We've all been there before. We write a function that has to deal with some edge case, and we use the throw
keyword in order to handle this situation:
type ResponseData = {
statusCode: number
responseBody?: ResponseBody
}
const makeHttpRequest = async (url: string): Promise<ResponseData> => {
if (!isUrl(url)) {
throw new Error(
'Invalid string passed into `makeHttpRequest`. Expected a valid URL.'
)
}
// ...
// other business logic here
// ...
return { ... } // ResponseData
}
Now imagine a month later, you are working on your project when you or a colleague forget to wrap makeHttpRequest
inside of a try / catch
block.
Two things happen here:
-
The compiler is no longer able to tell you whether your code is safe from runtime errors. In other words, using
throw
is not typesafe. And is just as dangerous asany
. They both dilute the benefits of using TypeScript in the first place. -
Because neither the compiler nor the types tell you that
makeHttpRequest
can fail (read: throw), you will eventually get a runtime error. This is a waste of time, money and happiness for everyone. People begin to ask why they're using TypeScript if the compiler isn't helping them with something so basic as adding atry / catch
block.
So the question is:
How do we encode failability into the typesystem?
First, let's begin by acknowledging that throw
is not typesafe. We must use a different approach in order to get the TypeScript compiler on our side.
What if we had a type
or interface
which represented the outcome of a computation that might fail?
Our type would represent two simple outcomes:
- Success case: Which would return / contain contain a "success value" (i.e.
ResponseData
in the case ofmakeHttpRequest
) - Failure case: Which would return / contain helpful information about why the failure occurred
Let's call our type somethething intuitive like, Result
. Let's call our Success variant Ok
and our Failure variant Err
.
Thus, if we were to formalize our type into code, it would look something like this:
type Result<T, E>
= Ok<T, E> // contains a success value of type T
| Err<T, E> // contains a failure value of type E
Going back to our makeHttpRequest
function, we would want to encode the potential for failure into the typesystem.
Thus makeHttpRequest
would have the following signature:
makeHttpRequest(url: string): Promise<Result<ResponseData, Error>>
And the function definition would look something like this:
// utility functions to build Ok and Err instances
const ok = <T, E>(value: T): Result<T, E> => new Ok(value)
const err = <T, E>(error: E): Result<T, E> => new Err(error)
const makeHttpRequest = async (url: string): Promise<Result<ResponseData, Error>> => {
if (!isUrl(url)) {
return err(new Error(
'Invalid string passed into `makeHttpRequest`. Expected a valid URL.'
))
}
// ...
// other business logic here
// ...
return ok({ ... }) // Ok(ResponseData)
}
Of course err(new Error('...'))
seems a little tedious. But here are some things you should know:
-
The argument of the
err
function must be of typeE
, or you'll get a compile error (type mismatch) between the type inside oferr
and the return type ofmakeHttpRequest
(where theE
type is represented as anError
instance).- Relatedly, I just chose
Error
as the type forE
for the sake of simplicity ... meaningE
could be anything you want! More on this in a bit!
- Relatedly, I just chose
-
The user of
makeHttpRequest
can use this function without fear that it might randomly throw. No more runtime errors 🚀 -
The author of the
makeHttpRequest
function also doesn't have to worry about writing and updating documentation every time a new edge case appears that would cause the function to throw an error. All of the behaviour of the function is encoded in the return type. Relatedly, the type serves as documentation now: "makeHttpRequest
is an asynchronous function that can either succeed withResponseData
or fail with aError
."
... "But wait, how do I get the T
value or E
value that is wrapped inside of a Result<T, E>
?"
Great question. Let me show you how. We're going to use a package I made [aptly] named neverthrow
.
> npm install neverthrow
import { ok, err, Result } from 'neverthrow'
// we'll keep this simple
type ResponseBody = {}
interface ResponseData {
statusCode: number
responseBody?: ResponseBody
}
const makeHttpRequest = async (
url: string
): Promise<Result<ResponseData, Error>> => {
if (!isUrl(url)) {
return err(new Error(
'Invalid string passed into `makeHttpRequest`. Expected a valid URL.'
))
}
// ...
// other business logic here
// ...
return ok({ ... }) // Ok(ResponseData)
}
So we're currently at the same place we were at with the last code snippet, except this time we're using the neverthrow
package.
If you were to read through the neverthrow
documentation you'd see that a Result
has a .map
method which takes the T
value inside of a Result
and converts it into anything you want.
Here's an example:
import { makeHttpRequest } from './http-api.ts'
const run = async () => {
// unwrap the Promise
// at this point
// we have a Result<ResponseData, Error>
const result = await makeHttpRequest('https://jsonplaceholder.typicode.com/todos/1')
result.map(responseData => {
console.log(responseData)
})
}
run()
But wait, what if the result variable contains an E
value? in other words, it's an Err
instead of an Ok
.
Well, again, the docs for neverthrow
show you how to handle this situation too ... Just use mapErr
!
import { makeHttpRequest } from './http-api.ts'
const run = async () => {
// unwrap the Promise
// at this point
// we have a Result<ResponseData, Error>
const result = await makeHttpRequest('https://jsonplaceholder.typicode.com/todos/1')
result.mapErr(errorInstance => {
console.log(errorInstance)
})
}
run()
The most beautiful thing about Result
s is that they are chainable! Here's the above code in a more realistic example:
import { makeHttpRequest } from './http-api.ts'
const run = async () => {
// unwrap the Promise
// at this point
// we have a Result<ResponseData, Error>
const result = await makeHttpRequest('https://jsonplaceholder.typicode.com/todos/1')
result
.map(responseData => {
// do something with the success value
})
.mapErr(errorInstance => {
// do something with the failure value
})
}
run()
There is a lot more you can do with a Result
type (check out the docs), but map
ing is the most important part of the API.
Making Your Types More Intuitive
If you start using Result
a lot in your return types, you might notice two things:
-
The meaning of the
Result
s is not very clear- Example: The
Result
of a databse query might be something likePromise<Result<T, DbError>>
while theResult
of a network call might be something likePromise<Result<T, NetworkError>>
.
- Example: The
-
The types are really long and verbose. For example, above we had a
Promise<Result<ResponseData, Error>>
... this is a somewhat intimidating type!
To solve both issues, you could leverage type aliases!
Here's an example. Instead of having a function return a generic Result
with a DbError
as the E
type, why not alias this type to something much more intuitive?
type DbResult<T> = Result<T, DbError>
Now your function can just return Promise<DbResult<T>>
. It's a lot more succinct!
Further, your type now encodes meaning. The above type says that there's something async going on that could fail and I know that it's dealing with the database. Neat!
Here's a real-world example from one of my projects:
handler: (req: Request, res: SessionManager) => DecodeResult<Promise<RouteResult<T>>>
So handler
is a function that does a few things:
- It does some decoding / deserializing of incoming data
- It then does something async which generates a
RouteResult
with some data of typeT
I know exactly what's going to happen by only reading the types. And the beautiful thing is that I won't ever get runtime errors because none of my code throws (and all the 3rd party libs I depend on have been wrapped to return Result
s as well).
Summary
-
Avoid using
throw
if you can.- Users of your API are not required to
catch
(the compiler doesn't enforce it). This means you will eventually hit a runtime error ... it's just a matter of time - Using
throw
forces you to maintain documentation which will eventually become stale - If you do manage to maintain your documentation, a lot of people won't bother reading completely through the documentation, and won't realize that your functions throw in certain scenarios
- Users of your API are not required to
-
Encode the potential for failure into your types using
Result
s- The types are self documenting, and they cannot become "stale"
- Users are given a friendly API that lets them deal with failable values in a safe way (using
map
andmapErr
)
-
There's a fully-tested and type-checked npm package for this called
neverthrow
... try it out!
Liked this post? Consider buying me a coffee :)