Opinionated list of best practices and conventions I use in startups.
For the last several years in production, we have been making good and bad decisions that impacted our developer experience dramatically. Some of them are worth sharing.
- Project Structure
- Async Routes
- Pydantic
- Dependencies
- Miscellaneous
- Follow the REST
- FastAPI response serialization
- If you must use sync SDK, then run it in a thread pool.
- ValueErrors might become Pydantic ValidationError
- Docs
- Set DB keys naming conventions
- Migrations. Alembic
- Set DB naming conventions
- SQL-first. Pydantic-second
- Set tests client async from day 0
- Use ruff
- Bonus Section
There are many ways to structure a project, but the best structure is one that is consistent, straightforward, and free of surprises.
Many example projects and tutorials divide the project by file type (e.g., crud, routers, models), which works well for microservices or projects with fewer scopes. However, this approach didn't fit our monolith with many domains and modules.
The structure I found more scalable and evolvable for these cases is inspired by Netflix's Dispatch, with some minor modifications.
fastapi-project
├── alembic/
├── src
│ ├── auth
│ │ ├── router.py
│ │ ├── schemas.py # pydantic models
│ │ ├── models.py # db models
│ │ ├── dependencies.py
│ │ ├── config.py # local configs
│ │ ├── constants.py
│ │ ├── exceptions.py
│ │ ├── service.py
│ │ └── utils.py
│ ├── aws
│ │ ├── client.py # client model for external service communication
│ │ ├── schemas.py
│ │ ├── config.py
│ │ ├── constants.py
│ │ ├── exceptions.py
│ │ └── utils.py
│ └── posts
│ │ ├── router.py
│ │ ├── schemas.py
│ │ ├── models.py
│ │ ├── dependencies.py
│ │ ├── constants.py
│ │ ├── exceptions.py
│ │ ├── service.py
│ │ └── utils.py
│ ├── config.py # global configs
│ ├── models.py # global models
│ ├── exceptions.py # global exceptions
│ ├── pagination.py # global module e.g. pagination
│ ├── database.py # db connection related stuff
│ └── main.py
├── tests/
│ ├── auth
│ ├── aws
│ └── posts
├── templates/
│ └── index.html
├── requirements
│ ├── base.txt
│ ├── dev.txt
│ └── prod.txt
├── .env
├── .gitignore
├── logging.ini
└── alembic.ini
- Store all domain directories inside
src
foldersrc/
- highest level of an app, contains common models, configs, and constants, etc.src/main.py
- root of the project, which inits the FastAPI app
- Each package has its own router, schemas, models, etc.
router.py
- is a core of each module with all the endpointsschemas.py
- for pydantic modelsmodels.py
- for db modelsservice.py
- module specific business logicdependencies.py
- router dependenciesconstants.py
- module specific constants and error codesconfig.py
- e.g. env varsutils.py
- non-business logic functions, e.g. response normalization, data enrichment, etc.exceptions.py
- module specific exceptions, e.g.PostNotFound
,InvalidUserData
- When package requires services or dependencies or constants from other packages - import them with an explicit module name
from src.auth import constants as auth_constants
from src.notifications import service as notification_service
from src.posts.constants import ErrorCode as PostsErrorCode # in case we have Standard ErrorCode in constants module of each package
FastAPI is an async framework, in the first place. It is designed to work with async I/O operations and that is the reason it is so fast.
However, FastAPI doesn't restrict you to use only async
routes, and the developer can use sync
routes as well. This might confuse beginner developers into believing that they are the same, but they are not.
Under the hood, FastAPI can effectively handle both async and sync I/O operations.
- FastAPI runs
sync
routes in the threadpool and blocking I/O operations won't stop the event loop from executing the tasks. - If the route is defined
async
then it's called regularly viaawait
and FastAPI trusts you to do only non-blocking I/O operations.
The caveat is if you fail that trust and execute blocking operations within async routes, the event loop will not be able to run the next tasks until that blocking operation is done.
import asyncio
import time
from fastapi import APIRouter
router = APIRouter()
@router.get("/terrible-ping")
async def terrible_ping():
time.sleep(10) # I/O blocking operation for 10 seconds, the whole process will be blocked
return {"pong": True}
@router.get("/good-ping")
def good_ping():
time.sleep(10) # I/O blocking operation for 10 seconds, but in a separate thread for the whole `good_ping` route
return {"pong": True}
@router.get("/perfect-ping")
async def perfect_ping():
await asyncio.sleep(10) # non-blocking I/O operation
return {"pong": True}
What happens when we call:
GET /terrible-ping
- FastAPI server receives a request and starts handling it
- Server's event loop and all the tasks in the queue will be waiting until
time.sleep()
is finished- Server thinks
time.sleep()
is not an I/O task, so it waits until it is finished - Server won't accept any new requests while waiting
- Server thinks
- Server returns the response.
- After a response, server starts accepting new requests
GET /good-ping
- FastAPI server receives a request and starts handling it
- FastAPI sends the whole route
good_ping
to the threadpool, where a worker thread will run the function - While
good_ping
is being executed, event loop selects next tasks from the queue and works on them (e.g. accept new request, call db)- Independently of main thread (i.e. our FastAPI app),
worker thread will be waiting for
time.sleep
to finish. - Sync operation blocks only the side thread, not the main one.
- Independently of main thread (i.e. our FastAPI app),
worker thread will be waiting for
- When
good_ping
finishes its work, server returns a response to the client
GET /perfect-ping
- FastAPI server receives a request and starts handling it
- FastAPI awaits
asyncio.sleep(10)
- Event loop selects next tasks from the queue and works on them (e.g. accept new request, call db)
- When
asyncio.sleep(10)
is done, servers finishes the execution of the route and returns a response to the client
Warning
Notes on the thread pool:
- Threads require more resources than coroutines, so they are not as cheap as async I/O operations.
- Thread pool has a limited number of threads, i.e. you might run out of threads and your app will become slow. Read more (external link)
The second caveat is that operations that are non-blocking awaitables or are sent to the thread pool must be I/O intensive tasks (e.g. open file, db call, external API call).
- Awaiting CPU-intensive tasks (e.g. heavy calculations, data processing, video transcoding) is worthless since the CPU has to work to finish the tasks, while I/O operations are external and server does nothing while waiting for that operations to finish, thus it can go to the next tasks.
- Running CPU-intensive tasks in other threads also isn't effective, because of GIL. In short, GIL allows only one thread to work at a time, which makes it useless for CPU tasks.
- If you want to optimize CPU intensive tasks you should send them to workers in another process.
Related StackOverflow questions of confused users
- https://stackoverflow.com/questions/62976648/architecture-flask-vs-fastapi/70309597#70309597
- Here you can also check my answer
- https://stackoverflow.com/questions/65342833/fastapi-uploadfile-is-slow-compared-to-flask
- https://stackoverflow.com/questions/71516140/fastapi-runs-api-calls-in-serial-instead-of-parallel-fashion
Pydantic has a rich set of features to validate and transform data.
In addition to regular features like required & non-required fields with default values, Pydantic has built-in comprehensive data processing tools like regex, enums, strings manipulation, emails validation, etc.
from enum import Enum
from pydantic import AnyUrl, BaseModel, EmailStr, Field
class MusicBand(str, Enum):
AEROSMITH = "AEROSMITH"
QUEEN = "QUEEN"
ACDC = "AC/DC"
class UserBase(BaseModel):
first_name: str = Field(min_length=1, max_length=128)
username: str = Field(min_length=1, max_length=128, pattern="^[A-Za-z0-9-_]+$")
email: EmailStr
age: int = Field(ge=18, default=None) # must be greater or equal to 18
favorite_band: MusicBand | None = None # only "AEROSMITH", "QUEEN", "AC/DC" values are allowed to be inputted
website: AnyUrl | None = None
Having a controllable global base model allows us to customize all the models within the app. For instance, we can enforce a standard datetime format or introduce a common method for all subclasses of the base model.
from datetime import datetime
from zoneinfo import ZoneInfo
from fastapi.encoders import jsonable_encoder
from pydantic import BaseModel, ConfigDict
def datetime_to_gmt_str(dt: datetime) -> str:
if not dt.tzinfo:
dt = dt.replace(tzinfo=ZoneInfo("UTC"))
return dt.strftime("%Y-%m-%dT%H:%M:%S%z")
class CustomModel(BaseModel):
model_config = ConfigDict(
json_encoders={datetime: datetime_to_gmt_str},
populate_by_name=True,
)
def serializable_dict(self, **kwargs):
"""Return a dict which contains only serializable fields."""
default_dict = self.model_dump()
return jsonable_encoder(default_dict)
In the example above, we have decided to create a global base model that:
- Serializes all datetime fields to a standard format with an explicit timezone
- Provides a method to return a dict with only serializable fields
BaseSettings was a great innovation for reading environment variables, but having a single BaseSettings for the whole app can become messy over time. To improve maintainability and organization, we have split the BaseSettings across different modules and domains.
# src.auth.config
from datetime import timedelta
from pydantic_settings import BaseSettings
class AuthConfig(BaseSettings):
JWT_ALG: str
JWT_SECRET: str
JWT_EXP: int = 5 # minutes
REFRESH_TOKEN_KEY: str
REFRESH_TOKEN_EXP: timedelta = timedelta(days=30)
SECURE_COOKIES: bool = True
auth_settings = AuthConfig()
# src.config
from pydantic import PostgresDsn, RedisDsn, model_validator
from pydantic_settings import BaseSettings
from src.constants import Environment
class Config(BaseSettings):
DATABASE_URL: PostgresDsn
REDIS_URL: RedisDsn
SITE_DOMAIN: str = "myapp.com"
ENVIRONMENT: Environment = Environment.PRODUCTION
SENTRY_DSN: str | None = None
CORS_ORIGINS: list[str]
CORS_ORIGINS_REGEX: str | None = None
CORS_HEADERS: list[str]
APP_VERSION: str = "1.0"
settings = Config()
Pydantic is a great schema validator, but for complex validations that involve calling a database or external services, it is not sufficient.
FastAPI documentation mostly presents dependencies as DI for endpoints, but they are also excellent for request validation.
Dependencies can be used to validate data against database constraints (e.g., checking if an email already exists, ensuring a user is found, etc.).
# dependencies.py
async def valid_post_id(post_id: UUID4) -> dict[str, Any]:
post = await service.get_by_id(post_id)
if not post:
raise PostNotFound()
return post
# router.py
@router.get("/posts/{post_id}", response_model=PostResponse)
async def get_post_by_id(post: dict[str, Any] = Depends(valid_post_id)):
return post
@router.put("/posts/{post_id}", response_model=PostResponse)
async def update_post(
update_data: PostUpdate,
post: dict[str, Any] = Depends(valid_post_id),
):
updated_post = await service.update(id=post["id"], data=update_data)
return updated_post
@router.get("/posts/{post_id}/reviews", response_model=list[ReviewsResponse])
async def get_post_reviews(post: dict[str, Any] = Depends(valid_post_id)):
post_reviews = await reviews_service.get_by_post_id(post["id"])
return post_reviews
If we didn't put data validation to dependency, we would have to validate post_id
exists
for every endpoint and write the same tests for each of them.
Dependencies can use other dependencies and avoid code repetition for the similar logic.
# dependencies.py
from fastapi.security import OAuth2PasswordBearer
from jose import JWTError, jwt
async def valid_post_id(post_id: UUID4) -> dict[str, Any]:
post = await service.get_by_id(post_id)
if not post:
raise PostNotFound()
return post
async def parse_jwt_data(
token: str = Depends(OAuth2PasswordBearer(tokenUrl="/auth/token"))
) -> dict[str, Any]:
try:
payload = jwt.decode(token, "JWT_SECRET", algorithms=["HS256"])
except JWTError:
raise InvalidCredentials()
return {"user_id": payload["id"]}
async def valid_owned_post(
post: dict[str, Any] = Depends(valid_post_id),
token_data: ddict[str, Any] = Depends(parse_jwt_data),
) -> dict[str, Any]:
if post["creator_id"] != token_data["user_id"]:
raise UserNotOwner()
return post
# router.py
@router.get("/users/{user_id}/posts/{post_id}", response_model=PostResponse)
async def get_user_post(post: dict[str, Any] = Depends(valid_owned_post)):
return post
Dependencies can be reused multiple times, and they won't be recalculated - FastAPI caches dependency's result within a request's scope by default,
i.e. if valid_post_id
gets called multiple times in one route, it will be called only once.
Knowing this, we can decouple dependencies onto multiple smaller functions that operate on a smaller domain and are easier to reuse in other routes.
For example, in the code below we are using parse_jwt_data
three times:
valid_owned_post
valid_active_creator
get_user_post
,
but parse_jwt_data
is called only once, in the very first call.
# dependencies.py
from fastapi import BackgroundTasks
from fastapi.security import OAuth2PasswordBearer
from jose import JWTError, jwt
async def valid_post_id(post_id: UUID4) -> Mapping:
post = await service.get_by_id(post_id)
if not post:
raise PostNotFound()
return post
async def parse_jwt_data(
token: str = Depends(OAuth2PasswordBearer(tokenUrl="/auth/token"))
) -> dict:
try:
payload = jwt.decode(token, "JWT_SECRET", algorithms=["HS256"])
except JWTError:
raise InvalidCredentials()
return {"user_id": payload["id"]}
async def valid_owned_post(
post: Mapping = Depends(valid_post_id),
token_data: dict = Depends(parse_jwt_data),
) -> Mapping:
if post["creator_id"] != token_data["user_id"]:
raise UserNotOwner()
return post
async def valid_active_creator(
token_data: dict = Depends(parse_jwt_data),
):
user = await users_service.get_by_id(token_data["user_id"])
if not user["is_active"]:
raise UserIsBanned()
if not user["is_creator"]:
raise UserNotCreator()
return user
# router.py
@router.get("/users/{user_id}/posts/{post_id}", response_model=PostResponse)
async def get_user_post(
worker: BackgroundTasks,
post: Mapping = Depends(valid_owned_post),
user: Mapping = Depends(valid_active_creator),
):
"""Get post that belong the active user."""
worker.add_task(notifications_service.send_email, user["id"])
return post
FastAPI supports both sync
and async
dependencies, and there is a temptation to use sync
dependencies, when you don't have to await anything, but that might not be the best choice.
Just as with routes, sync
dependencies are run in the thread pool. And threads here also come with a price and limitations, that are redundant, if you just make a small non-I/O operation.
See more (external link)
Developing RESTful API makes it easier to reuse dependencies in routes like these:
GET /courses/:course_id
GET /courses/:course_id/chapters/:chapter_id/lessons
GET /chapters/:chapter_id
The only caveat is to use the same variable names in the path:
- If you have two endpoints
GET /profiles/:profile_id
andGET /creators/:creator_id
that both validate whether the givenprofile_id
exists, butGET /creators/:creator_id
also checks if the profile is creator, then it's better to renamecreator_id
path variable toprofile_id
and chain those two dependencies.
# src.profiles.dependencies
async def valid_profile_id(profile_id: UUID4) -> Mapping:
profile = await service.get_by_id(profile_id)
if not profile:
raise ProfileNotFound()
return profile
# src.creators.dependencies
async def valid_creator_id(profile: Mapping = Depends(valid_profile_id)) -> Mapping:
if not profile["is_creator"]:
raise ProfileNotCreator()
return profile
# src.profiles.router.py
@router.get("/profiles/{profile_id}", response_model=ProfileResponse)
async def get_user_profile_by_id(profile: Mapping = Depends(valid_profile_id)):
"""Get profile by id."""
return profile
# src.creators.router.py
@router.get("/creators/{profile_id}", response_model=ProfileResponse)
async def get_user_profile_by_id(
creator_profile: Mapping = Depends(valid_creator_id)
):
"""Get creator's profile by id."""
return creator_profile
If you think you can return Pydantic object that matches your route's response_model
to make some optimizations,
then it's wrong.
FastAPI firstly converts that pydantic object to dict with its jsonable_encoder
, then validates
data with your response_model
, and only then serializes your object to JSON.
from fastapi import FastAPI
from pydantic import BaseModel, root_validator
app = FastAPI()
class ProfileResponse(BaseModel):
@model_validator(mode="after")
def debug_usage(self):
print("created pydantic model")
return self
@app.get("/", response_model=ProfileResponse)
async def root():
return ProfileResponse()
Logs Output:
[INFO] [2022-08-28 12:00:00.000000] created pydantic model
[INFO] [2022-08-28 12:00:00.000020] created pydantic model
If you must use a library to interact with external services, and it's not async
,
then make the HTTP calls in an external worker thread.
We can use the well-known run_in_threadpool
from starlette.
from fastapi import FastAPI
from fastapi.concurrency import run_in_threadpool
from my_sync_library import SyncAPIClient
app = FastAPI()
@app.get("/")
async def call_my_sync_library():
my_data = await service.get_my_data()
client = SyncAPIClient()
await run_in_threadpool(client.make_request, data=my_data)
If you raise a ValueError
in a Pydantic schema that is directly faced by the client, it will return a nice detailed response to users.
# src.profiles.schemas
from pydantic import BaseModel, field_validator
class ProfileCreate(BaseModel):
username: str
@field_validator("password", mode="after")
@classmethod
def valid_password(cls, password: str) -> str:
if not re.match(STRONG_PASSWORD_PATTERN, password):
raise ValueError(
"Password must contain at least "
"one lower character, "
"one upper character, "
"digit or "
"special symbol"
)
return password
# src.profiles.routes
from fastapi import APIRouter
router = APIRouter()
@router.post("/profiles")
async def get_creator_posts(profile_data: ProfileCreate):
pass
Response Example:
- Unless your API is public, hide docs by default. Show it explicitly on the selected envs only.
from fastapi import FastAPI
from starlette.config import Config
config = Config(".env") # parse .env file for env variables
ENVIRONMENT = config("ENVIRONMENT") # get current env name
SHOW_DOCS_ENVIRONMENT = ("local", "staging") # explicit list of allowed envs
app_configs = {"title": "My Cool API"}
if ENVIRONMENT not in SHOW_DOCS_ENVIRONMENT:
app_configs["openapi_url"] = None # set url for docs as null
app = FastAPI(**app_configs)
- Help FastAPI to generate an easy-to-understand docs
- Set
response_model
,status_code
,description
, etc. - If models and statuses vary, use
responses
route attribute to add docs for different responses
- Set
from fastapi import APIRouter, status
router = APIRouter()
@router.post(
"/endpoints",
response_model=DefaultResponseModel, # default response pydantic model
status_code=status.HTTP_201_CREATED, # default status code
description="Description of the well documented endpoint",
tags=["Endpoint Category"],
summary="Summary of the Endpoint",
responses={
status.HTTP_200_OK: {
"model": OkResponse, # custom pydantic model for 200 response
"description": "Ok Response",
},
status.HTTP_201_CREATED: {
"model": CreatedResponse, # custom pydantic model for 201 response
"description": "Creates something from user request ",
},
status.HTTP_202_ACCEPTED: {
"model": AcceptedResponse, # custom pydantic model for 202 response
"description": "Accepts request and handles it later",
},
},
)
async def documented_route():
pass
Explicitly setting the indexes' namings according to your database's convention is preferable over sqlalchemy's.
from sqlalchemy import MetaData
POSTGRES_INDEXES_NAMING_CONVENTION = {
"ix": "%(column_0_label)s_idx",
"uq": "%(table_name)s_%(column_0_name)s_key",
"ck": "%(table_name)s_%(constraint_name)s_check",
"fk": "%(table_name)s_%(column_0_name)s_fkey",
"pk": "%(table_name)s_pkey",
}
metadata = MetaData(naming_convention=POSTGRES_INDEXES_NAMING_CONVENTION)
- Migrations must be static and revertable. If your migrations depend on dynamically generated data, then make sure the only thing that is dynamic is the data itself, not its structure.
- Generate migrations with descriptive names & slugs. Slug is required and should explain the changes.
- Set human-readable file template for new migrations. We use
*date*_*slug*.py
pattern, e.g.2022-08-24_post_content_idx.py
# alembic.ini
file_template = %%(year)d-%%(month).2d-%%(day).2d_%%(slug)s
Being consistent with names is important. Some rules we followed:
- lower_case_snake
- singular form (e.g.
post
,post_like
,user_playlist
) - group similar tables with module prefix, e.g.
payment_account
,payment_bill
,post
,post_like
- stay consistent across tables, but concrete namings are ok, e.g.
- use
profile_id
in all tables, but if some of them need only profiles that are creators, usecreator_id
- use
post_id
for all abstract tables likepost_like
,post_view
, but use concrete naming in relevant modules likecourse_id
inchapters.course_id
- use
_at
suffix for datetime_date
suffix for date
- Usually, database handles data processing much faster and cleaner than CPython will ever do.
- It's preferable to do all the complex joins and simple data manipulations with SQL.
- It's preferable to aggregate JSONs in DB for responses with nested objects.
# src.posts.service
from typing import Any
from pydantic import UUID4
from sqlalchemy import desc, func, select, text
from sqlalchemy.sql.functions import coalesce
from src.database import database, posts, profiles, post_review, products
async def get_posts(
creator_id: UUID4, *, limit: int = 10, offset: int = 0
) -> list[dict[str, Any]]:
select_query = (
select(
(
posts.c.id,
posts.c.slug,
posts.c.title,
func.json_build_object(
text("'id', profiles.id"),
text("'first_name', profiles.first_name"),
text("'last_name', profiles.last_name"),
text("'username', profiles.username"),
).label("creator"),
)
)
.select_from(posts.join(profiles, posts.c.owner_id == profiles.c.id))
.where(posts.c.owner_id == creator_id)
.limit(limit)
.offset(offset)
.group_by(
posts.c.id,
posts.c.type,
posts.c.slug,
posts.c.title,
profiles.c.id,
profiles.c.first_name,
profiles.c.last_name,
profiles.c.username,
profiles.c.avatar,
)
.order_by(
desc(coalesce(posts.c.updated_at, posts.c.published_at, posts.c.created_at))
)
)
return await database.fetch_all(select_query)
# src.posts.schemas
from typing import Any
from pydantic import BaseModel, UUID4
class Creator(BaseModel):
id: UUID4
first_name: str
last_name: str
username: str
class Post(BaseModel):
id: UUID4
slug: str
title: str
creator: Creator
# src.posts.router
from fastapi import APIRouter, Depends
router = APIRouter()
@router.get("/creators/{creator_id}/posts", response_model=list[Post])
async def get_creator_posts(creator: dict[str, Any] = Depends(valid_creator_id)):
posts = await service.get_posts(creator["id"])
return posts
Writing integration tests with DB will most likely lead to messed up event loop errors in the future. Set the async test client immediately, e.g. httpx
import pytest
from async_asgi_testclient import TestClient
from src.main import app # inited FastAPI app
@pytest.fixture
async def client() -> AsyncGenerator[TestClient, None]:
host, port = "127.0.0.1", "9000"
async with AsyncClient(transport=ASGITransport(app=app, client=(host, port)), base_url="http://test") as client:
yield client
@pytest.mark.asyncio
async def test_create_post(client: TestClient):
resp = await client.post("/posts")
assert resp.status_code == 201
Unless you have sync db connections (excuse me?) or aren't planning to write integration tests.
With linters, you can forget about formatting the code and focus on writing the business logic.
Ruff is "blazingly-fast" new linter that replaces black, autoflake, isort, and supports more than 600 lint rules.
It's a popular good practice to use pre-commit hooks, but just using the script was ok for us.
#!/bin/sh -e
set -x
ruff --fix src
ruff format src
Some very kind people shared their own experience and best practices that are definitely worth reading. Check them out at issues section of the project.
For instance, lowercase00 has described in details their best practices working with permissions & auth, class-based services & views, task queues, custom response serializers, configuration with dynaconf, etc.
If you have something to share about your experience working with FastAPI, whether it's good or bad, you are very welcome to create a new issue. It is our pleasure to read it.