GitHub user wolvery created a discussion: DISCUSS Durable executor launch 
records + typed 409 for stale /run attempts (PR #70931, refs #69760)

Hi all,

I'd like the list's input on which of two approaches to pursue for #69760
before either is hardened for review. I've prototyped both:

Minimal (KubernetesExecutor, equality check):
https://github.com/apache/airflow/pull/69782
Durable redesign (all executors, RFC/Draft):
https://github.com/apache/airflow/pull/70931
Issue: #69760

## Problem

When a task instance transitions out of its original launch (scheduler
failover, pod adoption, stuck-queued reschedule, orphan reset, or a manual
clear), the executor originally assigned may still call the Execution API
`/run` for that task. Today the launch identity lives only on the mutable
`TaskInstance` row, so a late/stale launch surfaces as an opaque
`404 not_found` — indistinguishable from a genuinely unknown task. That
consumes a retry and/or blocks clean requeueing.

## Two candidate designs

Both share the same goal — reject a stale/superseded launch at `/run` — but
differ in where the launch identity lives and how far the change reaches.

### Option A — reuse `external_executor_id`, equality check (PR #69782)

- No schema change. Reuses the existing scheduler capability for executors
that
set `pre_assigns_external_executor_id = True`: at QUEUED the scheduler
writes a
fresh `external_executor_id` UUID to the `task_instance` row.
- Opts KubernetesExecutor into that path, threads the token through the K8s
workload lifecycle (preserves it across queued events so the scheduler
doesn't
overwrite it with the job id, annotates the pod, revalidates the DB row
just
before pod creation).
- `/run` rejects a worker whose token no longer equals the current TI token.
- Scope: KubernetesExecutor. Identity lives on the mutable/deletable TI row;
it's an equality check, not a lifecycle.

### Option B — durable launch-record table, state machine (PR #70931,
Draft/RFC)

- New durable `task_instance_launch` table (with migration) storing an
immutable
token independent of the mutable/deleted `TaskInstance` row, with an
explicit
`active -> consumed | superseded` lifecycle and guarded transitions.
- Executor-agnostic (Celery / K8s / Edge). The launch record is written
atomically at `SCHEDULED -> QUEUED`.
- Supersession is first-class (stuck-queued reschedule, failed adoption,
orphan
reset, clear/next-try); successful adoption preserves the token.
- `/run` returns a typed `409 stale_executor_launch` for a known *terminal*
token; unknown tokens still `404`. Gated behind Execution API version
`v2026_06_30` (Cadwyn) so older Task SDK clients keep the legacy `404`.
- Task SDK maps `409` to a new `TaskInstanceSupersededError`; a superseded
worker logs once and exits 0 rather than failing.

## Open questions for the list

1. Which direction fits Airflow best — the minimal equality check on the
existing field (A), or the durable, executor-agnostic launch record (B)?
Is the extra table + migration in B justified, or is A's "no schema
change,
reuse `external_executor_id`" preferable and B a later follow-up?
2. Is a typed `409 stale_executor_launch` on `/run` the contract you'd
want, and
is gating it behind a new Execution API version the accepted
compatibility
story here? Or is A's simpler reject sufficient?
3. Should this stay KubernetesExecutor-scoped (A) or land executor-agnostic
from the first cut (B)?
4. Does either warrant an AIP, or is a PR + this thread sufficient given
it's an
additive change (schema/table in B) plus a versioned API behaviour?

Both PRs have Testing sections (unit + a live KubernetesExecutor
kind-cluster
validation on #70931) and #70931 carries a `significant` newsfragment. I'll
consolidate onto whichever approach the list favours and close the other.
Development was AI-assisted (disclosed in the PRs).

Extra:
task_instance_history is per-attempt, while the
proposed record is per-launch, so I don’t believe they are equivalent.

A single task attempt can be launched more than once—for example, after a
stuck-queued reschedule. Each launch has a different token, but
task_instance_history stores only one row per attempt and may be deleted
with the TI.

The launch table stores each token and atomically marks it consumed or
superseded. DAG/task/run/try fields identify the attempt, but only the
token identifies which launch remains valid.

The distinction is:

- task_instance_history: audit snapshot of a completed or replaced attempt;
- task_instance_launch: durable identity and atomic state for each executor
launch.

Just to be clear: saying the current 404 necessarily
consumes a retry was too broad. The main issue is that it conflates an
unknown task with a known obsolete launch; the typed 409 makes that
distinction explicit.

For context, KubernetesExecutor is spawning duplicate, stale task pods that
later fail at /run with 404. These pods may trigger unnecessary node
provisioning, consume cluster capacity, and delay legitimate tasks from
starting. The launch record lets Airflow identify the obsolete launch
explicitly, while scheduler-side token validation can prevent stale pods
from being created in the first place.



P.s.: I used fable 5 for developing this.
Thanks,
Guilherme

GitHub link: https://github.com/apache/airflow/discussions/72473

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to