Locust best practices

Locust Scenario Builder

Enforces a consistent, production-ready pattern for Locust load test scripts in Python, covering HTTP/REST APIs with realistic user behavior, custom load shapes, and proper response validation.

Output Format

When producing or fixing a locustfile, always deliver three things:

1. A complete, runnable locustfile — never a partial snippet.
2. The exact run command with all flags needed (--headless, -u, -r, -t, --host).
3. A one-line explanation of the wait_time strategy chosen and why it fits the load goal.

---

Step 1 — Gather Context

Ask only what is unknown:

• Goal: New locustfile · Fix existing locustfile · Specific API question
• Protocol: HTTP/REST (default) · WebSocket · gRPC
• Workload: Fixed concurrent users or fixed arrival rate (RPS)?
• Auth: None · Login on start · Token from environment variable?
• Data: Static · CSV file · Generated per user?

Only load references/SHAPES.md when the user asks about custom load profiles, multi-stage ramps, spike simulation, or LoadTestShape.

Only load references/PROTOCOLS.md when the user mentions WebSocket, gRPC, or testing non-HTTP protocols. Contains custom client wrappers, events.request.fire() pattern, and protocol compatibility notes.

Only load references/DESIGN-PATTERNS.md when the user asks about project structure, multiple user classes, TaskSet, shared auth helpers, CSV data loading at scale, modular locustfiles, or multi-file organization.

Only load references/RUNNING.md when the user asks about headless mode, CI/CD integration, exit codes, SLA-based build failures, distributed testing, master/worker setup, multiple cores, logging to file, or custom log levels.

---

Step 2 — Apply the 4-Block Pattern

Every locustfile must follow this structure. Generate the complete skeleton first, then fill in details.

Block 1 → Imports + config    locust imports, constants, env vars
Block 2 → User class          HttpUser with host, wait_time, on_start, @task methods
Block 3 → Shape (optional)    LoadTestShape for custom ramp profiles
Block 4 → Run command         locust CLI with all flags as a comment at the bottom

---

Step 3 — HttpUser Rules

3.1 Always use catch_response=True for validation

Never rely on HTTP status alone. Always validate the response body:

# Wrong — only checks HTTP status, misses business logic errors
self.client.get("/api/users")

# Correct — validates response content
with self.client.get("/api/users", catch_response=True) as response:
    if response.status_code != 200:
        response.failure(f"Expected 200, got {response.status_code}")
    elif "users" not in response.json():
        response.failure("Missing 'users' key in response")

3.2 Group dynamic URLs with name=

Without grouping, each unique URL creates a separate stats entry — flooding the report:

# Wrong — creates N separate stats rows
for user_id in range(100):
    self.client.get(f"/user/{user_id}")

# Correct — groups all under one stats entry
for user_id in range(100):
    self.client.get(f"/user/{user_id}", name="/user/[id]")

3.3 Use on_start for auth — never share tokens across users

Each virtual user must authenticate independently:

# Wrong — shared token is a race condition and not realistic
shared_token = None

class MyUser(HttpUser):
    def on_start(self):
        global shared_token  # ❌ shared state across users

# Correct — each user gets its own token
class MyUser(HttpUser):
    def on_start(self):
        response = self.client.post("/auth/login", json={
            "username": os.environ["TEST_USER"],
            "password": os.environ["TEST_PASSWORD"],
        })
        self.token = response.json()["access_token"]

    @task
    def get_profile(self):
        self.client.get("/profile", headers={"Authorization": f"Bearer {self.token}"})

3.4 Choose wait_time based on the load goal

Goal	wait_time	Why
Simulate real users with think time	between(1, 5)	Realistic pacing
Maintain fixed RPS regardless of latency	constant_throughput(2)	Open model — 2 tasks/sec per user
Reproduce exact user pacing	constant(1)	Fixed interval

Never use wait_time = constant(0) — zero think time generates 10–100× more load than real users and produces unrealistic results.

3.5 Use @task weight to reflect real traffic distribution

class ShopUser(HttpUser):
    wait_time = between(1, 3)

    @task(10)   # 10× more common — most users just browse
    def browse_products(self):
        self.client.get("/products")

    @task(3)    # 3× — some users search
    def search(self):
        self.client.get("/search?q=shoes", name="/search")

    @task(1)    # 1× — few users actually purchase
    def checkout(self):
        self.client.post("/checkout", json={"cart_id": self.cart_id})

---

Step 4 — Standard Locustfile Template

import os
import json
from locust import HttpUser, task, between, events

# ---------------------------------------------------------------------------
# Block 1 — Config
# ---------------------------------------------------------------------------
BASE_URL = os.environ.get("TARGET_HOST", "http://localhost:8080")

# ---------------------------------------------------------------------------
# Block 2 — User class
# ---------------------------------------------------------------------------
class ApiUser(HttpUser):
    host = BASE_URL
    wait_time = between(1, 3)

    def on_start(self):
        """Runs once per virtual user before any task executes."""
        with self.client.post(
            "/auth/login",
            json={
                "username": os.environ["TEST_USER"],
                "password": os.environ["TEST_PASSWORD"],
            },
            catch_response=True,
        ) as response:
            if response.status_code == 200:
                self.token = response.json()["access_token"]
            else:
                response.failure(f"Login failed: {response.status_code}")
                self.token = None

    def _auth_headers(self):
        return {"Authorization": f"Bearer {self.token}"}

    @task(3)
    def get_items(self):
        with self.client.get(
            "/api/items",
            headers=self._auth_headers(),
            catch_response=True,
        ) as response:
            if response.status_code != 200:
                response.failure(f"GET /api/items failed: {response.status_code}")

    @task(1)
    def create_item(self):
        with self.client.post(
            "/api/items",
            json={"name": "test-item", "value": 42},
            headers=self._auth_headers(),
            catch_response=True,
        ) as response:
            if response.status_code not in (200, 201):
                response.failure(f"POST /api/items failed: {response.status_code}")

# ---------------------------------------------------------------------------
# Block 3 — Events (optional)
# ---------------------------------------------------------------------------
@events.test_start.add_listener
def on_test_start(environment, **kwargs):
    print("Test started — verifying environment variables...")
    for var in ["TEST_USER", "TEST_PASSWORD"]:
        if not os.environ.get(var):
            raise ValueError(f"Missing required env var: {var}")

# ---------------------------------------------------------------------------
# Run command (headless):
# TEST_USER=alice TEST_PASSWORD=secret locust -f locustfile.py \
#   --headless -u 50 -r 5 -t 5m --host http://your-api.com \
#   --html report.html --csv results
# ---------------------------------------------------------------------------

---

Step 5 — Common Mistakes

1. No catch_response=True — silent failures

Without it, a 500 response with a valid HTTP connection still counts as success in Locust's stats. Always use catch_response=True and call response.failure() explicitly.

2. Missing name= on dynamic URLs

Every unique URL path creates a new row in the stats table. Parameterized paths like /user/123, /user/456 explode the report. Use name="/user/[id]".

3. Shared mutable state between users

Locust users run in greenlets (gevent), not threads, but shared state still causes race conditions in logic. Always store per-user state (self.token, self.cart_id) as instance attributes.

When loading credentials from CSV, store all fields as instance attributes in on_start — not just one:

# Wrong — only stores username, password is lost
self.username = row["username"]

# Correct — store every field the user will need
self.username = row["username"]
self.password = row["password"]

4. wait_time = constant(0) or no wait_time

Omitting think time generates unrealistic load and will saturate your test machine before the target system. Always define a meaningful wait_time.

5. Hardcoded host and credentials

Never hardcode host, usernames, or passwords. Always use environment variables:

host = os.environ.get("TARGET_HOST", "http://localhost:8080")

6. Asserting only status code, not response body

A 200 with {"error": "token expired"} will pass a status-only check. Always validate the response body for business logic correctness.

7. Using FastHttpUser without understanding the tradeoff

FastHttpUser uses geventhttpclient instead of requests. It is faster but does not support all requests features (cookies jar, some auth helpers). Only use it when you have proven a throughput bottleneck in the test runner itself, not the target system.

---

Step 6 — Run Commands Reference

# Basic headless run — 50 users, 5/sec ramp, 5 minutes
locust -f locustfile.py --headless -u 50 -r 5 -t 5m --host https://api.example.com

# With HTML report and CSV output
locust -f locustfile.py --headless -u 100 -r 10 -t 10m \
  --host https://api.example.com \
  --html report.html --csv results

# Run specific user class only
locust -f locustfile.py --headless -u 50 -r 5 -t 5m \
  --host https://api.example.com ApiUser

# With environment variables
TARGET_HOST=https://api.example.com \
TEST_USER=alice \
TEST_PASSWORD=secret \
locust -f locustfile.py --headless -u 50 -r 5 -t 5m

# Open web UI (no --headless) — useful for interactive exploration
locust -f locustfile.py --host https://api.example.com

Flag reference:

Flag	Meaning
-u	Peak number of concurrent users
-r	Spawn rate — users added per second
-t	Total run time (e.g. 30s, 5m, 2h)
--headless	No web UI — run and exit automatically
--host	Target base URL
--html	Generate HTML report at path
--csv	Generate CSV stats files at prefix
--only-summary	Suppress per-interval stats output