Files
conti-docs/backend/10-testing.md
T
Guangfei.Zhao 8c0fcd84e8 feat: Enhance documentation on layering, object naming conventions, and API design
- Added object naming conventions (PO/DAO/BO/DTO/VO) in 02-layering.md to clarify terminology and usage within the team.
- Updated 06-api-design.md to include MapStruct for DTO and entity conversion, providing examples and configuration details.
- Expanded 07-config-governance.md with local development instructions and strategies for running without K8s, including two recommended approaches.
- Included K8s probe configuration details in 08-observability.md for liveness and readiness checks.
- Clarified CI/CD processes in 09-build-deploy.md, detailing environment distinctions and deployment strategies for local, Dev, UAT, and Prod.
- Introduced ArchUnit for architectural testing in 10-testing.md, ensuring adherence to defined layering rules and coverage verification with Jacoco.
2026-08-13 15:19:05 +08:00

11 KiB
Raw Blame History

10. 测试策略

决策

JUnit 5 + MockK 做单元测试,Testcontainers 做集成测试,WireMock 做外部依赖打桩,分层对应 02-layering.md

分层测试策略

  • domain 层(有的话):纯单元测试,不起 Spring 容器,mock 掉 repository/client 接口,验证业务规则本身。
  • application 层@SpringBootTest 或轻量 slice testmock 掉 infrastructure 层。
  • infrastructure 层:用 Testcontainers 起真实数据库跑 repository 测试,避免 H2 和生产数据库行为差异导致的假通过。
  • api 层@WebMvcTest 验证参数校验、异常处理、响应结构是否符合 06-api-design.md 的约定。
  • f6-integration / mini-clients:对外部依赖用 WireMock 打桩,覆盖超时/重试/熔断路径。

domain 层单元测试示例(对应 02 里的换票场景)

// domains/webview-ticket/src/test/kotlin/.../IssueWebviewTicketServiceTest.kt
class IssueWebviewTicketServiceTest {
    private val ticketRepository = mockk<WebviewTicketRepository>()
    private val fixedClock = Clock.fixed(Instant.parse("2026-01-01T00:00:00Z"), ZoneOffset.UTC)
    private val service = IssueWebviewTicketService(ticketRepository, fixedClock)

    @Test
    fun `复用未过期的已签发票据`() {
        val existing = WebviewTicket("t1", storeId = 1, userId = 1, status = TicketStatus.ISSUED,
            expiresAt = fixedClock.instant().plusSeconds(60))
        every { ticketRepository.findActiveTicket(1, 1) } returns existing

        val result = service.issue(userId = 1, storeId = 1)

        assertEquals("t1", result.ticketId)
        verify(exactly = 0) { ticketRepository.save(any()) } // 复用场景不应该重新签发
    }

    @Test
    fun `没有有效票据时签发新票据`() {
        every { ticketRepository.findActiveTicket(1, 1) } returns null
        every { ticketRepository.save(any()) } just Runs

        val result = service.issue(userId = 1, storeId = 1)

        assertEquals(TicketStatus.ISSUED, result.status)
        verify(exactly = 1) { ticketRepository.save(any()) }
    }
}

不起 Spring 容器、不连数据库,纯 JVM 内存跑完,这类测试应该是数量最多、跑得最快的一层。

infrastructure 层集成测试示例(Testcontainers

// domains/identity-store/src/test/kotlin/.../StoreJpaRepositoryTest.kt
@DataJpaTest
@Testcontainers
class StoreJpaRepositoryTest {

    companion object {
        @Container
        @JvmStatic
        val postgres = PostgreSQLContainer("postgres:16")

        @JvmStatic
        @DynamicPropertySource
        fun props(registry: DynamicPropertyRegistry) {
            registry.add("spring.datasource.url", postgres::getJdbcUrl)
            registry.add("spring.datasource.username", postgres::getUsername)
            registry.add("spring.datasource.password", postgres::getPassword)
        }
    }

    @Autowired lateinit var repository: StoreJpaRepository

    @Test
    fun `按 code 查询门店`() {
        repository.save(StoreEntity(name = "示例门店", code = "S001", status = StoreStatus.ACTIVE))

        val found = repository.findByCode("S001")

        assertNotNull(found)
        assertEquals("示例门店", found?.name)
    }
}

用真实 PostgreSQL(而不是 H2)跑测试,是因为 Flyway migration 里的 SQL 方言、JPA 对特定数据库函数的行为,H2 不一定能完全模拟,容易出现"H2 测试通过、生产环境报错"的假阳性。

api 层测试示例(@WebMvcTest

@WebMvcTest(StoreController::class)
class StoreControllerTest {
    @Autowired lateinit var mockMvc: MockMvc
    @MockkBean lateinit var storeAppService: StoreAppService

    @Test
    fun `切换门店参数为空时返回参数校验错误`() {
        mockMvc.post("/api/v1/stores/switch") {
            contentType = MediaType.APPLICATION_JSON
            content = """{}"""
        }.andExpect {
            status { isBadRequest() }
            jsonPath("$.code") { value("INVALID_PARAM") } // 对应 06-api-design.md 的 ApiResult 结构
        }
    }
}

外部依赖打桩示例(WireMock,覆盖 F6 超时/熔断路径)

@Testcontainers
class F6ApiClientResilienceTest {
    companion object {
        @Container
        @JvmStatic
        val wireMock = WireMockContainer("wiremock/wiremock:3.9.1")
    }

    @Test
    fun `F6 响应超时后走 fallback 返回降级数据`() {
        wireMock.stubFor(
            get(urlPathEqualTo("/f6/procurement/list"))
                .willReturn(aResponse().withFixedDelay(5000).withStatus(200)) // 模拟超过 timeout 配置的慢响应
        )

        val result = f6ApiClient.fetchProcurementList(storeId = 1).block()

        assertTrue(result!!.degraded) // 验证 05-integration-layer.md 里配置的 2s 超时 + fallback 生效
    }
}

ArchUnit:把分层规则变成可执行的测试

01-project-structure.md02-layering.md 里定的依赖方向规则(domain 不依赖 Spring/JPA、domains 之间不互相依赖、api 不直接依赖 infrastructure),光靠 code review 肉眼盯着容易漏,模块一多更盯不过来。用 ArchUnit 把这些规则写成测试,每次构建自动检查:

// build.gradle(专门放架构测试的模块,或加进 bootstrap 的 test 依赖)
testImplementation 'com.tngtech.archunit:archunit-junit5:1.3.0'
// architecture-test/src/test/kotlin/.../LayeringRulesTest.kt
class LayeringRulesTest {
    private val classes = ClassFileImporter().importPackages("com.continental.retailapp")

    @Test
    fun `domain 层不能依赖 Spring 或 JPA`() {
        noClasses()
            .that().resideInAPackage("..domain..")
            .should().dependOnClassesThat().resideInAnyPackage("org.springframework..", "jakarta.persistence..")
            .check(classes)
    }

    @Test
    fun `domains 之间不能互相依赖(bff-orchestration 除外)`() {
        slices()
            .matching("com.continental.retailapp.(*)..")
            .should().notDependOnEachOther()
            .ignoreDependency(
                DescribedPredicate.describe("来自 bff-orchestration") { it.resideInAPackage("..bffOrchestration..") },
                DescribedPredicate.alwaysTrue(),
            ) // bff-orchestration 允许依赖多个 domain,是唯一的例外,见 02-layering.md
            .check(classes)
    }

    @Test
    fun `api 层不能直接依赖 infrastructure 层`() {
        noClasses()
            .that().resideInAPackage("..api..")
            .should().dependOnClassesThat().resideInAPackage("..infrastructure..")
            .check(classes)
    }
}

这类测试一次写好,覆盖的是全代码库范围的结构性规则,跑在 CI Validation 阶段(见 09-build-deploy.md),比"等 code review 时人工发现某个 domain 偷偷 import 了另一个 domain 的 entity"要可靠得多,而且不区分改动大小——哪怕只加了一行 import,只要违反规则就会立刻挂红。

CI 里跑 Testcontainers 的前提条件

infrastructure 层的集成测试(前面 StoreJpaRepositoryTest 那个例子)依赖 Testcontainers 起真实容器,这要求执行 ./gradlew test 的 GitLab Runner 本身能起 Docker 容器,不是随便一个 Runner 都能跑,两种常见配置:

方式一:Docker-in-Dockerdind),托管 Runner 的默认选择

# .gitlab-ci.yml
unit-integration-test:
  stage: validate
  image: eclipse-temurin:21-jdk
  services:
    - docker:24-dind
  variables:
    DOCKER_HOST: tcp://docker:2375
    DOCKER_TLS_CERTDIR: ""
  script:
    - ./gradlew test --no-daemon

方式二:挂载宿主机 Docker socket,适合 09-build-deploy.md 里那种自建 self-hosted Runnerazure-vnet-runner

# GitLab Runner 的 config.toml
[[runners]]
  [runners.docker]
    privileged = true
    volumes = ["/var/run/docker.sock:/var/run/docker.sock", "/cache"]

方式二没有 dind 的嵌套虚拟化开销,跑起来更快,但要求 Runner 能访问宿主机的 Docker socket(等价于 Runner 对宿主机有较高权限),只适合放在我们自己管控的 self-hosted Runner 上;如果以后接入公共共享 Runner 跑这一类测试,只能用方式一,不应该在共享 Runner 上开 Docker socket 权限。

覆盖率门禁(Jacoco

// build.gradle
plugins {
    id 'jacoco'
}

jacocoTestReport {
    dependsOn test
    reports {
        xml.required = true   // CI 里给 GitLab 覆盖率可视化用
    }
}

jacocoTestCoverageVerification {
    violationRules {
        rule {
            limit {
                minimum = 0.70
            }
        }
    }
}

check.dependsOn jacocoTestCoverageVerification   // 覆盖率不达标,./gradlew check 直接失败
# .gitlab-ci.ymlCI Validation 阶段追加覆盖率门禁
unit-integration-test:
  stage: validate
  script:
    - ./gradlew test jacocoTestCoverageVerification --no-daemon
  artifacts:
    reports:
      coverage_report:
        coverage_format: cobertura
        path: '**/build/reports/jacoco/test/jacocoTestReport.xml'

阈值定在 70% 而不是追求 90%+:domain 层因为纯逻辑、mock 成本低,覆盖率天然会很高;infrastructure/api 层的诉求是"关键路径别漏测"而不是"每一行都要覆盖",统一定一个较低的整体阈值,作用是拦住完全没写测试就合入的代码,而不是逼着每个模块都卷到很高的数字——那样容易导致为了凑覆盖率写没有意义的测试。

附录:为什么 domain 层用 mock、infrastructure 层坚持用真实依赖

这是测试金字塔的实际落地取舍:越往下层(domain)测试数量应该越多、跑得越快,因为业务规则的分支组合往往很多(各种边界条件),用 mock 把依赖都隔离掉才能便宜地把每个分支都测到;越往上/往基础设施层,测试数量应该越少但真实度要求越高,因为这一层要验证的恰恰是"我们对某个具体技术(JPA、真实数据库、真实 HTTP 依赖)的假设是否成立"——如果这一层也用 mock,等于假设了"这个假设是对的",那测试就失去了意义。

Testcontainers 和 WireMock 的共同点是:它们让"跑得慢、需要真实环境"的测试仍然可以在 CI 里可重复地跑起来(每次测试起一个全新的容器,跑完销毁,不依赖某个共享的、状态可能被污染的测试环境)。

待补充

  • 是否需要和 APP 端做端到端契约测试(比如引入 Pact)。

参考链接