fix(checkpointer): use AsyncConnectionPool for postgres to prevent stale connection errors (#3223)

[WillemJiang]

Fixes #3223, #3254

Replace AsyncPostgresSaver.from_conn_string() with an explicit
AsyncConnectionPool that has check_connection enabled, so dead idle
connections are detected and replaced on checkout instead of raising
OperationalError.

[Copilot]

Pull request overview

This PR addresses long-lived async Postgres checkpointer failures by switching from a single connection created via AsyncPostgresSaver.from_conn_string() to an explicitly managed psycopg_pool.AsyncConnectionPool configured to validate connections on checkout, reducing “connection is closed” OperationalErrors in idle scenarios.

Changes:

• Replace async Postgres checkpointer creation to use AsyncConnectionPool (with dict_row row factory and connection checking) instead of from_conn_string().
• Add a unit test asserting the async Postgres checkpointer path constructs the saver using a connection pool.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
backend/packages/harness/deerflow/runtime/checkpointer/async_provider.py	Build async Postgres checkpointer using AsyncConnectionPool and pass the pool into AsyncPostgresSaver.
backend/tests/test_checkpointer.py	Add an async test to validate the Postgres checkpointer uses a connection pool rather than from_conn_string().

[rayhpeng]

The direction looks good to me: replacing the single async Postgres connection with AsyncConnectionPool, enabling checkout validation via check_connection, and adding TCP keepalive should make the async checkpointer much more resilient to stale idle connections. CI is green, and the earlier async context manager test issue appears to be addressed.

I’m good with merging this. Two non-blocking suggestions:

1. Consider extracting the duplicated Postgres pool construction into a small helper, since both the legacy checkpointer path and the unified database path now need to stay in sync.
2. Consider adding a test for the unified database.backend == "postgres" path as well, since the PR changes that path too.

Also, the PR body mentions Fixes #3223, while the latest commit mentions Fixes #3254; it may be worth clarifying whether this PR is intended to close both issues.

[WillemJiang]

@rayhpeng Here's a summary of the changes 8f8bc60

async_provider.py — Extracted two helpers from the duplicated postgres code:

• _build_postgres_pool(conn_string) — constructs the AsyncConnectionPool with keepalive + connection checking (was copy-pasted in both _async_checkpointer and _async_checkpointer_from_database)
• _ensure_postgres_imports() — validates and returns the postgres dependencies (was also duplicated)

Both paths now call these shared helpers, so pool kwargs stay in sync by construction.

test_checkpointer.py — Added test_database_postgres_uses_connection_pool that verifies the unified database.backend == "postgres" path creates an AsyncConnectionPool with keepalive, constructs the saver with
conn=pool, and calls setup(). Mirrors the existing legacy-path test.

BTW, #3254 is the same as #3223