Let's say I have a long idempotent pipeline with 20 tasks. It takes an hour to run. It fails on the final quick task due to an intermittent third party API failure which outlasts my retry settings. To resolve this problem, I just need to retry the final task, which uses data gathered from all the prior tasks as inputs. If I could retry just this task, it would be resolved in 20 seconds. However, it sounds like I'd need to rerun the entire pipeline from the top.