Fix async engine missing pool_recycle and pool_pre_ping configuration

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866

[kaxil]

The async SQLAlchemy engine was created with zero pool configuration while the sync engine got pool_size, pool_recycle, pool_pre_ping, and max_overflow from [database] config.

How this is hit

The async engine serves all Execution API requests that go through async session dependencies. The API server (gunicorn) creates the async engine at startup via _configure_async_session(). Without pool health settings:

• pool_recycle=-1 (SQLAlchemy default) -- connections are never recycled, so connections that exceed PostgreSQL's idle_in_transaction_session_timeout or pgbouncer's server_idle_timeout sit dead in the pool
• pool_pre_ping=False (SQLAlchemy default) -- dead connections are never detected before checkout, so the first query on a stale connection fails with a closed-connection error

The sync engine has had these settings since day one. The async engine missed them when _configure_async_session() was extracted from configure_orm() in PR #51920.

Fix

Read the same [database] config section values (pool_size, pool_recycle, pool_pre_ping, max_overflow) for the async engine. Also respect SQL_ALCHEMY_POOL_ENABLED=False by using NullPool, matching the sync engine behavior for pgbouncer setups that require it.

The create_async_metadata_engine() signature gains an optional engine_args parameter (defaulting to None for backward compatibility with existing airflow_local_settings.py overrides). Upstream PR should update the cluster-policies.rst docs to show the new signature.

[pierrejeambrun]

Thanks for the PR!

Just one comment

[pierrejeambrun]

Maybe we can have a 'common' function that builds the 'common' part only of the engine args.

And then let the async/sync specific parts like you have now.

But that's not super important, if you want to go with full duplication that's not a big deal.

[vatsrahul1001]

Manual backport #66866