dotnet-testing-strategy（点网测试策略）

dotnet-testing-strategy

Decision framework for choosing the right test type, organizing test projects, and selecting test doubles in.NET applications. Covers unit vs integration vs E2E trade-offs with concrete criteria, naming conventions, and when to use mocks vs fakes vs stubs.

Out of scope: Test project scaffolding (directory layout, xUnit project creation, coverlet setup, editorconfig overrides) is owned by [skill:dotnet-add-testing]. Code coverage tooling and mutation testing are covered by [skill:dotnet-test-quality]. CI test reporting and pipeline integration -- see [skill:dotnet-gha-build-test] and [skill:dotnet-ado-build-test].

Prerequisites: Run [skill:dotnet-project-analysis] to understand the solution structure before designing a test strategy.

Cross-references: [skill:dotnet-xunit] for xUnit v3 testing framework features, [skill:dotnet-integration-testing] for WebApplicationFactory and Testcontainers patterns, [skill:dotnet-snapshot-testing] for Verify-based approval testing, [skill:dotnet-test-quality] for coverage and mutation testing, [skill:dotnet-add-testing] for test project scaffolding.

Test Type Decision Tree

Use this decision tree to determine which test type fits a given scenario. Start at the top and follow the first matching criterion.

Does the code under test depend on external infrastructure?
  (database, HTTP service, file system, message broker)
|
+-- YES --> Is the infrastructure behavior critical to correctness?
|           |
|           +-- YES --> Does it need the full application stack (middleware, auth, routing)?
|           |           |
|           |           +-- YES --> E2E / Functional Test
|           |           |           (WebApplicationFactory or Playwright)
|           |           |
|           |           +-- NO  --> Integration Test
|           |                       (WebApplicationFactory or Testcontainers)
|           |
|           +-- NO  --> Unit Test with test doubles
|                        (mock the infrastructure boundary)
|
+-- NO  --> Is this pure logic (calculations, transformations, validation)?
            |
            +-- YES --> Unit Test (no test doubles needed)
            |
            +-- NO  --> Unit Test with test doubles
                        (mock collaborator interfaces)

Concrete Criteria by Test Type

Cost-Benefit Guidance

• Prefer unit tests for business logic. They run fast, pinpoint failures precisely, and have no infrastructure requirements.
• Use integration tests to verify infrastructure boundaries work correctly. A repository unit test with a mocked DbContext proves nothing about actual SQL generation -- use a real database via Testcontainers.
• Use E2E tests sparingly for critical paths only. They are slow, brittle, and expensive to maintain. Cover the happy path and one or two critical failure scenarios.
• The testing pyramid is a guideline, not a rule. Some applications (CRUD APIs with minimal logic) benefit from more integration tests than unit tests. Match the strategy to the application's complexity profile.

Test Organization

Project Naming Convention

Mirror the src/ project structure under tests/ with a suffix indicating test type:

MyApp/
  src/
    MyApp.Domain/
    MyApp.Application/
    MyApp.Api/
    MyApp.Infrastructure/
  tests/
    MyApp.Domain.UnitTests/
    MyApp.Application.UnitTests/
    MyApp.Api.IntegrationTests/
    MyApp.Api.FunctionalTests/
    MyApp.Infrastructure.IntegrationTests/

• *.UnitTests -- isolated tests, no external dependencies
• *.IntegrationTests -- real infrastructure (database, HTTP, file system)
• *.FunctionalTests -- full application stack via WebApplicationFactory

See [skill:dotnet-add-testing] for creating these projects with proper package references and build configuration.

Test Class Organization

One test class per production class. Place test files in a namespace that mirrors the production namespace:

// Production: src/MyApp.Domain/Orders/OrderService.cs
// Test:       tests/MyApp.Domain.UnitTests/Orders/OrderServiceTests.cs
namespace MyApp.Domain.UnitTests.Orders;

public class OrderServiceTests
{
    // Group by method, then by scenario
}

For large production classes, split test classes by method:

// OrderService_CreateTests.cs
// OrderService_CancelTests.cs
// OrderService_RefundTests.cs

Test Naming Conventions

Use the Method_Scenario_ExpectedBehavior pattern. This reads naturally in test explorer output and makes failures self-documenting:

public class OrderServiceTests
{
    [Fact]
    public void CalculateTotal_WithDiscountCode_AppliesPercentageDiscount()
    {
        // ...
    }

    [Fact]
    public void CalculateTotal_WithExpiredDiscount_ThrowsInvalidOperationException()
    {
        // ...
    }

    [Fact]
    public async Task SubmitOrder_WhenInventoryInsufficient_ReturnsOutOfStockError()
    {
        // ...
    }
}

Alternative naming styles (choose one per project and stay consistent):

Arrange-Act-Assert Pattern

Every test follows the AAA structure. Keep each section clearly separated:

[Fact]
public async Task CreateOrder_WithValidItems_PersistsAndReturnsOrder()
{
    // Arrange
    var repository = new FakeOrderRepository();
    var service = new OrderService(repository);
    var request = new CreateOrderRequest
    {
        CustomerId = "cust-123",
        Items = [new OrderItem("SKU-001", Quantity: 2, UnitPrice: 29.99m)]
    };

    // Act
    var result = await service.CreateAsync(request);

    // Assert
    Assert.NotNull(result);
    Assert.Equal("cust-123", result.CustomerId);
    Assert.Single(result.Items);
    Assert.True(repository.SavedOrders.ContainsKey(result.Id));
}

Guideline: If you cannot clearly label the three sections, the test may be doing too much. Split into multiple tests.

Test Doubles: When to Use What

Terminology

Decision Guidance

Do you need to verify HOW a dependency was called?
|
+-- YES --> Do you need a working implementation too?
|           |
|           +-- YES --> Spy (record calls on a fake)
|           +-- NO  --> Mock (NSubstitute / Moq)
|
+-- NO  --> Do you need the dependency to DO something realistic?
            |
            +-- YES --> Fake (in-memory implementation)
            +-- NO  --> Stub (return canned values)

Example: Stub vs Mock vs Fake

// STUB: Returns canned data -- verifying the code under test's logic
var priceService = Substitute.For<IPriceService>();
priceService.GetPriceAsync("SKU-001").Returns(29.99m);  // canned return

var total = await calculator.CalculateTotalAsync(items);
Assert.Equal(59.98m, total);  // assert on the result, not the call

// MOCK: Verifies interaction -- ensuring a side effect happened
var emailSender = Substitute.For<IEmailSender>();

await orderService.CompleteAsync(order);

await emailSender.Received(1).SendAsync(             // assert on the call
    Arg.Is<string>(to => to == order.CustomerEmail),
    Arg.Any<string>(),
    Arg.Any<string>());

// FAKE: In-memory implementation -- realistic behavior without infrastructure
public class FakeOrderRepository : IOrderRepository
{
    public Dictionary<Guid, Order> Orders { get; } = new();

    public Task<Order?> GetByIdAsync(Guid id, CancellationToken ct = default)
        => Task.FromResult(Orders.GetValueOrDefault(id));

    public Task SaveAsync(Order order, CancellationToken ct = default)
    {
        Orders[order.Id] = order;
        return Task.CompletedTask;
    }
}

When to Prefer Fakes Over Mocks

• Domain-heavy applications: Fakes give more realistic behavior for complex interactions. An in-memory repository catches bugs that mocks miss (e.g., duplicate key violations).
• Overuse of mocks is a test smell. If a test has more mock setup than actual assertions, consider whether a fake would be clearer and more maintainable.
• Integration boundaries are better tested with real infrastructure via [skill:dotnet-integration-testing] than with mocks. A mocked DbContext does not verify that your LINQ translates to valid SQL.

Testing Anti-Patterns

1. Testing Implementation Details

// BAD: Breaks when refactoring internals
repository.Received(1).GetByIdAsync(Arg.Is<Guid>(id => id == orderId));
repository.Received(1).SaveAsync(Arg.Any<Order>());
// ... five more Received() calls verifying the exact call sequence

// GOOD: Test the observable outcome
var result = await service.ProcessAsync(orderId);
Assert.Equal(OrderStatus.Completed, result.Status);

2. Excessive Mock Setup

// BAD: Mock setup is longer than the actual test
var repo = Substitute.For<IOrderRepository>();
var pricing = Substitute.For<IPricingService>();
var inventory = Substitute.For<IInventoryService>();
var shipping = Substitute.For<IShippingService>();
var notification = Substitute.For<INotificationService>();
var audit = Substitute.For<IAuditService>();
// ... 20 lines of .Returns() setup

// BETTER: Use a builder or fake that encapsulates setup
var fixture = new OrderServiceFixture()
    .WithOrder(testOrder)
    .WithPrice("SKU-001", 29.99m);
var result = await fixture.Service.ProcessAsync(testOrder.Id);

3. Non-Deterministic Tests

Tests must not depend on system clock, random values, or external network. Inject abstractions:

// BAD: Uses DateTime.UtcNow directly
public bool IsExpired() => ExpiresAt < DateTime.UtcNow;

// GOOD: Inject TimeProvider (.NET 8+)
public bool IsExpired(TimeProvider time) => ExpiresAt < time.GetUtcNow();

// In test
var fakeTime = new FakeTimeProvider(new DateTimeOffset(2025, 6, 15, 0, 0, 0, TimeSpan.Zero));
Assert.True(order.IsExpired(fakeTime));

Key Principles

• Test behavior, not implementation. Assert on observable outcomes (return values, state changes, published events), not internal method calls.
• One logical assertion per test. Multiple Assert calls are fine if they verify one logical concept (e.g., all properties of a returned object). Multiple unrelated assertions indicate the test should be split.
• Keep tests independent. No test should depend on another test's execution or ordering. Use fresh fixtures for each test.
• Name tests so failures are self-documenting. A failing test name should tell you what broke without reading the test body.
• Match test type to risk. High-risk code (payments, auth) deserves integration and E2E coverage. Low-risk code (simple mapping) needs only unit tests.
• Use TimeProvider for time-dependent logic (.NET 8+). It is the framework-provided abstraction; do not create custom IClock interfaces.

Agent Gotchas

1. Do not mock types you do not own. Mocking HttpClient, DbContext, or framework types leads to brittle tests that do not reflect real behavior. Use WebApplicationFactory or Testcontainers instead -- see [skill:dotnet-integration-testing].
2. Do not create test projects without checking for existing structure. Run [skill:dotnet-project-analysis] first; duplicating test infrastructure causes build conflicts.
3. Do not use Thread.Sleep in tests. Use Task.Delay with a cancellation token, or better, use FakeTimeProvider.Advance() to control time deterministically.
4. Do not test private methods directly. If a private method needs its own tests, it should be extracted into its own class. Test through the public API.
5. Do not hard-code connection strings in integration tests. Use Testcontainers for disposable infrastructure or WebApplicationFactory for in-process testing -- see [skill:dotnet-integration-testing].

References

• .NET Testing Best Practices
• Unit testing C# with xUnit
• Integration tests in ASP.NET Core
• NSubstitute documentation
• TimeProvider in.NET 8