5 Things we've learned building large APIs with FastAPI

FastAPI follows the UNIX philosophy of "do one thing, and do it well". By using Starlette and Pydantic, FastAPI provides you with powerful tools to build a beautiful API. This gives you a lot of freedom to decide how you organise the rest of your codebase. The level of freedom also comes with some challenges once a codebase grows.

In this talk we explore some of the common challenges in building FastAPI apps, and share how we solved them:

1. Minimizing the global state with dependency injection
2. Reducing the complexity of testing a FastAPI app
3. Pitfalls of async FastAPI
4. Supporting multiple authentication schemes
5. Modelling the relations between patch, response and database models

1. 5 Things we've learned building large APIs with FastAPI @ InvestSuite Maarten Huijsmans
2. My last public presentation … in primary school
3. @lukin0110 Who am I? Maarten Huijsmans Python Developer @ InvestSuite twitter/lukin0110 github.com/lukin0110 linkedin.com/in/mhuijsmans
4. A suite of modern wealthtech solutions Built with Python and FastAPI Wealthtech as a service investsuite.com
5. @lukin0110 • Relatively new python web framework • Built on top of Starlette (ASGI) • 43.8k Github stars • Created by Sebastián Ramirez (@tiangolo) • Excellent documentation ❤ https://fastapi.tiangolo.com/
6. 5 Things we've learned building large APIs with FastAPI @lukin0110 I assume that you’re familiar with: • The basics of FastAPI • Pydantic • Web / API development • Python 😎 Assumptions
7. 5 Things we've learned building large APIs with FastAPI @lukin0110 1. 🥳 Minimizing the global state with dependency injection 2. 󰟜 Reducing the complexity of testing a FastAPI app 3. 😎 Pitfalls of async Python/FastAPI 4. 🔐 Supporting multiple authentication schemes 5. 💾 Modelling the relations between patch, response and database models Goal of the next 30 minutes
8. Minimizing the global state with dependency injection 🥳 5 Things we've learned building large APIs with FastAPI
9. 🥳 1. Minimizing the global state with dependency injection Global state, que?
10. 🥳 1. Minimizing the global state with dependency injection
11. 🥳 1. Minimizing the global state with dependency injection • Generally bad practice • You easily end up there because FastAPI itself lives in the global state • Hard to patch in unit tests Global state, que?
12. 🥳 1. Minimizing the global state with dependency injection Idea? Avoid global state
13. 🥳 1. Minimizing the global state with dependency injection Inject state with Depends
14. 🥳 1. Minimizing the global state with dependency injection Make injected state configurable & reusable
15. Reducing the complexity of testing a FastAPI app 󰟜 5 Things we've learned building large APIs with FastAPI
16. 󰟜 2. Reducing the complexity of testing a FastAPI app See the previous learning 󰤇 Reduce/avoid the global state. It will make you happy 😎 Lesson learned
17. 󰟜 2. Reducing the complexity of testing a FastAPI app • Override configs • Mock/patch functions • Setup a TestClient Things to deal with when testing an API …
18. 󰟜 2. Reducing the complexity of testing a FastAPI app
19. 󰟜 2. Reducing the complexity of testing a FastAPI app Use BaseSettings to centralize app config
20. 󰟜 2. Reducing the complexity of testing a FastAPI app Override dependencies
21. 󰟜 2. Reducing the complexity of testing a FastAPI app Pure python classes / functions • Forces you to not rely directly on environment variables • Dependencies are cached • Endpoints can be lightweight wrappers around business logic
22. Pitfalls of async Python/FastAPI 😎 5 Things we've learned building large APIs with FastAPI
23. 😎 3. Pitfalls of async Python/FastAPI • asyncio is not as easy as it looks like • GIL: Global Interpreter Lock • “Sync” libraries Async python is an event loop. A single-thread in a single-process. The pitfalls
24. 😎 3. Pitfalls of async Python/FastAPI Blocking code Solution!
25. 😎 3. Pitfalls of async Python/FastAPI Global Interpreter Lock and CPU-Blocking
26. 😎 3. Pitfalls of async Python/FastAPI Async python Async python
27. 😎 3. Pitfalls of async Python/FastAPI • A lot of python libraries are not asyncio compatible. Or not battle tested. • You easily end up with a blocking application. Not always clear what is blocking it in a large codebase. Lessons learned
28. 😎 3. Pitfalls of async Python/FastAPI • Remove all `async` prefixes => application becomes multi-threaded • Offloading CPU-bound operations to: ◦ a ProcessPoolExecutor => moves GIL troubles to another process and FastAPI remains responsive ◦ AWS Lambda => multi-process by default ◦ Queuing systems (RabbitMQ, Redis, Kafka, etc) and process asynchronously on another server. What can you do …
29. 😎 3. Pitfalls of async Python/FastAPI https://fastapi.tiangolo.com/async/ Read the documentation carefully
30. Supporting multiple authentication schemes 🔐 5 Things we've learned building large APIs with FastAPI
31. 🔐 4. Supporting multiple authentication schemes Be able to support 2 different auth schemes at the same time. • JWT • API Keys Want did we want
32. 🔐 4. Supporting multiple authentication schemes • Provides a lot of utilities ◦ For API Keys, Basic Auth, Oauth2, … • Utilities integrate very well with OpenAPI specs and Swagger/Redoc • Extensive documentation • Various code samples FastAPI & Security
33. 🔐 4. Supporting multiple authentication schemes • Not with batteries included • Actual implementation/integration of flows you need to do yourself • FastAPI auth is different from Starlette auth Things to keep in mind
34. 🔐 4. Supporting multiple authentication schemes Middleware based (Starlette) Dependency based (FastAPI)
35. 🔐 4. Supporting multiple authentication schemes
36. 🔐 4. Supporting multiple authentication schemes A simple solution auto_error=False However… • Need to repeat this on every endpoint. • If you specify it on router level you don’t have access to a user anymore.
37. 🔐 4. Supporting multiple authentication schemes Maybe a middleware approach is better suited for authentication. • Decouple auth schemes from routers/views. • User is available on the request object. • Flexibility at router level to decide which schemes to use. Lessons learned
38. Modelling the relations between patch, response and database models 💾 5 Things we've learned building large APIs with FastAPI
39. 💾 5. Modelling the relations between patch, response and database models • Automatic type checking • Validation of input and output models out of the box • Generates beautiful documentation. OpenAPI, Swagger, & ReDoc FastAPI + Pydantic = Powerful 🚀
40. 💾 5. Modelling the relations between patch, response and database models http://localhost:8000/docs
41. 💾 5. Modelling the relations between patch, response and database models 1. Used inheritance to define patch, response and DB models 2. Fully documented our DB models 3. Exposed a subset of each DB model in the API How did we build CRUD APIs?
42. 💾 5. Modelling the relations between patch, response and database models Patch Model ⊆ Response Model ⊆ DB Model
43. 💾 5. Modelling the relations between patch, response and database models
44. 💾 5. Modelling the relations between patch, response and database models • DB models are hard to read in code • Hard to change an API contract without a DB migration 󰤇 • Hard to change a DB model without breaking the API contract However, some lessons learned here 🤯
45. 💾 5. Modelling the relations between patch, response and database models • Investigating decoupling DB models from API models 😎 • Investigating “Annotated” fields to reuse documentation Next steps
46. 💾 5. Modelling the relations between patch, response and database models Annotated Experimental 🤓
47. Thanks for your patience 🤓 Questions? twitter/lukin0110 github.com/lukin0110 linkedin.com/in/mhuijsmans maarten.huijsmans@investsuite.com