For the record, I don’t mind if err != nil, and all things being equal, I would rather not remove that boilerplate, since errors are just values and all that jazz. I know many, many long-time gophers that agree.

However, if we’re going to change the language, this is a much better change than the proposed try in my opinion. This maintains legibility of code, maintains ease of finding exit points in the code, still reduces boilerplate, and encourages people to add context to an error.

IMO it removes the major faults of try, which are the high likelihood of missing a call to try embedded in the right hand side somewhere when reading code, and the problem of encouraging nesting of functions which is distinctly bad for code readability.

My main reason for disagreeing with that is that it couples the fmt-package into the language. i.e. to actually add this to the spec, you’d also need to define what fmt does in the spec - and IMO, fmt is far too large (conceptually) to be part of the language proper. So, IMO, the multiple-argument form you are suggesting should be removed.

This is many ways worse than try() function proposal in my opinion because it doesn’t really solve the underlying goal of reducing boilerplate code.

A function with multiple error check looks like this:

func Foo() (err error) {
	var file1, err1 = open("file1.txt")
	catch(err1)
	defer file1.Close()
	var file2, err2 = open("file2.txt")
	catch(err2)
	defer file1.Close()
}

Vs. originally proposed try() function:

func Foo() (err error) {
	var file1 = try(open("file1"))
	defer file1.Close()
	var file2 = try(open("file1"))
	defer file2.Close()
}

Later looks more concise and elegant in my opinion.

@ubikenobi It’s always trade-offs. In my opinion, the reduction provided by try costs far too much in readability. More so when the extra “flexibility” try permits is weighed. The clarity that comes from that single newline (and possibly two columns of text) is worthwhile. The surety that comes from reduced flexibility? Even more critically so.

@ngrilly

Why? This is one of the main use case for similar constructs in Rust, Swift, Haskell, etc.

I understand one of Go’s fundamental principles to be that errors are not exceptional but rather common, and that the code to deal with them (i.e. the sad path) is at least as important, and often more important, than the code to accomplish the main work (i.e. the happy path). I also understand that the relative reliability and predictability of Go programs comes in large part from this principle, that errors are lifted up front and center for every expression that can generate them, forcing—or at least strongly encouraging—the programmer to deal with them in situ.

Error chaining subverts this principle in the name of expediency. I don’t disagree that in some cases this can be useful. But I would hate to see Go become a language where it is common. We would lose one of its most important strengths.

One variation would be:

• to start, catch only takes one argument
• don’t allow catch to take additional arguments for format string and args
• after experience or experimentation, consider including a handler function as an optional second argument, such as:

  decorate := func(err error) error { return fmt.Errorf("foo failed: %v", err) }

  ...

  f, err := os.Open(config)
  catch(err, decorate)          // decorate called in error case

or using helper functions that could live somewhere in the standard library:

  catch(err, fmt.Annotatef("foo failed for %v", arg1))      

That would be more parallel to the current proposal for try as a builtin.

edit: part of the rationale for a handler function is it means a builtin does not directly rely on fmt semantics, which is the concern @Merovius expressed better in the next comment.

If the error solution does not support chaining then its value is reduced. If it does not eliminate the extra lines of code between statements, its value is reduced.

I fully agree with this.

If y’all want to manifest something other than try() in 1.14, you need to first prove that try() isn’t useful for the great majority of existing code.

The Go team has told us that if we can’t prove that, the merits of alternatives are moot.

EDIT: This is a fact. Down-voting facts doesn’t make them fiction.

@apghero why is it a nonstarter? That’s basically the whole difference between this and the try proposal. I do not want a function on the right hand side of a complicated statement to be able to exit. I want it to have to be the only thing on the line so it is hard to miss.

The most common error handling code I have written is the classic:

        ..., err := somefunc1()
        if err != nil {
                return err
        }
        ..., err := somefunc2()
        if err != nil {
                return err
        }

Like @ubikenobi, I think the catch solution (or the on solution just suggested) adds little value:

        ..., err := somefunc1()
        catch(err)
        ..., err := somefunc2()
        catch(err)

while try lets you focus on the code, not the errors:

        ... := try(somfunc1())
        ... := try(somfunc2())

Just as well quickly learned that r is probably an io.Reader and w is probably an io.Writer, we will quickly learn to ignore the try while reading code.

The try proposal really shines when you want to have chaining types and still support errors. Consider:

        v, err := Op1(input)
        if err != nil {
                return err
        }
        v, err = v.Op2()
        if err != nil {
                return err
        }
        result, err := v.Op3()
        if err != nil {
                return err
        }

With try this becomes:

        result := Op1(input).Op2().Op3()

This would actually encourage this paradigm more, which I think would be a good thing.

Catch, on the other hand, make it:

        v, err := Op1(input)
        catch(err)
        v, err = v.Op2()
        catch(err)
        result, err := v.Op3()
        catch(err)

Quite honestly, if the error solution does not support chaining then its value is reduced. If it does not eliminate the extra lines of code between statements, its value is reduced.

One scenario where try is clearly not helpful and the check proposal has a very slight advantage over straight Go is:

        ..., err := somefunc1()
        if err != nil {
                return ..., fmt.Errorf("somefunc1: %v", err)
        }
        ..., err := somefunc2()
        if err != nil {
                return ..., fmt.Errorf("somefunc2: %v", err)
        }

While catch would add some slight value, it certainly does not carry its own weight vs the existing paradigm give it’s other issues.

I can contrive a solution that uses try, but it is worse than the original code because each return requires different processing.

As mentioned, the try solution can easily be extended in the future by adding additional parameters, though they need to be readable.

As for the name try, I am not super enamored with it, it will make many people think “where is the catch?”. I could suggest must, instead, but right now we associate must with a panic. I don’t know a better word choice so try is okay. I certainly don’t want something like returnOnError even though it is pretty descriptive!

I think this is not much better than the try proposal. All we need is to be able to write if err { ... } without having to put != nil.

I support this proposal.

It is important that any change to error handling allows each error in a function to be annotated explicitly and independently from other errors. Returning the naked error, or using the same annotation for every error return in a function, may be useful in some exceptional circumstances, but are not good general practices.

I also support using a keyword instead of a builtin. If we add a new way to exit a function, it is important that it isn’t nestable within other expressions. If we have to wait until further improvements to the language allow this to be accomplished without breaking the Go 1 compatibility promise, then we should do that, rather than make the change now.

edit: @networkimprov’s #32611 also satisfies my requirements, and I support it as well.

At first glance, this is generally what try and check should have been. check is a great name (and way to avoid the baggage of catch), whether meaning “verify” or “halt/slow progress”.

Why not use existing keywords like:

func ReadFile(filename string) ([]byte, error) {
    f, err := os.Open(filename)
    return if err != nil
    defer f.Close()
    return ioutil.ReadAll(f)
}

1. No new language rule for implicit bool(err != nil)
2. So other conditions than err != nil are also possible:

    f, err := os.Open(filename)
    return if errors.Is(err, os.ErrNotExist)

1. return without named values is already in the language
2. Only implicit magic is returning err when used in the if statement, but explicit return values could be used like:

    return x, y, err if err != nil

A more complicated example doesn’t seem to have advantages in terms of readability compared to the traditional if clause, though:

    f, err := os.Open(filename)
    return nil, errors.Wrap(err, "my extra info for %q", filename) if errors.Is(err, os.ErrNotExist)

I think this is not much better than the try proposal. All we need is to be able to write if err { ... } without having to put != nil.

Thank you sir, I do think Go should allow implicit casting from nil interface to bool.

@daved Inconsistencies and unnecessary code clutter also negatively impact readability in my opinion. This is why I think catch() proposal is less readable and less safe compared to try() proposal. I’ll try to illustrate my three main concerns with a different example below.

• It enables inconsistent ways (example1/2) to call catch which really hinders control flow in my opinion.
• The compiler can only validate callee function to have an error as the last parameter, not the called function.
• Simple coding mistakes like calling catch in incorrect order (example 2) can lead to hard to find/ dangerous bugs.

//Example 1
func Dispense() (err error) {
	var motErr= SetMotorSpeed(5)
	catch(motErr)
	var fanErr= StartFan(5)
	catch(fanErr)
	var tempErr =SetTemperature(200)
	catch(tempErr)
	var itErr = DropItem()
	catch(itErr)
}
//Example 2
func Dispense() (err error)
	var motErr= SetMotorSpeed(5)
	var speedErr= StartFan()
	var tempErr =SetTemperature(200)
	var printErr = DropItem()
	catch(motErr)
	catch(fanErr)
	catch(tempErr)
	catch(itErr)
}

As appose to catch(), with try() function (example below):

• I gain consistency by ensuring every try() call behaves the same way. No need to declare err variables, or return error returns in the correct order.
• Ensures both called and callee function(or method) define an error as the last return parameter.
• With unnecessary code clutter motErr, catch() gone, makes machine logic vastly clear and easier to read in my opinion.
• My machine code runs safer by guaranteed return on err on try(). SetTemperature() not possible without successfully calling StartFan().

func Dispense() (err error){
	try(SetMotorSpeed(5))
	try(StartFan(5))
	try(SetTemperature(200))
	try(DropItem())
}

I would like to point out that when flexibility is needed err!=nil pattern works better than both catch() and try() proposal IMO. The main goal of the original try() as a function proposal s to remove unnecessary boilerplate when flexibility is NOT needed.

You can simplify this a bit. Since it expects an error value, you can give it any error value. So then it can be catch err or catch fmt.Errorf(“can’t open config for %s”, user) — no need to implicitly use this error generating function and you can return any kind of error that makes sense in your situation. Though I don’t like the use of catch. It would be perfectly legal in your scheme to say something like this but it looks weird!

    If !inrange(low, value, high) {
        catch OutOfRangeError
    }

Except for the name, this would be a perfectly legitimate use at the where error conditions are first detected and even better than having to say

    If !inrange(low, value, high) {
        return nil, OutOfRangeError
    }

On IBM systems on can use a macro ABEND to abnormally terminate a task. It was a bit like panic but the name was suggestive of its intended use. catch doesn’t have that. Neither did try. I thought of returnOnError or errorReturn or just ereturn or abreturn or may be even raise as in C. Any of these work for me better than catch or try but none seem ideal.

@fabian-f

Why not just make macros a feature and let users implement their own shorthands?

Because we have no time to learn every other library’s own flavor of Go just to be able to comprehend what is it doing. We’ve been there, done that @ C, we’ve been suffocated by it in C++. RiP Rust was such a wünderkid until got the feature-fever and passed away. I wish not the same with Go.

Could even shorthand the decoration #define catch(a, msg) if a != nil { return fmt.Errorf("%s: %w", msg, a) }

…and possibly will send user keys to the moon in yet other decoration you’ll skim over.

@ubikenobi As far as I understand, your example (where only errors are returned) can be duplicated with this proposal.

func Dispense() (err error){
	catch(SetMotorSpeed(5))
	catch(StartFan(5))
	catch(SetTemperature(200))
	catch(DropItem())
}

The main goal of the original try() as a function proposal s to remove unnecessary boilerplate when flexibility is NOT needed.

And yet it is designed to facilitate flexibility. If what you suggest is true, it’s flexibility can/should be reduced to only that which is necessary.

@pborman

… we will quickly learn to ignore the try while reading code.

That’s the problem.

Also, function chaining is not something I’m on board to promote. It’s only really useful for accumulator style types. If it’s, otherwise, some machine being controlled, then the relevant functions don’t need to return errors at all since “errors are value” and whatnot. I would use a container type that either provides a final error checking function, or use an error channel with the “body” being called in a goroutine.

@pborman

if the error solution does not support chaining then its value is reduced.

It seems you and I want different things. I specifically made this proposal because I don’t want to encourage people to jam more logic into a single line. I don’t want to encourage method chaining. I don’t want to hide away error handling. Having to write catch(err) every time your function might exit is good IMO. I want to be very explicit about what can fail and where the function can exit.

@pborman

[try] would actually encourage this paradigm more, which I think would be a good thing.

I don’t think it’s a good idea to encourage returning ~the naked~ an un-annotated error. As a general rule, errors should be individually annotated with context before being yielded up the callstack. Although if err != nil { return err } is the trope often cited in Hacker News comments or whatever, it isn’t great practice.

A more flexible one-extra-line proposal is #32611

err = f()
on err, <single_statement>

EDIT: And here is a catch proposal where catch does what you’d expect: #27519

This still would take up a keyword: catch, but only barely improve on the try variant.

The point is to reduce the footgun rating of a solution, not to offer an “improved” novelty.

Many of the arguments again try is that it not improves anything, just dumbifies error handling. This catch can just be written as a package that provides this as a single function. As @networkimprov stated, as long as there’s no proof that would improve the try case it’s void of use.

What for users that use custom errors that use wrapping functions?

This was covered in comments.

Comments are personal discussions, not part of the proposal, the proposal only has catch with error formatting.

If we plan on taking up a whole keyword, can we at least attach some function to it?

Now I’m not sure if this is a troll post. Most keywords serve exactly one function. return, if, etc.

Return takes multiple values and it’s not a function, it has a space after it, clearly marking it as a keyword. Reserved items should be space separated just so it’s clear they’re not a local function.

All of this can be replaced by having a local package called: errorHandler with 1 public function: Catch. No need to change the language for the sake of changing it (same does technically count for try() proposal).

@thepudds

f, err := os.Open(config)
check(fmt.ErrorHandlerf(err, "error opening config for user %s", user))

FWIW, this code would also work with try. The changes between the other proposal and this one would seem to be: a) rename try to check and b) remove the return value.

The former seems fairly minor.

The latter seems more significant and will essentially require the builtin to be used in a statement context. Rolling out try (or check) in two separate phases as you suggest actually sounds pretty good to me.

This seems to be nothing more than an equivalent to a macro such as the following:

#define catch(a) if a != nil { return }

Why not just make macros a feature and let users implement their own shorthands? Could even shorthand the decoration

#define catch(a, msg) if a != nil { return fmt.Errorf("%s: %w", msg, a) }

@Chillance The problem with macros is that tools (especially refactoring tools) have to expand the code to understand the semantics of the code, but must produce code with the macros unexpanded. It’s hard. Look at this for example in C++: http://scottmeyers.blogspot.com/2015/11/the-brick-wall-of-c-source-code.html. I guess it’s a little bit less hard in Rust than in C++ because they have hygienic macros. If the complexity issue can be overcomed, I could change my mind. Must be noted it changes nothing to the discussion about try, since Rust which has macros, has implemented a strict equivalent of try as a built-in macro 😉

@Chillance I’m personally opposed to adding hygienic macros to Go. People are complaining about try making the code harder to read (I disagree and think it makes it easier to read, but that’s another discussion), but macros would be strictly worse from this point of view. Another major issue with macros is that it makes refactoring, and especially automated refactoring, so much harder, because you usually have to expand the macro to refactor. I repeat myself, but this is a whole can of worms. try is 1 year old toy compared to this.

@nikolay

All we need is to be able to write if err { ... } without having to put != nil.

It’s a much larger change than try. It would imply to change the language spec to accept implicit casting from nil to bool, not just in if, but everywhere (an exception for if would be really weird). The implicit casting could introduce hard to track bugs. And we would still need to change gofmt to format this on 1 line instead on 3.

I came here to propose a very similar solution. I do think this proposal is simpler and clearer than the try(…) proposal put forward by the Go team, and would have little impact on existing code except to cut a lot of boilerplate.

I’d use check as the name, catch has too many other connotations (and is used elsewhere by the calling function, not the one throwing). I would prefer if you could pass an optional error as a second argument, to allow neatly wrapping errors in one line. That would get rid of problems with requiring fmt etc. and make the proposal simpler and still allow inline wrapping. You could even use a function returning error as the second argument then and replace some of the line noise caused by the if err := … ; err != nil { idiom.

To show some examples, which I think it improves:

f, err := os.Open(filename)
if err != nil {
	return …, err  // zero values for other results, if any
}

becomes:

f, err := os.Open(filename)
check(err) 

And the much longer example from the overview goes from:

func CopyFile(src, dst string) error {
	r, err := os.Open(src)
	if err != nil {
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}
	defer r.Close()

	w, err := os.Create(dst)
	if err != nil {
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}

	if _, err := io.Copy(w, r); err != nil {
		w.Close()
		os.Remove(dst)
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}

	if err := w.Close(); err != nil {
		os.Remove(dst)
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}
}

to this shorter example where the extra 1 line for error handling helps a lot with clarity when compared with the try example IMO, and even optionally replace the if err := xxx; err != nil { pattern with something simpler:

func CopyFile(src, dst string) error {
	r, err := os.Open(src)
        check(err, fmt.Errorf("copy %s %s: %v", src, dst, err))
	defer r.Close()

	w, err := os.Create(dst)
        check(err, fmt.Errorf("copy %s %s: %v", src, dst, err))

       
        _, err :=  io.Copy(w, r)
        check(err, func() error {
		w.Close()
		os.Remove(dst)
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}())

        err := w.Close()
        check(err, func() error {
		os.Remove(dst)
		return fmt.Errorf("copy %s %s: %v", src, dst, err)
	}())
}

The main reason I like this proposal is that it doesn’t obfuscate the function calls like os.Open(src), w.Close() as try would, and doesn’t encourage chaining or wrapping as try would.

The error handling starts on the line after, as now, but can be shorter if required (even in the case of annotation, but especially when there is no annotation). I don’t want fewer lines of code at the expense of clarity, so one extra line at times vs try is a win in my book.

This proposal gets rid of the bit people don’t like (verbose error handling), and leaves the rest of the code alone.

@pborman,

Just as well quickly learned that r is probably an io.Reader and w is probably an io.Writer, we will quickly learn to ignore the try while reading code.

The try proposal really shines when you want to have chaining types and still support errors. Consider:

        v, err := Op1(input)
        if err != nil {
                return err
        }
        v, err = v.Op2()
        if err != nil {
                return err
        }
        result, err := v.Op3()
        if err != nil {
                return err
        }

With try this becomes:

        result := Op1(input).Op2().Op3()

Actually it becomes

         result := try(try(try(Op1(input)).Op2()).Op3())

Finding out which try failed would be a trying exercise! In this single expression there are quite a few basic blocks! It would be very hard to learn to “unsee” the trys.

@jesse-amano The must keyword you propose doesn’t differentiate between a standard error returned by a function call, and a real panic triggered by a bug in your code.

People keep repeating try will make our code unreadable, but Rust …

All due respect to Rust, please don’t use it as an example of readability. For it’s own goals, it is clearly a great language. For the benefits that Go provides and I am drawn to, there is no relation whatsoever.

@natefinch, I think you should support no changes to the language in that case. It does everything you want already (expect that you still need the if statement, which is even more precise if err != && err != io.EOF ...)

I would support no change to the language over check as I believe check has no real benefits over how Go is today. My feeling is if we are going to add something to Go, the try proposal at least seems like a choice that provides actual benefit at little cost. I actually like it and am not opposed to it going in. If it doesn’t go in, I am okay, as long as none of the other proposals I have seen do not go in, either.

@pborman

Probably should not call this a naked return

You’re right, I’ve edited to call this an un-annotated error.

If your error handling becomes to complicated (unique handling per error)

My claim is that this is not complicated error handling, this is [what should be] the default way to handle errors, and that anything less is shortcuts that shouldn’t be encouraged in the general case.

IMO, this doesn’t really improve on if err != nil { ... } fmt on one line.

This still would take up a keyword: catch, but only barely improve on the try variant. While I do think this is a step forward, it still only allows “lazily” error handling.

What for users that use custom errors that use wrapping functions? What for users that want to log for developers & operators and return human-understandable errors to the users?

if err != nil {
    log.Println("Hey operating, big error, here's the stack trace: ...")
    return myerrorpackage.WrappedErrorConstructor(http.StatusInternalServerError, "Some readable message")
}

If we plan on taking up a whole keyword, can we at least attach some function to it? Counter proposal to the whole try or catch: do try/catch with check & handle (better keywords imho), scope specific.

Suggestion is part of the whole mix of error handling proposals: scoped check/handle proposal

WRT Backwards compatibility (catch <error> isn’t backwards compatible)… I’d be more than happy to wait for something like this until modules becomes the default build mode ala https://blog.golang.org/go2-next-steps

@provPaulBrousseau, IMHO catch or its equivalent shouldn’t be used if you are returning an error as well as a legit value. [This is another reason I bemoan the lack of a sum type in Go. With sum type you’d use it in most cases where an error or a legit value is returned but not both. For examples such as yours, a tuple or a product type would be used. In such a language catch would only work if the parent function returned a sum type]

But if you do use catch, it would return “”, err.

(Ignoring its use as giving up the processor to another coroutine or thread) yield is better than catch! returnOnError or returnIfError exactly captures the semantics but it is a bit long.

returnIf can also work provided error values have err or error in their name.

returnIf(OutOfRangeError)

So, it can’t be a keyword, because that’s not backwards compatible. So it would need to be catch(OutOfRangeError) …

I do agree that the name might need some tweaking. Naming is hard, and I welcome suggestions for another name that might be more appropriate.

The problem is really that it is most accurately called maybeReturn

How about yield? Where it’s keep going, or stop if something is in the way? Also yield is to produce, like it produces the error.

f, err := os.Open(config)
yield(err) 
defer f.Close()

and then it could be

yield(OutOfRangeError)

which looks better to me