActionState` entry after failover. We do not need to restore child `ActionTask`s. We can retain only unfinished root callers and let them issue the calls again. The regenerated child tasks then reconcile against `ActionState`: an incomplete action reruns and reuses its persisted durable-call results, while a completed action skips its body and replays the framework-visible events needed to reconstruct the remaining action graph. Calls to a deeper-level sub-agent initiated by that completed action can be skipped because their results have already been consumed by the action. Without `ActionState`, the unfinished subtree is recomputed. This makes the recovery strategy clear and simple. I have sent the code to @yunfengzhou-hub for review. Do you think this strategy is feasible? -- 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]
