Hi Djeeraj,

I left a comment in the AIP, but it would be great if you could include
some concrete examples of the workflow this serves. I understand why the
workarounds are subpar and Airflow's limitations, but it's not obvious to
me when I'd want to use this. A concrete workflow you run today would make
the case much stronger.

Thanks,
Constance

On Thu, Jul 30, 2026 at 2:28 PM Dheeraj Turaga <[email protected]>
wrote:

> Dear Airflow Community,
>
> Id like to start a discussion on AIP-115: On-Demand Task Sections
>
>
> https://cwiki.apache.org/confluence/spaces/AIRFLOW/pages/440305022/AIP-115+On-Demand+Task+Sections
>
> TL;DR - AIP-115: On-Demand Task Sections
>
> Some Dags contain work that is expensive, slow, risky, or only
> occasionally needed. Today an author can skip it (branching /
> short-circuit) and lose any first-class way to run it afterward,
> block the whole run on a human decision (HITL approval), or tell
> users to clear the right tasks by hand. GitLab's manual pipeline
> jobs are the closest existing analogue; Airflow has no Dag-level
> equivalent.
>
> AIP-115 proposes a ManualGateOperator that lets an author mark a
> section of a Dag as manual-only
>
> Scheduled runs skip past the gate and complete normally - nothing
> waits on a human. The gated tasks are shown distinctly in the grid
> and graph, and the gate offers a "Run manual section" action.
> Triggering it runs that section for the specific Dag run you chose,
> through the normal scheduler and executor path, with the usual logs,
> retries, XComs, callbacks, and trigger-rule behavior. There's an
> equivalent API endpoint for programmatic use.
>
> The result is an optional branch of a pipeline that's discoverable
> from the Dag definition, visible in the UI, auditable, and runnable
> per-run on demand - rather than a convention users have to know
> about and reproduce by clearing tasks manually. The change is
> additive: existing skipping, branching, and HITL behavior are
> unchanged.
>
> The proposal covers the section-boundary semantics, task states, API
> shape, and alternatives considered in detail
>
> ------
> Would love to hear your thoughts & feedback
>
> Regards,
> Dheeraj
>

Reply via email to