Fix asset event scheduling race condition and consistency

[dingo4dev]

This PR addresses a race condition in the Asset-based scheduling logic where AssetEvent records created by concurrent tasks can be "missed" or "orphaned" by the Scheduler.

The issue stems from the Visibility Gap between a database flush() (which generates the created_at timestamp) and the final commit() (which makes the record visible to other sessions).

This change ensures that the synchronization between the AssetEvent and the AssetDagRunQueue is atomic and that the timestamps are aligned to prevent the drift caused by performance issue.

• Manually assigns the same timestamp to both the AssetEvent and the ADRQ before the flush.
• Commit and populate the asset event record in database immediately, so that the event can be accessible by job scheduler
• Prevent creating dag by empty asset event as the asset event is trigger in prev_dag_run

related: #54659, #56750, #56749

---
config:
  theme: redux-dark-color
  look: neo
---
sequenceDiagram
    participant J1 as Task Instance 1
    participant J2 as Task Instance 2
    participant JS as Job Scheduler
    participant DB as Database

    Note over J1,DB: T0

    Note left of J1: T1
    J1-->>DB: Asset Event 1 (AE1) & commit
    activate J1
    activate DB
    Note left of J1: T2
    J2-->>DB: Asset Event 2 (AE2) & commit
    activate J2
    activate DB

    Note left of J1: T3
    J2-->>-DB: Asset Dag Run Queue (ADRQ)<br/>created_at = AE2.timestamp<br/>(flush & commit)
    deactivate DB
    Note left of J1: T4
    activate JS
    activate DB
    JS-->>DB: Fetch created_at from ADRQ
    JS-->>DB: Fetch events (AE1, AE2)
    JS->>JS: Create DAG with run_after = ADRQ created_at (AE2.timestamp)
    deactivate DB
    deactivate JS
    Note left of J1: T5
    J1-->>DB: Asset Dag Run Queue (ADRQ)<br/>created_at = AE1.timestamp<br/>(flush & commit)
    deactivate J1
    deactivate DB
    Note left of J1: T6
    activate JS
    activate DB
    JS-->>DB: Fetch created_at from ADRQ
    JS->>JS: No event found<br/>(AE1.timestamp < prev DAG run_after)
    deactivate DB
    deactivate JS

Loading

---

Was generative AI tooling used to co-author this PR?

• Yes (please specify the tool below)

[Copilot]

Pull request overview

This PR attempts to fix a race condition in asset-based scheduling where AssetEvent records created by concurrent tasks can be missed by the Scheduler due to a visibility gap between database flush() and commit() operations. The fix synchronizes timestamps between AssetEvent and AssetDagRunQueue records by manually assigning timestamps and committing the asset event immediately.

Changes:

• Modified asset event registration to commit immediately after creating AssetEvent, ensuring early visibility to the scheduler
• Updated AssetDagRunQueue creation to use the asset event's timestamp for synchronization across both PostgreSQL and slow-path implementations
• Changed scheduler logic to conditionally create DAG runs only when asset events exist, and to delete ADRQ entries more selectively based on timestamps

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 7 comments.

File	Description
airflow-core/src/airflow/jobs/scheduler_job_runner.py	Modified DAG run creation to be conditional on asset_events existence and updated ADRQ deletion to filter by timestamp
airflow-core/src/airflow/assets/manager.py	Changed from session.flush() to session.commit() for early AssetEvent persistence and updated ADRQ timestamp synchronization logic