Harden test/conftest.py and close the IPv6 and SASL test gaps

[jaysonsantos]

Part of the repository modernization effort.

Background

test/conftest.py starts real memcached processes. The fixtures are not robust.

Problem 1: no start check

test/test_tls.py:39-40 checks p.poll() and calls pytest.skip(...). The four fixtures in test/conftest.py do no such check.

If memcached is absent from PATH, every test fails with an opaque ConnectionRefusedError. A clear skip message is better.

Problem 2: fixed ports

The fixtures use port 11211, port 5000 (test/conftest.py:27), and port 5001 (test/test_tls.py:25). No collision check runs. Two runners on one host break each other. A leftover memcached process breaks the run.

Problem 3: fixed sleep

The fixtures call time.sleep(0.1) after start. A slow start causes a flaky failure.

Fix: poll the port or the socket until it accepts a connection. Use a timeout.

Problem 4: deprecated pytest API

test/conftest.py:12,23,33,45 use pytest.yield_fixture. This is a deprecated alias of @pytest.fixture.

Problem 5: stale unix socket

The unix socket path is /tmp/memcached.sock. See test/test_socket.py:11. If p.kill() leaves the socket file, the next run fails to bind.

Coverage gaps

IPv6 has no end-to-end test

test/conftest.py:44-56 starts a real server with -l::1. No test connects to it.

The only ::1 references live in test/test_server_parsing.py:30-60. Those tests call Protocol.split_host_port(). They parse strings. They never open a socket.

The fixture runs on every session and tests nothing.

Fix: add a real IPv6 round trip test, or delete the fixture. Do not keep an unused fixture.

SASL auth has no integration test

test/test_auth.py:16-61 mocks Protocol._get_response and feeds canned bytes. It tests the client state machine. It never starts a memcached with auth enabled.

No fixture starts an auth-enabled server.

Fix: add a fixture that starts memcached with SASL enabled. Add one test that authenticates and does a set and a get.

Distributed hashing is thin

test/test_distributed_client_hashing.py holds 12 lines. It asserts that one key maps to one server, ten times.

It never tests ring rebalance. That is the interesting property of consistent hashing.

Fix: add a test that adds or removes a server and checks the key-to-server mapping.

Coverage that is adequate

• TLS: test/test_tls.py covers it. Run 34284112624 ran 48 of 48 TLS tests with 0 skips.
• Compression: test/test_compression.py covers zlib, a bz2 swap, a mismatch, and the mocked toggles.
• Unix socket: test/test_socket.py reuses MemcachedTests. It holds no socket-specific edge cases. This is thin but acceptable.

Acceptance criteria

• Every fixture in test/conftest.py checks p.poll() after start. It calls pytest.skip() with a clear message when the process died or memcached is absent from PATH.
• Every fixture replaces time.sleep(0.1) with a connect-retry loop and a timeout.
• pytest.yield_fixture is replaced with @pytest.fixture at test/conftest.py:12,23,33,45.
• The unix socket file is removed on teardown.
• The IPv6 fixture has a test that does a set and a get over IPv6, or the fixture is deleted.
• A SASL-enabled memcached fixture exists. One test authenticates against it and does a set and a get.
• test/test_distributed_client_hashing.py tests the key-to-server mapping after a server changes.
• The suite passes locally and in CI.

Files to change

test/conftest.py, test/test_distributed_client_hashing.py, test/test_auth.py, test/test_server_parsing.py. New files are acceptable, for example test/test_ipv6.py and test/test_sasl_integration.py.

Order

Independent of every other modernization issue. Run it in parallel.

Related

Issue #250 reports that the unit tests are not successful. This work can close that report. Check it.

[jaysonsantos]

Tracked in #277.