[vishal-bala]

Motivation

No CI job passes --run-cluster-tests, so every test marked requires_cluster has been skipped on every run. Cluster-only regressions, and cluster tests that no longer test what they claim, have gone unnoticed as a result. This PR runs that suite nightly so the cluster code path gets regular coverage without slowing down pull requests.

Changes

Nightly cluster job

A new cluster-tests job in .github/workflows/test.yml runs pytest -n 0 --run-cluster-tests -m requires_cluster tests/ on the nightly schedule and on workflow_dispatch, and is skipped for pull requests and pushes. It runs on Python 3.11 across redis-py 6.x, 7.x and 8.x, because cluster replies differ by redis-py version: 8.x defaults to RESP3. The install step mirrors the test job's redis-py pinning. The run is serial because tests/cluster-compose.yml publishes fixed ports 7001–7006, which per-worker clusters under xdist would fight over.

Resolving host.docker.internal on Linux

Every cluster node announces host.docker.internal. That is the address the nodes gossip to each other, the one redis-cli --cluster create reconnects to while building the cluster, and the one the client is handed in MOVED redirects. So it has to resolve inside every container and on the host. Docker Desktop provides the name inside containers, but Linux Docker Engine, which GitHub's ubuntu runners use, does not. Every container in tests/cluster-compose.yml now maps it to host-gateway, and the job adds 127.0.0.1 host.docker.internal to the runner's /etc/hosts.

Cluster fixture no longer exports COMPOSE_PROJECT_NAME

redis_cluster_container in tests/conftest.py used to set COMPOSE_PROJECT_NAME to the cluster's project. The standalone redis_container fixture's testcontainers commands read that variable. So once a cluster test had run, redis_url looked for the standalone redis service in the cluster's project and raised ContainerIsNotRunning. Under -m requires_cluster a cluster test runs first, so the new job would have hit this on every leg. Every command in the cluster fixture already passes -p, so it now leaves the variable alone. This also stops the standalone container from leaking after a cluster run.

• CONTRIBUTING.md now says the test host must resolve host.docker.internal, and that CI runs the suite nightly.
• AGENTS.md and the pre-PR review references no longer say cluster tests never run in CI; they now say pull-request CI.

Notes

This PR is the top of a stack, #794 → #795 → #800 → this PR, because the new job would fail on main today. #794 fixes three cluster tests that silently used a single-node client, #795 fixes SearchIndex.info() and AsyncSearchIndex.info() on cluster clients, and #800 stops a cluster test from overrunning redis-py 8's per-node connection cap. The three fixes are independent; each branch merges the one below it only so that the stack is linear.

The job lives in test.yml rather than in a workflow of its own, so the nightly schedule stays in one place. The cost is that an on-demand run dispatches all of test.yml, including the service tests that use the provider API secrets.

Reviewer Brief

Start here: tests/cluster-compose.yml. The job only works if every container, and the runner, resolves the address the nodes announce, and this file is where most of that happens.

Where it could fail

• tests/conftest.py:84: the cluster fixture no longer exports COMPOSE_PROJECT_NAME. Check that nothing else relied on that variable pointing at the cluster project after the fixture had run.
• .github/workflows/test.yml, the cluster-tests install step: it copies the test job's redis-py case block, so a pin changed in only one copy leaves the nightly testing stale versions without failing.

Not verified: local runs against this compose file on macOS with the /etc/hosts entry (local runs used a stand-in topology announcing 127.0.0.1); Docker Desktop on Windows; more than one green run on Linux, though two of the failures this job first exposed were timing-dependent.

Beyond CI: a workflow_dispatch run on c220e6e3 (run 41057698156) formed the cluster on an ubuntu runner and passed all 22 cluster tests on redis-py 6.4.0, 7.4.1 and 8.1.0, with every other job in the workflow green. Before the conftest change, the job's test order made test_sync_client_validation error with ContainerIsNotRunning locally; after it, that test passes.

Since the last brief: merged the updated #795 and the new #800 below this PR in the stack; this PR's own diff is unchanged. The first dispatch, on a55375f0, failed three tests on redis-py 8.x, all fixed in those two PRs.

_{Reviewed 236b38de, fixes in a55375f0 · Claude Code (claude-opus-5-5) · lenses: correctness, security, design, tests, docs, Redis behaviour}