From 8a561b15214148e5fba9a7bbb7ebfb36539a1c09 Mon Sep 17 00:00:00 2001
From: Alexey Shalaev <75322386+AlexeyShalaev@users.noreply.github.com>
Date: Sun, 6 Sep 2026 23:06:56 +0300
Subject: [PATCH] fix: fail startup on an unreachable Redis, and count
 translated errors as errors

The work landed in #22. Its squash subject lost the Conventional Commit prefix --
my mistake on the merge, not the author's -- so release-please skipped the merge
and these fixes would never have reached a release. This commit carries the
record. It changes no code: #22 is already on master.

* retry_async_connection could not fail startup. It raised only when connect_func
  raised, while the health check returned False instead of raising, so the loop
  spun three times with no backoff and handed back a client that could not reach
  Redis. A falsy result is a retried failure now and the last attempt raises.
  Reproduced against a dead port.
* AsyncRedisProvider.get_redis was unreachable: it and get_redis_with_health_check
  provided the same dishka key and the second won. The choice moved to a
  constructor flag and both methods are reachable.
* Provider order was load-bearing and silent: the Redis provider supplied a None
  metrics collector, so registered after a metrics provider it overrode it and you
  got a plain client with no error. A flag turns the default off.
* The async instrumented client read pool._all_connections, which redis-py 8 does
  not have, so the pool-size gauge was always zero. It sums the two attributes the
  sync client already used.
* In all four clients the RuntimeError branch preceded the one that sets an error
  status, so a uvloop transport error translated to ConnectionError was counted as
  a success and never recorded.
* The three guide pages and the docs home documented a settings API that does not
  exist -- every snippet raised a validation error -- along with OpenTelemetry
  support, a wrong redis floor, a wrong Python floor and an extra that is not
  declared. All corrected against the code.