20 Jan 2020

Combining differing error types

Rust, Scala, and many other languages let you use a kind of or to represent errors. In Scala it might be Either<E, T>, and in Rust it’s likely to be Result<T, E>. The E represents an error, and the awkward part of this is chaining together results with different types for E. This post contains my notes on this, for Rust.


  • A short introduction to the Result type
  • Combining errors of the same type
  • Introducing ? for short-cutting
  • A mapping between errors of different type
  • Application-specific errors
  • The From type class for conversion between error types
  • The standard error trait

Links and further reading are all at the end, except for links to the Rust playground to try out the code I’m showing.

Result and its combinators

Converting strings into numbers is something that can go wrong. For example, trying to parse the string “bananas” as an unsigned 16-bit integer in Rust would fail:

fn main() {

  let number = "123".parse::<u16>();
  println!("{:?}", number);

  let error = "bananas".parse::<u16>();
  println!("{:?}", error);

Running that code (open in playground) outputs:

Err(ParseIntError { kind: InvalidDigit })

Rust encourages you to encode these recoverable errors by using Result. Both number and error have the same type: a Result<u16, ParseIntError>. I read this as “the result is either a u16 or a parse error”.

As we can see, a Result can have a value that is either Ok or an Err:

pub enum Result<T, E> {

You can pattern match on results:

let msg = match number {
    Ok(_n) => "thank you for entering a number",
    Err(_) => "please enter an number",
println!("Regarding {:?}, {}", number, msg);
// Regarding Ok(123), thank you for entering a number

You can also use a range of combinators to work with results. The common ones are:

mapApplies a function if the result is an Ok.
and_thenSequence this result with another result (flatMap or bind in other languages).
map_errApply a function if the result is an Err (leftMap in Scala’s Cats library).
orCombine two results, giving you the first Ok (if any).
or_elseTurn an Err into an Ok with a value you specify.
unwrap_orThe content of an Ok, or the default value you specify.

For example, here’s map:

let plus_1 = "99".parse::<u16>().map(|n| n + 1);
println!("{:?}", plus_1);
// Ok(100)

Combining errors of the same type

The and_then combinator can combine two results (playground):

use std::num::ParseIntError;

fn add(input1: &str, input2: &str) -> Result<u16, ParseIntError> {
        .and_then(|a| input2.parse::<u16>().map(|b| a + b))

fn main() {
    println!("Success result:  {:?}", add("1", "2"));
    println!("Example failure: {:?}", add("1", "boom"));

The output is:

Success result:  Ok(3)
Example failure: Err(ParseIntError { kind: InvalidDigit })

I find that straightforward, if a little tricky to read. Rust also has a fast-fail short cut for this, via the ? operator (playground):

fn add(input1: &str, input2: &str) -> Result<u16, ParseIntError> {
    let a: u16 = input1.parse()?;
    let b: u16 = input2.parse()?;
    Ok(a + b)

The ? at the end of the assignment to a and b is extracting the value from the Result. It can only do this for Ok values. Err values are immediately returned from add. ? is only applicable in the context of a Result return type (or Option, but that’s not our focus here).

The ? operator is almost equivalent to the and_then version or pattern matching, with a twist we’ll see later.

Notice that in the code above, all the types line up because we’re always working with the same error type: ParseIntError. That’s not always the case.

Combining errors of differing types

When there’s no relationship between different types of errors, we need to work harder to combine results. Suppose we want to parse a URL, then append a page number to it. The url crate provides a parser:

pub fn parse(input: &str) -> Result<Url, ParseError>

Let’s combine this with our number parsing. I’ve been explicit about the types of parsed_url and parsed_int to make the problem clearer hopefully (playground):


fn page_link(input_url: &str, input_page: &str) -> Result<String, ParseError> {
    let parsed_url: Result<Url, ParseError> = Url::parse(input_url);
    let parsed_int: Result<u16, ParseIntError> = input_page.parse::<u16>();

    parsed_url.and_then(|url| parsed_int.map(|page| format!("{}?page={}", url, page)))

The compiler tells us the problem is in the map: it expected ParseError but found struct ParseIntError. The problem is with me. I’ve said the type of page_link is a string or a ParseError, but what could I put in there to cater for both a URL parse error and a number parse error?

One solution is to pick one of the errors as a reasonable choice for page_link (playground):

use url;
use url::{ParseError, Url};

fn page_link(input_url: &str, input_page: &str) -> Result<String, ParseError> {
    let parsed_url: Result<Url, ParseError> = Url::parse(input_url);
    let parsed_int: Result<u16, ParseError> = 
        .map_err(|_err| ParseError::InvalidDomainCharacter);

    parsed_url.and_then(|url| parsed_int.map(|page| format!("{}?page={}", url, page)))

I’ve picked the url crate’s error for my function, and mapped number parsing into one of the error variations. I guess InvalidDomainCharacter is kind of sort of like a parsing error, but it’s a kludge.

It’s useful to be able to do this, but we can do better.

Application errors

I’m more likely to introduce my own application (or module) specific errors. Perhaps I’d define a Mishap enumeration with a couple of values:

enum Mishap {

Of course, I could have had just one catch-all error for parsing, or I could have included some context in my errors, but for now, I’m just marking two different kinds of errors.

With that, I can then have my function return a Mishap, and map_err lower-level errors into my error types (playground):

use Mishap::{BadPageNum, BadUrl};

fn page_link(input_url: &str, input_page: &str) -> Result<String, Mishap> {
    let parsed_url: Result<Url, Mishap> = 
        Url::parse(input_url).map_err(|_err| BadUrl);
    let parsed_int: Result<u16, Mishap> = 
        input_page.parse().map_err(|_err| BadPageNum);

    parsed_url.and_then(|url| parsed_int.map(|page| format!("{}?page={}", url, page)))

There’s a consistent pattern here: always map_err into Mishap, then sequence a computation at the end of the function.

This works as you might expect:

fn main() {
    println!("Success: {:?}", page_link("https://example.org", "5"));
    println!("Failure: {:?}", page_link("https://example.org", "five"));


Success: Ok("https://example.org/?page=5")
Failure: Err(BadPageNum)

Now that the error types all lineup, I can also use the ? short-cut too (playground):

fn page_link(input_url: &str, input_page: &str) -> Result<String, Mishap> {
    let url: Url = 
        Url::parse(input_url).map_err(|_err| BadUrl)?;
    let page: u16  = 
        input_page.parse().map_err(|_err| BadPageNum)?;

    Ok(format!("{}?page={}", url, page))

I think that’s a resaonably good way to get a readable function.


Inserting map_err for one-off cases is fine, but it is tedious if you’re doing this multiple times. Hopefully, you can isolate code to return one error type, and map_err it once. But if it is common, there’s a type class called From that eases the way.

A simplified version of From is:

pub trait From<T> {
    fn from(t: T) -> Self;

From captures a conversion from one type, T, to the implementing trait (Self). It’s not specific to error handling but helps us because ? calls from on errors, inferring the type we’re going to from the function signature.

The upshot is that we can implement a conversion from specific errors into our application error (playground):

impl From<ParseIntError> for Mishap {
    fn from(_err: ParseIntError) -> Mishap {

impl From<ParseError> for Mishap {
    fn from(_err: ParseError) -> Mishap {

And then omit the map_err calls:

fn page_link(input_url: &str, input_page: &str) -> Result<String, Mishap> {
    let url: Url = Url::parse(input_url)?;
    let page: u16 = input_page.parse()?;
    Ok(format!("{}?page={}", url, page))

To be clear, what’s happening here is the ? operators are causing fast-fail out of page_link. We can use ? here because the enclosing type is a Result. It compiles because ? invokes the From::from call to the inferred type (Mishap).

Without the From instances in place for url::ParseError and ParseIntError, the above function would not compile.

In practice

In an application, I end up doing a little of all these options: a few Froms when it makes sense and some map_err at other times. There are existing crates to help with this, such as automatically deriving From implementations for you. I’ve only started using these recently.

One I like the is a crate called thiserror, which I’ve used to wrap lower-level errors, and provide From implementations. It looks like this in use:

use thiserror::Error;

#[derive(Error, Debug)]
pub enum Mishap {
    Network(#[from] native_tls::Error),

    Imap(#[from] imap::error::Error),

    File(#[from] std::io::Error),

    #[error("Upload failed: {0}")]

The standard error trait

thiserror and other libraries also implement a standard Rust trait called std::error::Error. The intention is to capture some minimal error functionality. Functions like source() to give you lower-level errors, and to_string() for a message (via the Display trait, which is a kind of Show from other languages).

What that means is you don’t need to care about the specific error. It is not something I lean towards, but when glueing together code at a main-level, I see the value of not caring at that point what the error is, other than the fact I have an error.

Here’s how std::error:Error looks in the example I’ve been using (playground):

type Result<T> = std::result::Result<T, Box<dyn std::error::Error>>;

fn page_link(input_url: &str, input_page: &str) -> Result<String> {
    let url: Url = Url::parse(input_url)?;
    let page: u16 = input_page.parse::<u16>()?;
    Ok(format!("{}?page={}", url, page))

As the return type is getting long, I’ve introduced a type alias called Result<T>. The error type is the exciting part. Working from the inside:

  • We’re dealing with an error that implements the trait std::error::Error.
  • There’s some error at runtime that implements the trait, but we do not know at compile time what it is exactly. It is “dynamic”, in the sense that there needs to be extra information carried around to know how to dynamically dispatch method calls.
  • We also do not know the size of the error type. So we put it inside a Box, which is a pointer out to the heap.

I’ve found the dyn part easier to understand in comparison to the alternative. Consider:

fn known_at_compile_time<E>(err: E) -> String
    E: std::error::Error,

I’m more comfortable with code like this. We’ve written a generic function that takes any type E as an argument, with the constraint that there must be an implementation of std::error::Error for E. It means at compile-time, we know the specific type for E. That’s in contrast to the dyn Error which is stating something at runtime will implement the Error trait.

The result of using std::error::Error is pretty much the same:

fn main() {
        "Success:  {:?}",
        page_link("https://example.org", "5")
    // Success:  Ok("https://example.org/?page=5")

…but I can’t exhaustively match on my specific application errors. I usually want to be able to do that, but not all the time.

Notice that we do not need to implement From because the url crate does this, as does the internal ParseIntError. It’s a common practice for libraries to implement the std::error::Error trait.


Most of the time I’ll encode errors into a module (or application) specific enum. Then I’ll either map_err or use From to adapt errors as I chain computations together. Crates like thiserror reduce some of the boilerplate for this, and give a framework for describing errors.