kaxil commented on code in PR #73160:
URL: https://github.com/apache/airflow/pull/73160#discussion_r4043042013
##########
airflow-core/docs/core-concepts/resumable-tasks.rst:
##########
@@ -226,3 +226,42 @@ Comparison
- Airflow 2.2
- Airflow 3.3
- Airflow 3.2
+
+.. _concepts-resumable-tasks-retry-policies:
+
+Retry policies and durable execution
+-------------------------------------
+
+Durable execution, taken broadly, is fine-grained workflow control: rich
+primitives for retries, backoff, and fallbacks, configurable at the level
+of a single task, combined with state that survives a worker crash so a
+workflow can span days without losing progress. Airflow provides two
+primitives for this.
+
+A **retry policy** is the primitive for retries, backoff, and fallbacks. It
Review Comment:
"Fallbacks" isn't really a base RetryPolicy concept -- the base API is just
RETRY/FAIL/DEFAULT. It's specifically LLMRetryPolicy(fallback_rules=...), used
when the LLM classification call itself fails. Might be worth scoping this to
say the LLM-driven policy adds fallback support, rather than implying it's a
general retry-policy primitive.
##########
airflow-core/docs/core-concepts/resumable-tasks.rst:
##########
@@ -226,3 +226,42 @@ Comparison
- Airflow 2.2
- Airflow 3.3
- Airflow 3.2
+
+.. _concepts-resumable-tasks-retry-policies:
+
+Retry policies and durable execution
+-------------------------------------
+
+Durable execution, taken broadly, is fine-grained workflow control: rich
Review Comment:
This opens by folding retry policy into the definition of durable execution
("taken broadly... Airflow provides two primitives for this"), but the page
defines durable execution narrowly just above: "Durable execution is the
outcome; the task state store is how you get it." Every operator's "Durable
execution" section and task-state-store.rst use that same narrow meaning, and
common.ai's agent.rst frames it the other way around: "Durable execution only
helps when the task has retries configured" -- retries are a precondition, not
a component. Given the whole point of this section is that the two are
independent, it'd help to keep "durable execution" scoped to the state-store
meaning here too, e.g. "Long-running tasks typically need two separate
primitives together: a retry policy... and durable execution (the task state
store, as defined above)."
--
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]
