How do I stop iteration and return an error when Iterator::map returns a Result::Err?

Question

I have a function that returns a Result:

fn find(id: &Id) -> Result<Item, ItemError> {
    // ...
}

Then another using it like this:

let parent_items: Vec<Item> = parent_ids.iter()
    .map(|id| find(id).unwrap())
    .collect();

How do I handle the case of failure inside any of the map iterations?

I know I could use flat_map and in this case the error results would be ignored:

let parent_items: Vec<Item> = parent_ids.iter()
    .flat_map(|id| find(id).into_iter())
    .collect();

Result's iterator has either 0 or 1 items depending on the success state, and flat_map will filter it out if it's 0.

However, I don't want to ignore errors, I want to instead make the whole code block just stop and return a new error (based on the error that came up within the map, or just forward the existing error).

How do I best handle this in Rust?

Solution 1

Result implements FromIterator, so you can move the Result outside and iterators will take care of the rest (including stopping iteration if an error is found).

#[derive(Debug)]
struct Item;
type Id = String;

fn find(id: &Id) -> Result<Item, String> {
    Err(format!("Not found: {:?}", id))
}

fn main() {
    let s = |s: &str| s.to_string();
    let ids = vec![s("1"), s("2"), s("3")];

    let items: Result<Vec<_>, _> = ids.iter().map(find).collect();
    println!("Result: {:?}", items);
}

Playground

Solution 2

The accepted answer shows how to stop on error while collecting, and that's fine because that's what the OP requested. If you need processing that also works on large or infinite fallible iterators, read on.

As already noted, for can be used to emulate stop-on-error, but that is sometimes inelegant, as when you want to call max() or other consuming method. In other situations it's next to impossible, as when the consuming method is in another crate, such as itertools or Rayon¹.

Iterator consumer: try_for_each

When you control how the iterator is consumed, you can just use try_for_each to stop on first error. It accepts a closure that returns a Result, and try_for_each() will return Ok(()) if the closure returned Ok every time, and the first Err on the first error. This allows the closure to detect errors simply by using the ? operator in the natural way:

use std::{fs, io};

fn main() -> io::Result<()> {
    fs::read_dir("/")?.try_for_each(|e| -> io::Result<()> {
        println!("{}", e?.path().display());
        Ok(())
    })?;
    // ...
    Ok(())
}

If you need to maintain state between the invocations of the closure, you can also use try_fold. Both methods are implemented by ParallelIterator, so the same patter works with Rayon.

This approach requires that you control how the iterator is consumed. If that is done by code not under your control - for example, if you are passing the iterator to itertools::merge() or similar, you will need an adapter.

Iterator adapter: scan

The first attempt at stopping on error is to use take_while:

use std::{io, fs};

fn main() -> io::Result<()> {
    fs::read_dir("/")?
        .take_while(Result::is_ok)
        .map(Result::unwrap)
        .for_each(|e| println!("{}", e.path().display()));
    // ...
    Ok(())
}

This works, but we don't get any indication that an error occurred, the iteration just silently stops. Also it requires the unsightly map(Result::unwrap) which makes it seem like the program will panic on error, which is in fact not the case as we stop on error.

Both issues can be fixed by switching from take_while to scan, a more powerful combinator that not only supports stopping the iteration, but passes its callback owned items, allowing the closure to extract the error to the caller:

fn main() -> io::Result<()> {
    let mut err = Ok(());
    fs::read_dir("/")?
        .scan(&mut err, |err, res| match res {
            Ok(o) => Some(o),
            Err(e) => {
                **err = Err(e);
                None
            }
        })
        .for_each(|e| println!("{}", e.path().display()));
    err?;
    // ...
    Ok(())
}

If needed in multiple places, the closure can be abstracted into a utility function:

fn until_err<T, E>(err: &mut &mut Result<(), E>, item: Result<T, E>) -> Option<T> {
    match item {
        Ok(item) => Some(item),
        Err(e) => {
            **err = Err(e);
            None
        }
    }
}

...in which case we can invoke it as .scan(&mut err, until_err) (playground).

These examples trivially exhaust the iterator with for_each(), but one can chain it with arbitrary manipulations, including Rayon's par_bridge(). Using scan() it is even possible to collect() the items into a container and have access to the items seen before the error, which is sometimes useful and unavailable when collecting into Result<Container, Error>.

---

¹ Needing to use par_bridge() comes up when using Rayon to process streaming data in parallel:

fn process(input: impl BufRead + Send) -> std::Result<Output, Error> {
    let mut err = Ok(());
    let output = lines
        .input()
        .scan(&mut err, until_err)
        .par_bridge()
        .map(|line| ... executed in parallel ... )
        .reduce(|item| ... also executed in parallel ...);
    err?;
    ...
    Ok(output)
}

Again, equivalent effect cannot be trivially achieved by collecting into Result.

Solution 3

This answer pertains to a pre-1.0 version of Rust and the required functions were removed

You can use std::result::fold function for this. It stops iterating after encountering the first Err.

An example program I just wrote:

fn main() {
  println!("{}", go([1, 2, 3]));
  println!("{}", go([1, -2, 3]));
}

fn go(v: &[int]) -> Result<Vec<int>, String> {
    std::result::fold(
        v.iter().map(|&n| is_positive(n)),
        vec![],
        |mut v, e| {
            v.push(e);
            v
        })
}

fn is_positive(n: int) -> Result<int, String> {
    if n > 0 {
        Ok(n)
    } else {
        Err(format!("{} is not positive!", n))
    }
}

Output:

Ok([1, 2, 3])
Err(-2 is not positive!)

Demo

Solution 4

Handling nested .map() closure Result's

What if we have a .map() within a .map() within a .map()?

Here's an example for the specific case where the .map() operations are nested. The problem it solves is how to propagate a failure from the innermost closure while avoiding using .unwrap() which aborts the application.

This approach also enables using ? syntax at the outer layer to capture the error if one occurs, or unwrap the result to assign to a variable if no error occurred. ? can't otherwise be used from inside the closures.

.parse() as it's used below will return Result<T, ParseIntError>.

use std::error::Error;

const DATA: &str = "1 2 3 4\n5 6 7 8";

fn main() -> Result<(), Box<dyn Error>>
{
    let data = DATA.lines().map(|l| l.split_whitespace()
                                     .map(|n| n.parse() /* can fail */)
                                     .collect())
                           .collect::<Result<Vec<Vec<i32>>, _>>()?;
    println!("{:?}", data);
    Ok(())
}

Note that the outer .collect::<..>() generic expression specifies Result<Vec<Vec<..>>. The inner .collect() will be producing Results, which are stripped away by the outer Result as it takes the Ok contents and produces the 2-D vector.

Without relying heavily on type inference, the inner .collect() generic expression would look like this:

          .collect::<Result<Vec<i32>, _>>()) // <--- Inner.
    .collect::<Result<Vec<Vec<i32>>, _>>()?; // <--- Outer.

Using the ? syntax, the variable, data, will be assigned this 2-D vector; or the main() function will return a parsing error that originated from within the inner closure.

output:

[[1, 2, 3, 4], [5, 6, 7, 8]]

Taking it a step further, parse results nested three levels deep can be handled this way.

type Vec3D<T, E> = Result<Vec<Vec<Vec<T>>>, E>;

const DATA: &str = "1 2 | 3 4\n5 6 | 7 8";

fn main() -> Result<(), Box<dyn Error>>
{
    let data = DATA.lines()
                   .map(|a| a.split("|")
                             .map(|b| b.split_whitespace()
                                       .map(|c| c.parse()) // <---
                                       .collect())
                             .collect())
                   .collect::<Vec3D<i32,_>>()?;
    println!("{:?}", data);
    Ok(())
}

output:

[[[1, 2], [3, 4]], [[5, 6], [7, 8]]]

Or if a number couldn't be parsed, we'd get:

Error: ParseIntError { kind: InvalidDigit }

View more solutions

Comments

[Dogbert]

+1 This is awesome! (The example from my answer ported to this: is.gd/E26iv9)

[Kai Sellgren]

If my own type implements FromIterator, can I apply this same trick? Is this limited to FromIterator or could it be applied in a broader way? How does Rust know to use FromIterator?

[BurntSushi5]

@KaiSellgren Yes, you can apply this same trick. The key is in the type signature of collect, which is polymorphic on the return type, which must implement FromIterator. I don't know what you mean by "can it be applied in a broader way." Rust supports polymorphic return types... So, yes? (See the Rng and Default traits for more examples of return type polymorphism.)

[Kai Sellgren]

That makes sense, but how does it know to call from_iter() method? doc.rust-lang.org/std/iter/trait.FromIterator.html

[BurntSushi5]

@KaiSellgren from_iter is called in the collect method.

[U007D]

Use of collect() requires the iterator to be finite, correct? If so, how would a similar but infinite iterator be handled?

[BurntSushi5]

Sounds like a different question to me. Use a for loop?

[Good Night Nerd Pride]

What would you do in case of multiple map()s? If the first map() returns a Result, then the following map() has to accept a Result as well which can be annoying. Is there a way to achieve the same from the middle of map() chain? Short of just doing .map(...).collect<Result<Vec<_>, _>>()?.into_iter().map(...), of course.

[Shepmaster]

docs.rs/itertools/0.9.0/itertools/fn.process_results.html

[Shepmaster]

as when you want to sum() [...] the Ok items — that's already implemented in the standard library, using the same technique as the process_results method in itertools.

[user4815162342]

@Shepmaster I didn't know about process_results(), thanks. Its upside is that it doesn't require a separate error variable. Its downsides are that it's only available as a top-level function that calls you (possible issue when iterating over several things in parallel), and that it requires an external crate. The code in this answer is reasonably short, works with stdlib, and participates in iterator chaining.