Hi all, I've been reading the PR (flyte#5996) and noticed something about the output behavior of subtasks within

map_task

that seems counterintuitive: Current behavior: • Each subtask returns a List containing exactly one element •

map_task

aggregates these single-element lists into a final List of multiple elements My expected behavior: • Each subtask should return its output directly (not wrapped in a List) •

map_task

would collect these individual outputs into the final List I'm trying to understand: 1. Why was array_node designed this way initially? 2. Would returning direct values from subtasks (instead of single-element lists) be a better design? I believe the current design may be why this PR (flyte#5996) needed a workaround solution. Would changing the subtask output behavior to return direct values instead of single-element lists be a better approach?

this is just a display issue?

If that is only a display issue, I believe that the issue this Pull Request (#5996) addressed shouldn't exist. I am not familiar with the backend code here, so I need some time to verify that.

isn’t that only for offloaded literals? that’s not for all lists right?

but if the subtask's output spec is a scalar instead of a collection, the offloaded literals' type should be correct

but there are no offloaded literals right?

if the subtask of map_task returns a large value, it will be offloaded

yeah but i thought you were testing just with small values

@witty-action-93503, thanks for the investigation. You uncovered a bug in how we handle offloading of literals in the case of map tasks. In https://github.com/flyteorg/flytekit/pull/2872 we get the literal type for outputs from the task interface, but since the interface in the case of map tasks is

Collection[OutputType]

, instead of

OutputType

, we end up setting

inferredType

in that case to the wrong type. I'll follow up on this today.

thank you for the follow-up!