What

Adds celery/_redis_py_recursion_workaround.py, imported from celery/__init__.py, that backports redis-py PR #3557 for redis-py < 6.0.0b2. Faithful port of the upstream fix:

• Adds Connection.connect_check_health and Connection.on_connect_check_health.
• Threads check_health through every send_command in the on_connect chain.
• Changes send_packed_command's stale-socket reconnect from self.connect() to self.connect_check_health(check_health=False).
• Preserves the upstream transparent-reconnect contract.
• Adds a single WARNING-per-process observability log the first time the suppressed-check_health reconnect fires (subsequent reconnects log at DEBUG).

Why

On 2026-06-26T03:46:35Z–03:46:50Z, 171 Celery worker pods in production crashed with RecursionError('maximum recursion depth exceeded while calling a Python object') during a Redis Deployment Recreate rollout. Confirmed via prod cluster inspection: Redis ReplicaSet redis-6f9c8ffc9f was created at 2026-06-26T03:45:56Z, matching the first worker "Connection to broker lost" log at 03:45:55.820Z.

The bug is in redis-py 5.2.1's reconnect path (upstream issue #3745, fixed in PR #3557):

connect() -> on_connect()
  -> send_command("CLIENT","SETINFO","LIB-NAME", lib_name)
  -> send_packed_command(check_health=True)
    -> check_health() -> retry.call_with_retry(_send_ping, _ping_failed)
      -> _send_ping()                             # fails (any retryable error)
      -> _ping_failed() -> disconnect()           # _sock = None
      -> _send_ping()                             # retry
        -> send_command("PING", check_health=False)
        -> send_packed_command(check_health=False)
        -> if not self._sock: self.connect()      # <-- re-enters top of chain
        -> on_connect() -> CLIENT SETINFO -> check_health -> ...
        -> recursion (~166 levels until Python stack limit)

Trigger conditions (all set in every Celery service's broker_transport_options):

• health_check_interval = 30
• retry_on_timeout = True
• lib_name non-empty (redis-py defaults it to "redis-py")

Failure-mode agnostic: the recursion fires regardless of whether the initial _send_ping failure is TimeoutError, ConnectionError, BusyLoadingError, or a raw socket error. This fix breaks the cycle at the reconnect point, so all these variants surface as normal handleable errors.

Why not just upgrade redis-py

kombu==5.5.2 pins redis<=5.2.1. celery 5.5.x pins kombu<5.6. A redis-py bump alone is rejected by uv sync. Proper fix is a coordinated celery 5.6+ / kombu 5.6+ / redis-py 6.0+ upgrade, which also requires rebasing this fork onto upstream celery 5.6+ (forward-porting rbehal's spawn pool work, PR #1 --disable-prefetch, PR #2 GMLP-9012). That is tracked strategically. This PR is the tactical mitigation.

Idempotency / no-op conditions

• redis-py not installed → no-op
• redis-py ≥ 6.0.0b2 (detected via hasattr(Connection, 'connect_check_health')) → no-op
• Guards against double-patching on re-import

Validation

Two reproduction methods worked on the beta cluster:

• Deterministic in-pod repro: kubectl exec a Python script into a worker pod that monkeypatches redis.connection.Connection._send_ping to always raise TimeoutError after sending PING. Then client.ping() against live beta Redis. Unpatched → 5/5 RecursionError. Patched → 5/5 non-recursion.
• Live Redis DEBUG SLEEP + tracer script: temporarily enabled --enable-debug-command yes on the beta Redis Deployment, then redis-cli DEBUG SLEEP 180 (backgrounded) + a tracer script inside a worker pod using socket_timeout=0.2 so each recursion level ≈ 200 ms. Unpatched → RecursionError max_depth=996 in 13.42 s. Patched → non-recursion TimeoutError at max_depth=29 in 0.81 s. Redis deployment reverted after.

Test wheel 5.5.3+gumloop.0.1.5.gmlp9206.test1 is on beta cluster now. Do not promote that tag to staging/prod. Cut a clean 5.5.3+gumloop.0.1.6 on merge.

Caveats

• Only takes effect when celery is imported (celery/__init__.py applies the patch as import side-effect). For our Celery broker code paths this is universally true.
• App code that uses redis-py directly without importing celery first will NOT get the workaround; that path is covered by the coordinated redis-py upgrade tracked separately.

References

• redis-py PR documentation for 3.1 is not available on rtd celery/celery#3557: Fixed infinitely recursive health checks redis/redis-py#3557
• redis-py issue Consumer does not shutdown properly when embedded in gevent application celery/celery#3745: RecursionError in redis.asyncio when using retry_on_timeout or health_check_interval redis/redis-py#3745
• redis-py issue ignore_result=True in add_unlock_chord_task preventing all backend reporting? celery/celery#2563 (same recursion via max clients reached): Infinity retry on the exception of max number of clients reached redis/redis-py#2563
• Linear: GMLP-9206