Remove Redis Cluster support

[derekste]

Summary

• replace Redis Cluster clients and slot routing with standalone command and reader clients
• classify the first reachable endpoint once with CLUSTER INFO, reject detected Cluster mode, and never repeat detection during reconnects
• make reader bucketing connection-independent so owned, legacy, and generic readers register and recover while Redis is offline
• remove cross-slot fallbacks, Cluster startup/test helpers, and the Cluster-only CI smoke
• preserve {baseKey}:subKey, timestamps, payloads, standalone TCP/Unix sockets, authentication, and connection pooling

Compatibility

• removes the low-level RedisConnection::keyslot() method and copy() == -2 result
• keeps RedisAdapter, RA_Options, ping(key), and time(key) source surfaces otherwise unchanged
• requires Cluster deployments to migrate their endpoint and data to standalone Redis before upgrading
• leaves the library and wire-protocol versions unchanged

Validation

• C++17 Debug and C++20 Release builds
• 109 compatible tests passed with no failures in each C++17 and C++20 run; two Redis 7.4-only XREAD-tail cases skipped on local Redis 7.0.12
• ASan/UBSan: 109 passed, two Redis 7.4-only skips, no sanitizer findings
• TSan: 109 passed, two Redis 7.4-only skips, no race findings
• proxy coverage for one-time standalone classification, cached Cluster refusal, transport faults, trim failures, server rejection, and READONLY recovery
• YAML parsing, Python syntax checks, static Cluster-reference audit, and git diff --check

The local Redis 7.0.12 fixture cannot run the existing HEXPIRE watchdog case; the pinned Redis 7.4 CI fixture covers it.

Review requested from @wsulli and @bigsamich.

Closes #65

[Copilot]

🔵 Needs a closer look

Core connection replacement and concurrent reader recovery behavior warrant final human review despite comprehensive tests.

0 open findings

What changed in this PR

Removes Redis Cluster support while preserving standalone Redis compatibility and improving offline reader recovery.

Changes:

• Replaces Cluster routing with standalone command/reader clients and one-time mode detection.
• Makes reader registration and recovery connection-independent.
• Removes Cluster fixtures, CI coverage, and obsolete documentation.

File	Description
RedisConnection.hpp	Implements standalone clients and Cluster rejection.
RedisAdapter.cpp	Simplifies bucketing, copying, and reader recovery.
RedisAdapter.hpp	Updates reader and standalone API documentation.
connection_policy_test.cpp	Tests cached mode detection.
lifecycle_test.cpp	Tests offline reader recovery.
write_failure_test.py	Extends proxy classification and reconnect checks.
write_error_test.cpp	Adds constructor and Cluster-refusal scenarios.
test.cpp	Tests key schema and cross-base utilities.
recovery_test.cpp	Removes Cluster recovery coverage.
.github/​workflows/​test.yml	Removes Cluster CI smoke test.
scripts/​run-test-redis-cluster.py	Deletes private Cluster fixture.
create-cluster	Deletes legacy Cluster creation script.
cluster-start.sh	Deletes Cluster startup wrapper.
README.md	Documents standalone-only support.
docs/​api.md	Updates connection and recovery behavior.
docs/​building.md	Removes Cluster test instructions.
docs/​stream-subscriptions.md	Documents standalone subscription behavior.
docs/​redis-adapter-implementation-spec.md	Preserves the braced wire-key schema.
CHANGELOG.md	Records Cluster removal and compatibility impact.
.github/​ISSUE_TEMPLATE/​bug.yml	Requests supported connection details.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[bigsamich]

Reviewed at 95792a3. Looks good. The removal is clean: no Cluster references remain outside the new detection path and docs, reader_token() no longer depends on the connection, and CI (build-test, ASan, TSan) is green. Two minor items and a note:

1. A cached Cluster refusal still reconnects on every attempt (RedisConnection::connect()). After _initial_server == Unsupported, connect() still builds a pool and sends PING before acceptStandalone() returns false. On a Cluster endpoint, every failed write or read calls reconnect(0). The adapter keeps opening a fresh connection, as often as every ~100 ms, for the life of the process, and only the first attempt logs. Suggest returning false at the top of connect() once the cached result is Unsupported. The cluster-refusal proxy test's ping_probes == 2 would then become 1.

2. The mode is cached per RedisConnection object, not per endpoint. connect(opts) is public and can target a different host or socket; the new OfflineInitialConnectionDefersModeProbeUntilReachable test does exactly that. The result from the first reachable endpoint then applies to every later one. RedisAdapter always reconnects to _options.cxn, so the adapter is not affected. Either document this on connect() or key the cache on host/port/path.

3. Note: the PR drops the stop/restart-all-readers block from the reconnect thread. That also closes the "late reconnect thread restarts readers after ~RedisAdapter stopped them" path in #118 (part 2), which may be worth a line in the PR body. #118 stays open for readers restarted from callbacks via removeReader() and setDeferReaders(false).

[derekste]

Thanks for the review. I classified both requested changes as non-blocking follow-ups and filed them without changing the approved PR:

• Stop reconnect churn after a cached Redis Cluster rejection #133 tracks eliminating connection and PING churn after a cached Cluster rejection.
• Scope cached Redis server mode to the connection endpoint #134 tracks scoping the cached server-mode result to the endpoint, or explicitly constraining endpoint reuse.

The reconnect-thread reader-restart note is now recorded on #118. PR #132 removes that specific restart path, while #118 remains open for the callback-driven shutdown races.

All required checks on this PR are green, and there are no open inline review threads.