Register asset events before taking the task_instance row lock

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.

[Andrushika]

Register asset events before taking the task_instance row lock

Why

When a task succeeds, ti_update_state currently takes the task_instance row lock and then registers asset events while holding the lock.

Asset registration can be slow because it performs asset lookups, event inserts, and Dag-run queueing. When many tasks finish at the same time, concurrent requests can wait for the row lock while each request continues to use an API-server thread. In the case reported in #66853, this caused the API server to run out of memory and get OOMKilled.

What

This change registers the asset events before taking the task_instance row lock:

The asset events and the task state are still written in the same database transaction.

In this PR, we first read the state without the row lock, then perform the asset registration. This can hit a TOCTOU problem: the state may change in between, for example, when the worker retries. So before we actually change the state, we take the row lock and read the state one more time. The lock makes sure no one else can change it after this point. Then we decide what to do.

There are three cases:

1. The state is still RUNNING: the normal path. We change the state to SUCCESS and commit. The state and the events are saved together in one transaction.

2. The state is already SUCCESS: another request won the race and already saved its own events. We roll back so the events are not saved twice, and return 200.

3. The state is something else, such as FAILED: the task is no longer running, so the update is rejected with 409 and the transaction is rolled back. No events from this request are saved.

This locked re-check keeps the task state and asset events atomic: the events are committed only together with the SUCCESS state.

Benchmark

The task_instance row lock is held during one successful task completion, on Postgres. The task emits to N assets to make the registration slow. Median of 14 runs (lock holding time):

outlets	before	after
300	426 ms	2.0 ms
600	974 ms	2.1 ms

Before, the lock is held for the whole registration, so it grows with the number of outlets. After, the registration runs before the lock, so the lock is held only for the state update and stays flat.

This is an alternative to #66854.

related: #66853

---

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

• Yes (please specify the tool below)

Generated-by: Claude Code Opus 4.8 following the guidelines

---

• Read the Pull Request Guidelines for more information. Note: commit author/co-author name and email in commits become permanently public when merged.
• For fundamental code changes, an Airflow Improvement Proposal (AIP) is needed.
• When adding dependency, check compliance with the ASF 3rd Party License Policy.
• For significant user-facing changes create newsfragment: {pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.

[Andrushika]

Close this since #70951 is merged.