proposal: spec: the #id/catch error model, a rethink of check/handle

[networkimprov]

Please do not down-vote this post if you are against any new syntax for error handling.
Instead, vote in the first comment below. Thanks!

Having heard users' frustrations with Go1 error handling, the Go team has committed to delivering a new method. Ideally, a solution would stem from a familiar language. The Go2 Draft Design is fine for wrapping errors with context, and returning them succinctly. But its feel is novel, and it has significant drawbacks, discussed in Golang, how dare you handle my checks!

Besides returning errors, Go programs commonly:
a) handle an error and continue the function that received it, and
b) have two or more kinds of recurring error handling in a single function, such as:

{ log.Println(err) }
{ debug.PrintStack(); log.Fatal(err) }
{ if err == io.EOF { break } }
{ conn.Write([]byte("oops: " + err.Error())) } // e.g. a network message processor

There is indeed a long list of Requirements to Consider for Go2 Error Handling. The check/handle scheme accommodates a tiny subset of these, necessitating an awkward mix of Go1 & Go2 idioms, e.g.

handle err { return fmt.Errorf(..., err) }
v, err := f()
if err != nil {
   if isBad(err) {
      check err     // means 'throw'
   }
   // recover
}

Herein is a widely applicable approach to error handling, leveraging the C-family catch block. For the record, the author is grateful that Go does not provide C++ style exceptions, and this is not a plot to sneak them into the language through a side door :-)

The #id/catch Error Model

Let a catch identifier (catch-id) e.g. #err select a named handler. A single catch-id may appear in any assignment. A handler is known by its parameter name; the parameter can be of any type. A handler follows the catch-id(s) that trigger it and starts with catch <parameter>. Catch-ids are not variables and handler parameters are only visible within handlers, so there's no re-declaration of error variables.

These are not unique ideas. At last count, 17 posts on the feedback wiki suggest various ways to define and invoke named handlers, and 13 posts suggest invocation of handlers using assignment syntax.

func (db *Db) GetX(data []byte) (int, error) {
   n, #return := db.name()                       // return our own errors
   
   f, #err := os.Open(n)                         // this file's presence is optional
   defer f.Close()
   _, #err  = f.Seek(42, io.SeekStart)
   l, #err := f.Read(data)
   
   #return = db.process(data)
   
   catch err error {                             // handle OS errors here
      if !os.IsNotExist(err) { log.Fatal(err) }
      log.Println(n, "not found; proceding")
   }

   #return = db.index(data)                      // executes unless catch exits
   return l, nil
}

Several points are unresolved, see Open Questions below. Catch-id syntax is among them; #id is reminiscent of the URL form for goto id, but ?id, @id, and others are viable.

Advantages: similarity to established try/catch method (but without a try-block); clarity as to which handler is invoked for a given statement; certain statements may be skipped after an error occurs; handlers can return or continue the function.

Please help clarify (or fix) this proposal sketch, and describe your use cases for its features.

Feature Summary

Draft-2, 2018-09-19. Discussion following this comment below pertains to this draft.

These features meet a large subset of the Requirements to Consider for Go 2 Error Handling.

We can select one of several distinct handlers:

func f() error {
   v1, #fat := fatalIfError()     // a non-zero value for #id triggers corresponding catch
   v2, #wrt := writeIfError()

                                  // predefined handlers
   v3, #_   := ignoreIfError()    // or log the error in debug mode
   v4, #r   := returnIfError()    // aka #return
   v5, #p   := panicIfError()     // aka #panic
   
   catch fat       { log.Fatal(fat) }           // inferred parameter type
   catch wrt error { conn.Write(wrt.Error()) }
}

We can invoke a handler defined at package level (thanks @8lall0):

func f() error {
   #pkg = x()
}

catch pkg error {           // package-level handler; explicit type
   log.Println(pkg)
   return pkg               // return signature must match function invoking pkg handler
}

We can specify a type for implicit type assertion:

   f := func() error { return MyError{} }
   #err = f()
   catch err MyError { ... }

We can skip statements on error and continue after the handler:

   #err = f()
   x(1)                              // not called on error
   catch err { log.Println(err) }
   x(2)                              // always called

We can forward to a different handler (creates an explicit handler chain):

   #ret = x()
   if ... {
      #err = f()
      catch err {
         if ... { #ret = err }         // invoke alternate handler
         #ret = fmt.Errorf(..., err)   // invoke handler with alternate input
      }
   }
   catch ret { ... }

We can reuse catch-ids:

   #err = f(1)
   catch err { ... }
   #err = f(2)
   catch err { ... }

We can nest catch blocks:

   #era = f(1)
   catch era {
      #erb = f(2)
      catch erb { ... }      // cannot use 'era'; shadowing in catch disallowed
   }

We can see everything from the scope where a handler is defined, like closure functions:

   v1 := 1
   if t {
      v2 := 2
      #err = f()
      catch err { x(v1, v2) }
   }

We can still use Go1 error handling:

   v1, err := x()  // OK
   v2, err := y()  // but re-declaration might be abolished!

Open Questions

• What catch-id syntax? #id, ?id, @id, id!, $id, ...
  What style for predefined handlers? #r, #p, #_, #return, #panic, #nil, ...

• What handler definition syntax? catch id [type], catch id(v type), id: catch v [type], ...
  Infer parameter from previous stmt? #err = f(); catch { log.Println(err) }

• Invoke handler when ok=false for v, #ok := m[k]|x.(T)|<-c, etc?
  Pass a type error with context? v, #err := m[k]; catch { log.Println(err) }

• Treat parameter as const? catch err { err = nil } // compiler complains
  Lets forwarding skip test for nil: catch err { #ret = err }

• Require #id for return values of type error? proposal: spec: require return values to be explicitly used or ignored (Go 2) #20803

• Provide check functionality with f#id()? e.g. x(f1#_(), f2#err())
  If so, disallow nesting? x(f1#err(f2#err()))
  Allow position selector? f#id.0() tests first return value

• Provide more context to package-level handlers, e.g. caller name, arguments?
  catch (pkg error, caller string) { ... }

• Allow handlers in defer stack?

   defer last()              // skip if handler returns
   defer catch errd { ... }
   defer next#errd()         // skip if first() invokes handler
   defer first#errd()
  

• Allow multiple handler arguments?

   #val, #err = f()                 // return values assignable to catch parameter types
   catch (val T, err error) { ... } // either parameter could be non-zero
  

Disallowed Constructs

Declaring or reading a catch-id:

   var #err error           // compiler complains
   #err = f()
   if #err != nil { ... }   // compiler complains
   catch err { ... }

Multiple catch-ids per statement:

   #val, #err = f()   // compiler complains
   catch val { ... }  // if f() returns two non-zero values, which handler is executed?
   catch err { ... }

Shadowing of local variables in handlers:

func f() {
   if t {
      err := 2
      #err = f()            // OK; #err handler can't see this scope
   }
   pkg := 1                 // OK; #pkg handler (see above) can't see local variables
   err := 1
   #err = f()
   catch err { return err } // compiler complains; err==1 is shadowed
}

Self-invocation:

   #err = f(1)
   catch err {
      #err = f(2)        // compiler complains
   }
   #err = f(3)
   catch err { ... }

Unused handlers:

   catch err { ... }         // compiler complains
   #err = f()
   catch err { ... }
   #ret = f()
   catch err { return err }  // compiler complains
   catch ret { ... }
   catch ret { return ret }  // compiler complains

Discarded Ideas

Chain handlers with same catch-id in related scopes implicitly, as in the Draft Design:

func f() {
   v, #fat := x()
   if v != nice {                      // new scope
      #fat = y(&v)
      catch fat {                      // invoked 1st
         if ... { #fat = nil }         // can skip other handlers in chain
      }                                // no return/exit, continue along chain
   }
   catch fat { log.Fatal(fat) }        // invoked 2nd
}

Changelog

2018-09-19 draft-2 (discussion below)
a) Move implicit handler chain to new section "Discarded Ideas".
b) Make continuing after catch the default behavior.
c) Document catch-id reuse and nested catch block.
d) Disallow unused handlers (was "contiguous handlers with same catch-id") and self-invocation.
e) Add #_ predefined handler to ignore or log input.
f) Add implicit type assertion.

Why Not check/handle?

Please read Golang, how dare you handle my checks! for a discussion of each of the following points.

• No support for multiple distinct handlers.
• The last-in-first-out handle chain cannot continue a function.
• check is specific to type error and the last return value.
• The per-call unary check operator can foster unreadable constructions.
• The default handler makes it trivial to return errors without context.
• Handlers appear before the calls that trigger them, not in the order of operations.
• The handle chain is inapparent; one must parse a function by eye to discover it.

Also, there is relatively little support for the draft design on the feedback wiki.

Named Handlers Are Popular!

At last count, roughly 1/3rd of posts on the feedback wiki suggest ways to select one of several handlers:

1. @didenko github
2. @forstmeier gist
3. @mcluseau gist
4. @the-gigi gist
5. @PeterRK gist
6. @marlonche gist
7. @alnkapa github
8. @pdk medium
9. @gregwebs gist
10. @gooid github
11. @spakin gist
12. @morikuni gist
13. @AndrewWPhillips blogspot
14. @bserdar gist
15. @martinrode medium
16. @dpremus gist
17. @networkimprov this page

And the following posts suggest ways to invoke a handler with assignment syntax:

1. @oktalz gist
2. @pborman gist
3. @kd6ify blog
4. @rockmenjack github
5. @the-gigi gist
6. @8lall0 gist
7. @dpremus gist
8. @bserdar gist
9. @mcluseau gist
10. @didenko github
11. @gooid github
12. @Kiura gist
13. @networkimprov this page

/cc @rsc @mpvl @griesemer @ianlancetaylor @8lall0 @sdwarwick @kalexmills

Thanks for your consideration,
Liam Breck
Menlo Park, CA, USA

[networkimprov]

Go error handling works fine today, no new error specific syntax is necessary.

Some new general-purpose features (macros?) would ease error handling.

[PeterRK]

Mix of Go1 & Go2 idioms is not a problem.

[xiaoxubeii]

I don't like some kind of "Error-Handle", especially:

When an error check fails, it transfers control to the innermost handler, which transfers control to the next handler above it, and so on, until a handler executes a return statement.

It is hard to find which handle will catch the error unless count the code above. And it is not possible to specify the error handle.

[gregwebs]

I am surprised to see this posted here as a github issue. Can we have more transparency and guidance on the go 2 draft proposal feedback process? As it stands, we have only been directed to leave feedback on a wiki with no indication of future steps.

[ianlancetaylor]

Yes, we prefer feedback on the wiki. Thanks.

That said, this seem less like feedback on the design draft, and more like a completely different proposal.

[deanveloper]

A major issue with this is the same as with gotos.

func errorProne() {
    x, #err := oops() // goes straight to `catch err error`
    i := 10 // define variable i
    
    catch err error {
        fmt.Println(i) // can't print i, as it was never defined, even though it is in-scope
    }
}

Currently, goto gets around this by saying no variables may be defined inside any region between goto lbl and lbl: (although there are a couple proposals that are refining this to make things a bit easier)

Also, what about this?

func errorProne() {
    x, #err := oops() // err not nil
    y, #err2 := oops2() // err2 not nil

    fmt.Println(0)

    catch err error {
        fmt.Println(1)
    }
    
    fmt.Println(2)

    catch err2 error {
        fmt.Println(3)
        return
    }
    
    fmt.Println(4)
}

What order do these print in? Because it looks like the program runs top-down, does it print 1 3, or 1 2 4? I think the check/handle construct made this a bit clearer since it was extremely clear that the program does not run top-down, but because in this example, since the catches are defined below the #ids, it makes it a lot more difficult to recognize the flow of the program just by looking at it.

[DeedleFake]

What if it gets treated like a variable? catch err error would declare a new catch called err of type error, and only once it's in scope can you use it.

[networkimprov]

Re goto that will be fixed: #27165

Re errorProne(), there is no check/handle analog for #err, as handle cannot continue the function. Since #err2 returns, you write its catch immediately below the call; elsewhere should be flagged. Triggering a catch-id is like a throw, so oops() failure skips oops2(). Results: ok 0,2,4; #err 1,2,4; #err2 3

N.B. in #id/catch draft 1, you need #id = nil to continue, but that may become the default next draft.

@DeedleFake, yes the handler parameter is only accessible in that handler.

[networkimprov]

It is hard to find which handle will catch the error unless count the code above. And it is not possible to specify the error handle.

@xiaoxubeii, lots of people agree, see Named Handlers Are Popular above.

#id/catch draft 1 offers implicit (i.e. automatic) handler chains as in check/handle, and explicit ones where one handler forwards the error to another by name. However I believe implicit chaining should be dropped, because it is not obvious when chained handlers are far apart in a function.

[deanveloper]

Re goto that will be fixed: #27165

It's not "fixed", it's just expanded upon. You can now declare variables under that construct, but you can't use them outside of the range between the goto lbl and the lbl:. My example still should not compile, even though it looks intuitive.

Since #err2 returns, you write its catch immediately below the call; elsewhere should be flagged.

Assuming you mean "flagged" as in give a compiler error (as there are no warnings in Go), would the following be illegal?

func errorProne() error {
    x, #err := oops() // err not nil
    y, #err := oops2() // err not nil
    catch err error {
        return wrapErr("error in errorProne", err)
    }
}

Since catch err returns, it needs to be placed immediately after... So since there is code between the oops() call and the catch, it should fail to compile, correct? I'm not sure I'm understanding this.

If so, then all you've done is replaced if err != nil with catch err, which is definitely not a good solution.

Re errorProne(), there is no check/handle analog for #err, as handle cannot continue the function. Since #err2 returns, you write its catch immediately below the call; elsewhere should be flagged. Triggering a catch-id is like a throw, so oops() failure skips oops2(). Results: ok 0,2,4; #err 1,2,4; #err2 3

Okay, so I understand the flow now, but I'm not sure I like it. In order to understand it, my eyes need to dart all around the function in order to follow it, when it should be going a single direction.

The nice thing about check/handle was that the function went downward as normal, but then once an error occurs, it flows (block-wise) upward.

Going back to the original draft -

func f() error {
   #pkg = x()
   catch pkg { ... } // optional
}

catch pkg {                 // package-level handler
   log.Println(pkg.Error())
   return pkg               // return signature must match function invoking pkg handler
}

I don't like this - I want my returns to be within the function... a function's flow should not be determined by something outside of the function (ever). Also, how would this work with multi-value returns?

---

Again, my main gripe with the #id/catch proposal is that it gives the illusion that the function is running from the top-down. The check/handle very obviously shows that it does NOT run top-down, because the handler for an error needs to be declared above the check. I do like the concept of a named error handler, but I'm not sure that this is the way to do it. (Sorry if I'm coming off as mean or something, just trying to give constructive criticism)

[gregwebs]

@ianlancetaylor the majority of the feedback on the wiki would probably qualify as a new design. All of it solves the same problem overview scenario, and most contains a new control-flow operation and a technique for re-usable handlers.