Integration Testing

Integration Testing

Intro

Integration tests verify that your code works against real collaborators — databases, APIs, queues — not just mocks. The trick is keeping them fast, isolated, and reproducible. Disposable containers, per-test fixtures, and strict API mocks are the primary tools.

Overview

Testcontainers

Spin up real services as disposable containers for the duration of the test suite:

• Each suite gets its own container — no shared state.
• Use fixed image tags (not latest) for reproducibility.
• Set health checks or wait strategies so tests start only after the service is ready.
• Tear down after the suite (most libraries handle this via RAII or afterAll).

Supported in Python (testcontainers), Rust (testcontainers), JS/TS (testcontainers), Java (testcontainers), and Go (testcontainers-go).

Database fixtures

Three core strategies:

• Factories — generate objects with sensible defaults; override only what matters per test.
• Fixtures — predefined datasets loaded before tests; good for read-heavy tests.
• Seeds — baseline data loaded once for the suite; reset per test via transactions.

Cleanup approaches (pick one per project):

Default to transaction rollback. Fall back to truncation when you need multiple connections (e.g. concurrent access tests). Use disposable containers for full isolation in CI.

API mocking

Mock external HTTP APIs so integration tests run without network dependencies:

Best practices:

• Assert on what the code sent (method, headers, body), not just the response.
• Use strict mode — fail on any unexpected request.
• Record real API responses once and replay in CI (contract tests).

Snapshot and E2E

Snapshot tests capture function or component output and compare against a stored baseline. Good for serialized JSON, rendered HTML, CLI output, and error messages. Update intentionally (--update-snapshots / -u) and review diffs in PRs. Store snapshots next to tests in version control. Normalize volatile data (timestamps, random IDs) before snapshotting.

For E2E:

• Page Object Model — encapsulate page interactions; tests read like user stories.
• Arrange-Act-Assert — set up state, perform the action, verify the outcome.
• Test user journeys, not individual pages. One E2E test should cover a meaningful flow.
• Use dedicated test accounts and isolated environments. Never test against production data.

CI isolation and flaky tests

• Run integration tests in a separate CI job with longer timeouts.
• Use testcontainers or Docker Compose for backing services.
• Give each parallel worker a unique database/schema.
• Set per-test timeouts (not just per-job) to catch hangs.
• Cache container images in CI to speed up startup.

Common flaky-test causes and fixes:

• Shared state — per-test databases or transaction rollback.
• Timing/race conditions — replace sleep with explicit wait-for conditions.
• Non-deterministic ordering — randomize test order in CI to surface hidden dependencies.
• Network flakiness — mock all external APIs.
• Resource exhaustion — limit parallelism, close connections.

A flaky test that is ignored is worse than no test at all. Tag known flaky tests and track them.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• Sharing a single database across parallel test workers. Parallel workers reading and writing to the same schema produces ordering-dependent failures and data contamination. Give each worker a unique database or schema, or use per-test transaction rollback.
• Using sleep() to wait for async behavior. Fixed delays are both fragile (fail on slow CI runners) and wasteful (over-wait on fast machines). Replace with explicit wait-for conditions that poll until the expected state is reached or a timeout fires.
• Mocking the unit under test instead of its boundary. If the code being tested is itself replaced by a mock, the test verifies nothing. Mocks belong at the boundary — the external API, database, or queue — not inside the module under test.
• Asserting only on response shape, never on the request the code sent. Testing that "we got a 200" but not verifying the request body, headers, or method means outbound contract violations go undetected. Assert on what the code sent, not just what it received.
• Snapshotting timestamps or random IDs without normalization. Snapshots that contain created_at: 2026-04-08T10:23:11Z or id: "f3a2..." fail on every run. Normalize or mask volatile fields before snapshotting.
• Marking flaky tests as skipped without assigning an owner. A skipped test is a test that never catches anything. Tag the test with the owner's handle and a follow-up date; skipped tests must be fixed before a release.
• Using latest image tags for testcontainers. When an upstream container image is updated, CI breaks unpredictably. Pin testcontainer images to fixed tags (postgres:16-alpine, not postgres:latest) for reproducible CI runs.

Full reference

Testcontainers with Postgres (Python)

import pytest
from testcontainers.postgres import PostgresContainer
from sqlalchemy import create_engine, text

@pytest.fixture(scope="module")
def db_engine():
    with PostgresContainer("postgres:16-alpine") as pg:
        engine = create_engine(pg.get_connection_url())
        with engine.begin() as conn:
            conn.execute(text("CREATE TABLE users (id SERIAL, name TEXT)"))
        yield engine

def test_insert_and_query_user(db_engine):
    with db_engine.begin() as conn:
        conn.execute(text("INSERT INTO users (name) VALUES (:n)"), {"n": "Alice"})
        result = conn.execute(text("SELECT name FROM users")).fetchone()
    assert result[0] == "Alice"

API mocking with MSW (TypeScript)

import { http, HttpResponse } from "msw";
import { setupServer } from "msw/node";
import { fetchWeather } from "../src/weather";

const server = setupServer(
  http.get("https://api.weather.example/current", ({ request }) => {
    const url = new URL(request.url);
    if (url.searchParams.get("city") === "Berlin") {
      return HttpResponse.json({ temp: 18, condition: "cloudy" });
    }
    return new HttpResponse(null, { status: 404 });
  })
);

beforeAll(() => server.listen({ onUnhandledRequest: "error" }));
afterEach(() => server.resetHandlers());
afterAll(() => server.close());

test("returns weather for known city", async () => {
  const weather = await fetchWeather("Berlin");
  expect(weather.temp).toBe(18);
});

httpmock (Rust)

use httpmock::prelude::*;
use reqwest;

#[tokio::test]
async fn test_client_sends_auth_header() {
    let server = MockServer::start();
    let mock = server.mock(|when, then| {
        when.method(GET)
            .path("/api/data")
            .header("Authorization", "Bearer test-token");
        then.status(200)
            .json_body_obj(&serde_json::json!({"items": [1, 2, 3]}));
    });

    let client = ApiClient::new(&server.base_url(), "test-token");
    let result = client.fetch_data().await.unwrap();

    assert_eq!(result.items, vec![1, 2, 3]);
    mock.assert();
}

Fixture patterns

For deeper fixture and factory patterns across Python, TypeScript, and Rust, see references/test-fixtures.md. Key takeaways:

• pytest fixtures — use scope="session" for an engine and scope="function" with transaction rollback for a per-test session. factory_boy handles complex object graphs via SubFactory.
• TypeScript — pair beforeEach truncation with builder functions or a Builder class for complex fixtures. Order truncations to respect foreign keys.
• Rust — implement a TestUserBuilder with .insert(pool) or .insert_tx(&mut tx) methods. #[sqlx::test] plus per-test transactions gives automatic rollback on drop.

Database seeding strategies

• Minimal seed (recommended) — each test creates only the data it needs through factories. Tests stay independent and easy to read.
• Shared baseline seed — load reference data (countries, roles) once per suite; tests add their own records on top. Good for read-heavy suites.
• Snapshot seed — dump a known-good database state to a SQL file or container image and restore before each suite. Fastest for large datasets but harder to maintain.

Anti-patterns

• Sharing a single database across parallel test workers
• Using sleep() to wait for async behavior
• Mocking the unit under test instead of its boundary
• Asserting only on response shape, never on the request the code sent
• Snapshotting timestamps or random IDs without normalization
• Marking flaky tests as skipped without an owner or follow-up