test: upgrade pytest-asyncio and add loop scoping

[glevkovich]

Upgrade pytest-asyncio from 0.20.1 to 0.24.0+ and configure
event loop scope to prevent fixture crashes.

requirements.txt changes:

• pytest>=8.3,<9.0 (was >=7.1.2)
• pytest-asyncio>=0.24.0,<0.25.0 (was ==0.20.1)

conftest.py changes:

• Remove manual event_loop fixture override (deprecated in 0.24+)
• Rewrite async_client fixture to bypass async_pool. With
  session-scoped loops, the standalone ConnectionPool in async_pool
  retains connections with stale loop references when df_server
  restarts between test classes. Now async_client creates Redis
  directly with connection params, letting Redis manage its own
  internal pool and adding proper aclose() cleanup. This fixes 7
  failing tests (test_publish_stuck, test_nested_client_pause,
  test_cron_snapshot, test_set_cron_snapshot, test_basic_memory_usage,
  test_mixed_append, test_tiered_replication_with_hashes).

pytest.ini changes:

• Add asyncio_default_fixture_loop_scope=session

pytest-asyncio 0.24+ forbids manual event_loop overrides and
requires explicit loop scope configuration. We use session scope
because:

1. df_factory and other fixtures are class-scoped
2. Many async tests are module-level functions (not in classes)
3. session scope supports both patterns without UsageError

Upper bounds prevent pip from installing pytest 9.x and
pytest-asyncio 1.x, which conflict with dragonfly-fakeredis-tests.

snapshot_test.py changes:

• test_s3_snapshot, test_s3_save_local_dir: replace async_client
  fixture with df_server.client(). With session-scoped loops,
  the shared async_client pool holds stale connections from a
  previous df_server instance, causing "Future attached to a
  different loop". Creating a fresh client per test avoids this
  (same pattern used by test_s3_reload_snapshot_after_restart).

[Copilot]

Pull request overview

Upgrades the Python test harness dependencies to support newer pytest / pytest-asyncio behavior and adds explicit asyncio event loop scoping to avoid fixture scope conflicts in async tests.

Changes:

• Bump pytest to >=8.3,<9.0 and pytest-asyncio to >=0.24.0,<0.25.0 for the tests/dragonfly suite.
• Configure pytest-asyncio to use function-scoped default fixture event loops via asyncio_default_fixture_loop_scope=function.

Reviewed changes

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

File	Description
tests/pytest.ini	Adds explicit pytest-asyncio default fixture loop scope configuration.
tests/dragonfly/requirements.txt	Updates pytest/pytest-asyncio version constraints to newer supported ranges.

[augmentcode]

🤖 Augment PR Summary

Summary: Updates the Python integration test dependencies to use newer pytest/pytest-asyncio versions and adds explicit pytest-asyncio loop scoping.

Why: Aligns with pytest-asyncio 0.24+ strict loop-scope behavior to avoid fixture scope/loop-scope conflicts and resulting test failures.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 2 suggestions posted.

Comment augment review to trigger a new review at any time.

[vyavdoshenko]

lgtm

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

[vyavdoshenko]

I guess this is a mistake.
It should be:

asyncio_default_fixture_loop_scope=class

or even

asyncio_default_fixture_loop_scope=function

[glevkovich]

Not a mistake - tried both and they crash:

• function: crash with "Future attached to a different loop" because df_factory in tests/dragonfly/conftest.py is class-scoped and outlives the per-function loop.

• class: crash with with "UsageError: test_client_migrate is marked to be run in an event loop with scope class, but is not part of any class" - many tests are standalone functions, not inside a class.

session is the only scope that works for me, with both class-scoped fixtures and module-level test functions under pytest-asyncio 0.24's scoping rules.

[glevkovich]

Adding more context on the conftest.py changes :
Even with session scope, the old async_pool fixture was caching stale references to the loop across server teardowns, causing "Event loop is closed" and "Future attached to a different loop" errors during garbage collection.

So i decided to bypass async_pool and instantiating aioredis.Redis directly inside the async_client fixture, the client dynamically binds to the active loop every time and safely cleans up its sockets before Pytest tears the loop down. I hope that this combination finally makes the v0.24 upgrade stable - lets see if tests pass.

[glevkovich]

@vyavdoshenko please hold on with the review. This PR will take me more work than I thought to finalise (still lots of tests are failing so what I have no is probably not the final product). I need to fix more of tests/infra to make it work. Holding it for now...