fix(connection): convert sync Retry to async Retry in sync_to_async_redis - #817

Follow-up to #792, from this review comment.

Motivation

sync_to_async_redis passed the sync client's retry keyword through unchanged. The async connection awaits retry.call_with_retry(...), and a sync redis.retry.Retry returns the coroutine without awaiting it, so the retry policy never ran: a converted client failed on the first command after CLIENT KILL, where a native async client recovers.

Changes

• _async_connection_kwargs rebuilds a sync Retry as redis.asyncio.retry.Retry with the same backoff object, retry count and supported error set (_async_retry). retry=None and a missing keyword are left alone.
• Unit tests in tests/unit/test_sync_to_async_redis.py pin the conversion; an integration test in tests/integration/test_connection.py kills the converted client's connection and checks the next command succeeds.

Release Notes

Async paths that convert a caller's sync client (SemanticCache, AsyncSearchIndex, and so on) now keep that client's retry policy, so they reconnect after a dropped connection the way a native async client does.

Reviewer Brief

Start here: redisvl/redis/connection.py, _async_retry. A client converted from sync to async keeps its retry policy only if this rebuilds it, and _async_connection_kwargs calls it for every conversion, including a plain Redis(), which carries a Retry by default.

Where it could fail

• redisvl/redis/connection.py:383: _async_retry reads Retry._backoff and Retry._supported_errors, private attributes with no public getter in redis-py 6.3 through 8.1; if a future release renames one, every conversion raises AttributeError instead of quietly dropping the policy.
• redisvl/redis/connection.py:421: a user's Retry subclass is rebuilt as the stock async Retry, so an overridden call_with_retry is not carried over (it could not have run on the async connection before this change either).

Not verified: cluster and Sentinel clients (both are rejected before this code runs); redis-py versions between 6.3.0 and 8.1.0 other than 7.4.1 and 8.0.1.

Beyond CI: test_converted_client_recovers_after_client_kill failed 20 of 20 runs on main and passed 20 of 20 with this change, on both redis-py 6.3.0 and 8.1.0. The unit tests pass on redis-py 6.3.0, 7.4.1, 8.0.1 and 8.1.0, and hardcoding the retry count in _async_retry makes test_sync_retry_is_rebuilt_as_the_async_retry fail. A sync client built with retry_on_timeout=True or retry_on_error=[BusyLoadingError] converts to an async connection whose Retry includes those errors, on 6.3.0 and 8.1.0.

Since the last brief: 57e898e tightens the tests only: a non-default retry count of 5, two redundant tests and assertions dropped, and a clearer CLIENT KILL test docstring.

_{Reviewed 57e898e · Claude Code (claude-opus-5-5) · lenses: correctness, security, design, tests, docs, Redis behaviour}