Writing Unit Tests

Writing Unit Tests for SCT

Write isolated, fast unit tests that never contact external services. All new unit tests go in unit_tests/unit/ (not the root unit_tests/ directory).

Essential Principles

External Services Must Be Blocked

Unit tests must never make real network calls — not to AWS, GCE, Azure, Docker registries, or any external endpoint.

The unit_tests/unit/conftest.py autouse fake_remoter fixture blocks SSH connections automatically. But HTTP-based services (boto3, requests, REST APIs) are NOT auto-blocked. You must mock them explicitly using unittest.mock.patch, monkeypatch, or moto.

If a test is slow or flaky, the first suspect is an unmocked network call.

Use pytest, Not unittest.TestCase

All tests use pytest functions, fixtures, and assert — never unittest.TestCase or class Test*.

unittest.TestCase breaks pytest's fixture injection, autouse fixtures, parametrize, and parallel execution. SCT requires pytest-native style throughout unit_tests/unit/. This means no test classes at all — use flat module-level def test_* functions and group related tests with comment blocks (see AP-4).

Never duplicate test infrastructure — fake objects, base classes, runner stubs, and fixture setup code. Before adding anything new, check fake_cluster.py, unit_tests/unit/nemesis/__init__.py, execute_nemesis/__init__.py, and unit_tests/unit/conftest.py. Add concrete subclasses or attributes to existing structures; only create a new class hierarchy when the registry under test must be isolated from existing subclasses, and document the reason (see AP-5).

Tests Must Be Isolated and Parallel-Safe

Every test must pass independently, in any order, and in parallel.

SCT runs tests with pytest-xdist (-n2 by default) and pytest-random-order. Never rely on test execution order, shared mutable state, or global side effects. Use fixtures for setup/teardown, monkeypatch for environment variables, and tmp_path for file-based tests.

Special care for Singleton classes: SCT has classes with metaclass=Singleton (e.g. NodeLoadInfoServices, AdaptiveTimeoutStore subclasses) that persist mutable state across tests on the same worker process. Add an autouse fixture that clears the cache in teardown (post-yield only — never pre-yield). See pitfall P-16 for details.

Mock at the Boundary, Not the Logic

Mock external dependencies (network, file system, cloud APIs) — not internal SCT logic.

Mocking internal functions makes tests brittle and hides bugs. Mock at the outermost boundary: the HTTP call, the SSH command, the cloud SDK client. This tests the actual logic while isolating from infrastructure.

Never reimplement the code under test in a fake class. If you find yourself copying a method body from sdcm/ into a FakeFoo helper in your test, stop — you are testing the copy, not the real code. Always instantiate the real class and mock only its external I/O (network, file system, cloud APIs). See anti-pattern AP-6 for details.

No Inline Classes in Fixtures or Tests

Define helper classes at module level, not inside fixtures or test functions.

Inline classes (defined inside a function or fixture) are harder to read, cannot be reused, and make diffs confusing. Define helper classes at module level and instantiate them in fixtures. This keeps test code flat and scannable.

Use Events Fixtures for Event System Tests

Use events_function_scope fixture when tests publish or read SCT events — never manage EventsUtilsMixin manually.

The events_function_scope fixture (from unit_tests/conftest.py) creates a fully isolated events system per test — fresh temp directory, events device, and registry patcher. This prevents event leakage between tests. Access the raw events log via events_fixture.get_raw_events_log() and the events logger via events_fixture.get_events_logger(). Use events (module scope) only when many tests share expensive event setup and you are certain there is no cross-test interference.

When to Use

• Creating a new test file in unit_tests/unit/
• Adding test cases to an existing unit test module in unit_tests/unit/
• Mocking AWS, GCE, Azure, or other cloud services in tests
• Setting up pytest fixtures for SCT components
• Debugging a unit test that is failing, slow, or flaky
• Converting unittest-style tests to pytest style

When NOT to Use

• Writing integration tests that need Docker or real services — use the writing-integration-tests skill
• Running or configuring CI pipelines — edit Jenkins pipeline files directly
• Writing functional tests for K8s operators — see functional_tests/
• Fixing production code bugs — edit the source in sdcm/ directly

Quick Reference: Test Infrastructure

Autouse Fixtures (Always Active)

These fixtures from unit_tests/unit/conftest.py run automatically for every unit test:

Important: AWS, GCE, and Azure HTTP calls are NOT auto-blocked. You must mock them per-test using unittest.mock.patch, patch.object, monkeypatch, or moto. Common functions to patch include convert_name_to_ami_if_needed, find_scylla_repo, get_arch_from_instance_type, and KeyStore methods. Use patch.object(KeyStore, "method_name", ...) for KeyStore since it's imported via from sdcm.keystore import KeyStore in 20+ modules.

On-Demand Fixtures (from parent unit_tests/conftest.py)

Request these by name in your test function signature:

Test Markers

Quick Reference: Mocking Patterns

Pattern 1: monkeypatch for Environment Variables

def test_config(monkeypatch):
    monkeypatch.setenv("SCT_CLUSTER_BACKEND", "aws")
    monkeypatch.setenv("SCT_AMI_ID_DB_SCYLLA", "ami-123")
    config = SCTConfiguration()
    assert config.get("cluster_backend") == "aws"

Pattern 2: unittest.mock.patch for Functions

from unittest.mock import patch, MagicMock

def test_s3_download():
    with patch("sdcm.utils.common._s3_download_file") as mock_dl:
        mock_dl.return_value = "/tmp/file.tar.gz"
        result = download_from_cloud("s3://bucket/file.tar.gz")
        assert result == "/tmp/file.tar.gz"

Pattern 3: moto for Full AWS Service Mocking

import boto3
from moto import mock_aws

@mock_aws
def test_ec2_provisioning():
    ec2 = boto3.client("ec2", region_name="us-east-1")
    ec2.run_instances(ImageId="ami-12345", MinCount=1, MaxCount=1)
    instances = ec2.describe_instances()
    assert len(instances["Reservations"]) == 1

Pattern 4: FakeRemoter for Remote Commands

import re
from invoke import Result

def test_node_command(fake_remoter):
    # fake_remoter is the FakeRemoter CLASS (not an instance).
    # Setting result_map here sets a class attribute, affecting all instances.
    fake_remoter.result_map = {
        re.compile(r"nodetool status"): Result(stdout="UN 10.0.0.1", exited=0),
        re.compile(r"cat /etc/scylla/scylla.yaml"): Result(stdout="cluster_name: test", exited=0),
    }
    # Code that calls node.remoter.run("nodetool status") will get the fake result

Pattern 5: monkeypatch for Attribute Replacement

def test_custom_behavior(monkeypatch):
    monkeypatch.setattr("sdcm.utils.common.S3Storage.download_file", lambda *a, **kw: "/fake/path")
    result = some_function_that_downloads()
    assert result == "/fake/path"

Debugging Unit Tests

Test Hangs or Is Slow

1. Unmocked network call. Add -s flag to see stdout and check for connection attempts:

   uv run python -m pytest unit_tests/test_module.py::test_function -v -s
   

2. Unmocked wait.wait_for loop. If code uses SCT's wait_for, mock it or reduce the timeout.

Test Passes Alone but Fails in Parallel

1. Shared global state. Use fixtures instead of module-level variables.
2. Unpinned environment variables. Use monkeypatch.setenv not os.environ.
3. Conflicting FakeRemoter.result_map. The result_map is a class attribute — set it per test, not globally.

FakeRemoter Raises ValueError

The error No fake result specified for command: <cmd> means the code under test runs a remote command that FakeRemoter.result_map doesn't know about. Add the missing command pattern to result_map.

Test Fails With Import Errors

SCT has many dependencies. If a test fails with ModuleNotFoundError, ensure:

1. Your virtualenv is active: uv sync
2. The import is at the top of the file, not inline

Test Naming Convention

Use the pattern test_<function>_<scenario>_<expected>:

# Good — describes behavior, condition, and expectation
def test_parse_version_invalid_string_returns_none(): ...
def test_config_missing_backend_raises_value_error(): ...
def test_health_check_single_node_failure_does_not_block_others(): ...

# Bad — generic, unclear what's being tested
def test_parse(): ...
def test_config_1(): ...
def test_it_works(): ...

Running Tests

# Run all unit tests (excludes integration)
uv run sct.py unit-tests

# Run a specific test file
uv run sct.py unit-tests -t unit/test_config.py

# Run a specific test function with verbose output
uv run python -m pytest unit_tests/unit/test_config.py::test_function_name -v -s

# Run with parallel execution disabled (for debugging)
uv run python -m pytest unit_tests/unit/test_config.py -v -s -n0

# Run with coverage report
uv run python -m pytest unit_tests/unit/ --cov=sdcm --cov-report=term-missing

# Run specific file with coverage for a single module
uv run python -m pytest unit_tests/unit/test_config.py --cov=sdcm.sct_config --cov-report=term-missing

Reference Index

Success Criteria

A well-written SCT unit test:

• Lives in unit_tests/unit/ with a test_*.py filename
• Uses pytest style (assert, fixtures, @pytest.mark.parametrize) — not unittest
• Does NOT have @pytest.mark.integration marker
• Makes zero real network calls (all external services mocked)
• Uses monkeypatch for environment variables, not os.environ
• Uses tmp_path for temporary files, not hardcoded paths
• Passes in isolation, in parallel, and in random order
• Has all imports at the top of the file
• Follows Google docstring format for test docstrings
• Passes uv run sct.py pre-commit checks