Railway Oriented Programming
Railway Oriented Programming
Slides and videos explaining a functional approach to error handling
Any suggestions on how this could be implemented in Python?
First Google result for the search Railway Oriented Programming in python:
Repository: DavidVujic/pythonic-railway: Experimenting with Railway oriented programming and Python
YouTube video (5min): David Vujic - A Pythonic Railway?
Here is an alternative Piped link(s): https://piped.video/watch?v=357sRh9OVEg
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I'm open-source, check me out at GitHub.
Why not just throw an error?
Catch it higher up the chain, and return that as the response.
Which would essentially the "early return" methodology. The final return of a function is the happy path, everything else is tested for and thrown as an error before hand (or returns a different value depending on the structure).I don't know why further parts in a chain (or "2 track railway") of methods would want to process an error.
Like, why not just have the error instantly return to the client.Maybe, for things like input validation, having a list of errors would be useful (IE email is invalid, passwords don't match, password must be 8 characters, and accept TOS).
Even then, that is a validation step. The main chain of processing shouldn't care if there is 1 or 5 errors. Validation doesn't pass, return the validation error.The only reason I can see is if you have some response handler that can accept an error or data, and you don't want to write a try/except, or you don't want to abstract the try/except to some parent method, or you find that try/excepts have some critical performance penalty that is wholey unacceptable.
I feel like this 9 year old solution is looking for a problem, or is an example of working within strange constraints, or is just an idea/example for a specific talk.
Maybe it's just being purely functional for the sakw of being functional?Idk. Maybe I just don't get functional programming?
Because the reality is that people don't document errors (what exceptions can be raised by a method), don't explicitly handle all the cases (how many times have you returned a 500 in a Flask app from missing an error type in your
exceptblock/missing an errorhandler), or don't even think to put a try...except around fallible code (I see this a lot, anecdotally). TheEithermonad (or Rust'sResult, which I'm more familiar with) force you to do something with the error explicitly.Exceptional cases are just as normal a part of a program as any other flow, including the success flows. If you have 20 happy paths and 80 exception paths, and only cover those 20 happy paths, you've covered 20% of the cases.
Typically the error track of the railroads analogy is implemented as a short circuit/return, so in effect this behavior IS an early return, but elegantly masked as though you are just handling the happy path.
In rust, if you function returns a Result type (the Either type from the slides), you can track a ? On to the end of any function call that returns a Result. This gets de-sugared to a check for an error and returns from the function early if it is.
At that point, the caller of the function gets to choose if they want to handle the potential failure of your function, or just let it bubble up further by using the ? operator. You end up with the behavior you’re describing, but with the benefit of errors handled by the type system and code that looks cleaner by just (explicitly) handling the happy path.
The immediate use for this that jumps out at me is batch processing: you take n inputs and return n outputs, where
output[i]is the result of processinginput[i]. You cannot throw since you still have to process all of the valid input.This style also works for an actor model: loosely coupled operations which take an input message and emit an output message for the next actor in the chain. If you want to be able to throw an exception or terminate prematurely, you would have to configure an error sink shared by all of the actors and to get the result of an operation, you so have to watch for messages on both the final actor and the error sink.