limits > 2.0.3 breaks with Redis Sentinel

Hi, I am the maintainer of the Falcon-Limiter package (see https://github.com/zoltan-fedor/falcon-limiter) and one of my users has raised an issue about the Falcon-Limiter breaking when using limits version 2.3.0 with Redis Sentinel, while it works with llimits version 2.0.3. See the original issue at https://github.com/zoltan-fedor/falcon-limiter/issues/1

I did manage to reproduce the user’s issue in Falcon-Limiter after deploying Redis Sentinel onto my Kubernetes cluster.

Then I went and tested the limits library (without the Falcon-Limiter) and unfortunately I was able to reproduce the same here too, so it seems to me this is an issue with the limits library, not with how Falcon-Limiter makes calls to it. (correct me if I am wrong here)

Below is the test code I use:

from limits import storage, strategies, parse
redis_password = "xxxxxxx"

r_storage = storage.storage_from_string(f"redis+sentinel://:{redis_password}@127.0.0.1:26379/mymaster",
    sentinel_kwargs={"password": redis_password})
moving_window = strategies.MovingWindowRateLimiter(r_storage)
one_per_minute = parse("1/minute")

assert True == moving_window.hit(one_per_minute, "test_namespace", "foo")

When running this with limits version 2.3.0 I get the following error - while there is no error when using version 2.0.3.

Traceback (most recent call last):
  File "test-limits.py", line 9, in <module>
    assert True == moving_window.hit(one_per_minute, "test_namespace", "foo")
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/limits/strategies.py", line 84, in hit
    return self.storage().acquire_entry(  # type: ignore
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/limits/storage/redis.py", line 178, in acquire_entry
    return super()._acquire_entry(key, limit, expiry, self.storage, amount)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/limits/storage/redis.py", line 79, in _acquire_entry
    acquired = self.lua_acquire_window([key], [timestamp, limit, expiry, amount])
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/commands/core.py", line 4440, in __call__
    return client.evalsha(self.sha, len(keys), *args)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/commands/core.py", line 3891, in evalsha
    return self.execute_command("EVALSHA", sha, numkeys, *keys_and_args)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1176, in execute_command
    return conn.retry.call_with_retry(
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/retry.py", line 44, in call_with_retry
    fail(error)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1180, in <lambda>
    lambda error: self._disconnect_raise(conn, error),
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1166, in _disconnect_raise
    raise error
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/retry.py", line 41, in call_with_retry
    return do()
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1177, in <lambda>
    lambda: self._send_command_parse_response(
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1153, in _send_command_parse_response
    return self.parse_response(conn, command_name, **options)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/client.py", line 1192, in parse_response
    response = connection.read_response()
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/sentinel.py", line 61, in read_response
    return super().read_response(disable_decoding=disable_decoding)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/connection.py", line 800, in read_response
    response = self._parser.read_response(disable_decoding=disable_decoding)
  File "/home/myuser/.local/share/virtualenvs/test-limits-kWUTuU1U/lib/python3.8/site-packages/redis/connection.py", line 336, in read_response
    raise error
redis.exceptions.AuthenticationError: Authentication required.

Does limits version 2.3.0 require the password for Sentinel to be passed differently?

In case you want to reproduce the issue on your side, unfortunately you will need to deploy a Redis Sentinel cluster. There are additional details about that in the original ticket, but just a quick summary:

# deploy Redis Sentinel onto a Kubernetes cluster (or Minikube):
$ helm repo add bitnami https://charts.bitnami.com/bitnami
$ helm install redis-sentinel bitnami/redis --set sentinel.enabled=true

# port-forward the services
$ kubectl port-forward --namespace default svc/redis-sentinel 26379:26379
$ kubectl port-forward --namespace default svc/redis-sentinel 6379:6379

# get the Redis password
$ export REDIS_PASSWORD=$(kubectl get secret --namespace default redis-sentinel -o jsonpath="{.data.redis-password}" | base64 --decode

# you might also need to alias the redis-sentinel-node-0.redis-sentinel-headless.default.svc.cluster.local domain name to localhost in your host file

Based on this test (https://github.com/alisaifee/limits/blob/master/tests/storage/test_redis_sentinel.py#L33) this scenario should work, but obviously it doesn’t - likely the mocking is not able to capture the issue.

[alisaifee]

For future readers this issue exists for limits versions >2.0.3, <=2.3.0. Since the actual releases in that range happened in a reasonably short period of time, I’ll only be making a fix in 2.3 (i.e. 2.3.1 contains the fix to ensure backward compatibility)

[alisaifee]

Thanks for the report and my apologies for the backward incompatibility - this was entirely unintentional.

The behaviour in version < 2.1 was incorrect and had resulted in people having to workaround by supplying the password for the master/slave nodes in the uri and passing the password for the sentinel (correctly) in the sentinel_kwargs. I didn’t even realise that and as part of the refactoring ended up breaking this.

The “expected” way to pass the parameters when both the sentinel and the master/slave are password protected would be either:

uri = f"redis+sentinel://{sentinel_host}:{sentinel_port}"
storage_options = {"sentinel_kwargs": {"password": sentinel_password, "password": master_password}

or

uri = f"redis+sentinel://:{sentinel_password}@{sentinel_host}:{sentinel_port}"
storage_options = {"password": master_password}

At the moment I don’t have any clear idea on how to add a patch that would be backward compatible as well as encourage the correct behaviour - I’m wondering if this can be communicated via documentation/release notes.

Open to suggestions though!