Hi Amogh,

On Jarek's thought, I agree with him that just changing the pause behavior
is probably not a good idea.
I can see instances where a user may want to instantly pause a dag in one
case
but want to drain a dag in another case. I myself have such a need for
different dags.

For example, during airflow upgrades, I would like to gracefully drain my
whole instance before upgrade.
While during regular operation, I may want to prevent my dag from launching
further tasks.

The drain here is a transient which eventually converges to pause.

Draining keeps task scheduling enabled, and only automated dagrun creation
paths are gated.
*is_paused* is set to true only when the dagrun are completed. Manual
triggers are still allowed
simialr to the behavior when *is_paused* set to false.

It would be great to get your thoughts on the PR once you get a chance to
play with it.

Dheeraj

On Tue, Sep 8, 2026 at 11:19 PM Amogh Desai <[email protected]> wrote:

> Yep, *is_paused* does more than block new runs. The scheduler checks it
> directly when deciding which
> tasks to run next. That is what freezes downstream tasks today, and it is
> also why I would push back on
> "same is_paused but two buttons."
>
> If you are suggesting that "Pause and drain" sets *is_paused=True* right
> away, those two checks still need a way
> to know "this one is a soft pause, let it keep going." That means adding a
> second flag anyway, just hidden behind
> the same name. And *is_paused* shows up in other places like backfill pause
> logic and asset triggered runs. Each of
> those would need checking to make sure a "soft paused" Dag still acts like
> an active one everywhere except the one
> spot it should not. From the PR, it avoids all that by keeping
> *is_paused=False* during drain, so every existing check
> keeps working as it should. It only adds new checks in a few places that
> decide whether to start a new run.
>
> Thanks & Regards,
> Amogh Desai
>
>
> On Tue, Sep 8, 2026 at 4:29 PM Jarek Potiuk <[email protected]> wrote:
>
> > I like the idea Amogh. I think **just changing** pause behaviour might
> not
> > be good idea for compatibility (someone might rely on it's current
> > behaviour).
> >
> > But maybe we could have different behaviour of two
> > different buttons/behaviour: "Pause" (current behaviour - stop current
> > tasks and it's downstream), "Pause and drain" - set the Dag to "paused"
> but
> > let the running dag_runs to continue until completion.
> >
> > It might however require some scheduler changes - I think currently
> > is_paused is used to select run-eligible tasks ?
> >
> > J.
> >
> >
> > On Tue, Sep 8, 2026 at 7:41 AM Amogh Desai <[email protected]>
> wrote:
> >
> > > Thanks for picking a 4-year-old issue.
> > > The ask itself makes sense.
> > >
> > > One design question that might eventually come up in this thread / PR
> is:
> > > why do we need a new persisted state
> > > instead of just changing what is_paused does? i.e: keep gating new run
> > > creation on is_paused, but stop freezing
> > > downstream tasks in created runs. That would get you graceful draining
> > > without a schema change or migration.
> > >
> > > From my reading, doing that would silently change the semantics of
> > > *is_paused*. Today, freezes downstream tasks in-flight.
> > > Some operators pause specifically to stop a run's progress, not just to
> > > block new ones. Folding drain behavior into *is_paused*
> > > might change that for everyone already relying on the current freeze +
> > > pause behavior, with no option to opt out, yes?
> > > Worth clarifying.
> > >
> > > Thanks & Regards,
> > > Amogh Desai
> > >
> > >
> > > On Sun, Sep 6, 2026 at 3:15 AM Dheeraj Turaga <[email protected]
> >
> > > wrote:
> > >
> > > > Hey Everyone,
> > > >
> > > > I wanted to start this discussion to get feedback on PR #72407, which
> > > > proposes adding a "draining" scheduling state to Airflow.
> > > >
> > > > The Problem
> > > >
> > > > Currently, pausing a DAG stops it mid-flight. While running tasks are
> > > > allowed to finish, queued and downstream tasks remain stranded until
> > the
> > > > DAG is unpaused. There is currently no native way to stop starting
> new
> > > runs
> > > > while allowing those already in flight to complete.
> > > >
> > > > This is a significant pain point during upgrades and maintenance. The
> > > > current workaround as described in the four-year-old Issue #22006
> > > requires
> > > > users to manually rewrite every DAG schedule to None, wait for runs
> to
> > > > drain, and then restore the schedules afterward. This process is
> > invasive
> > > > and prone to error.
> > > >
> > > > Proposed Solution: The draining State
> > > >
> > > > The PR introduces a draining state that sits between active and
> paused.
> > > Key
> > > > behaviors include:
> > > >
> > > >   - No New Scheduled Runs: The scheduler creates no new runs for a
> > > draining
> > > > DAG (including scheduled, asset-triggered, and partitioned/rollup
> > paths).
> > > >   - Completion of In-Flight Runs: is_paused remains false during the
> > > drain,
> > > > meaning task instances in existing runs are still scheduled and
> finish
> > > > normally.
> > > >   - Automatic Convergence: Once no unfinished runs remain, the
> > scheduler
> > > > automatically moves the DAG to the paused state and writes a
> > > > drain_completed audit log entry.
> > > >
> > > > The core property of this feature is that draining is transient, not
> a
> > > > third resting state; it always converges to paused.
> > > >
> > > > Implementation Details
> > > >
> > > > Explicit run creation via manual triggers, TriggerDagRunOperator,
> asset
> > > > materialization, or backfills remains allowed during draining,
> > mirroring
> > > > the behavior of a paused DAG. An earlier revision that blocked these
> > > > actions was reverted to ensure draining is not stricter than the
> state
> > it
> > > > converges into.
> > > >
> > > > Links:
> > > >
> > > >   - PR: https://github.com/apache/airflow/pull/72407
> > > >   - Issue: https://github.com/apache/airflow/issues/22006
> > > >
> > > > Thanks,
> > > > Dheeraj
> > > >
> > >
> >
>

Reply via email to