errors: add ErrUnsupported

[ianlancetaylor]

UPDATE: This proposal has shifted from the original description to the one described in comments below.

• #41198 (comment)
• #41198 (comment)
• #41198 (comment)
• #41198 (comment)

Go has developed a pattern in which certain interfaces permit their implementing types to provide optional methods. Those optional methods are used if available, and otherwise a generic mechanism is used.

For example:

• io.WriteString checks whether an io.Writer has a WriteString method, and either calls it or calls Write.
• io.Copy checks the source for a WriterTo method, and then checks the destination for a ReaderFrom method.
• net/http.(*timeoutWriter).Push checks for a Push method, and returns ErrNotSupported if not found.

The io/fs proposal (#41190) proposes various other optional methods, such as ReadFile, where there is again a generic implementation if the method is not defined.

The use of WriterTo and ReaderFrom by io.Copy is awkward, because in some cases whether the implementation is available or not is only known at run time. For example, this happens for os.(*File).ReadFrom, which uses the copy_file_range system call which is only available in certain cases (see the error handling in https://golang.org/src/internal/poll/copy_file_range_linux.go). When os.(*File).ReadFrom is called, but copy_file_range is not available, the ReadFrom method falls back to a generic form of io.Copy. This loses the buffer used by io.CopyBuffer, leading to release notes like https://golang.org/doc/go1.15#os, leading in turn to awkward code and, for people who don't read the release notes, occasional surprising performance loss.

The use of optional methods in the io/fs proposal seems likely to lead to awkwardness with fs middleware, which must provide optional methods to support higher performance, but must then fall back to generic implementations with the underlying fs does not provide the method.

For any given method, it is of course possible to add a result parameter indicating whether the method is supported. However, this doesn't help existing methods. And in any case there is already a result parameter we can use: the error result.

I propose that we add a new value errors.ErrUnimplemented whose purpose is for an optional method to indicate that although the method exists at compile time, it turns out to not be available at run time. This will provide a standard well-understood mechanism for optional methods to indicate that they are not available. Callers will explicitly check for the error and, if found, fall back to the generic syntax.

In normal use this error will not be returned to the program. That will only happen if the program calls one of these methods directly, which is not the common case.

I propose that the implementation be simply the equivalent of

var ErrUnimplemented = errors.New("unimplemented operation")

Adding this error is a simple change. The only goal is to provide a common agreed upon way for methods to indicate whether they are not available at run time.

Changing ReadFrom and WriteTo and similar methods to return ErrUnimplemented in some cases will be a deeper change, as that could cause some programs that currently work to fail unexpectedly. I think the overall effect on the ecosystem would be beneficial, in avoiding problems like the one with io.CopyBuffer, but I can see reasonable counter arguments.

[darkfeline]

This seems to overlap with the existing way of testing whether a method/interface is supported:

foo, ok := bar.(Foo) // Does bar support Foo?

Will we end up in situations where callers need to do two checks for whether a method is supported (the type assertion and an ErrUnimplemented check)? I know we can forbid that by convention, but it seems like a really easy trap to fall into.

[ianlancetaylor]

This new error is intended to be used with type assertions. Type assertions test availablity at compile time, and the error tests availablity at run time. Something like io.copyBuffer would change to look like this:

	if wt, ok := src.(WriterTo); ok {
		written, err := wt.WriteTo(dst)
		if err != errors.ErrUnimplemented {
			return written, err
		}
		// WriteTo existed but was not actually implemented, carry on as though it did not exist.
	}

In other words, yes, for these cases you are expected to write two tests.

[darkfeline]

I see, this is about the implementation determining whether the method can be provided at runtime time, rather than the caller determining whether the "dynamic type" of a value implements the method at runtime. I see the value for this for existing code, but new code can use an interface type with a different implementation determined at runtime.

type Foo interface {}
type Bar interface {}
func NewFoo() Foo {
  if runtimeCheck() {
    return specialFoo{} // supports .(Bar)
  } else {
    return basicFoo{}
  }
}

Does this proposal include recommendations for the wider community on whether to use ErrUnimplemented or the above for new code?

[josharian]

Can we spell it without the Err prefix, since it is in the errors package?

[tooolbox]

The use of optional methods in the io/fs proposal seems likely to lead to awkwardness with fs middleware, which must provide optional methods to support higher performance, but must then fall back to generic implementations with the underlying fs does not provide the method.

Would it suffice for io/fs to define an ErrNotImplemented sentinel? It's not clear to me that it's a positive change to broadly sanction the concept of "methods that don't actually work at runtime".

[narqo]

(A minor suggestion) I feel Unimplemented is hard to read. Can I suggest

errros.NotImplemented

[mvdan]

The use of optional methods in the io/fs proposal seems likely to lead to awkwardness with fs middleware, which must provide optional methods to support higher performance, but must then fall back to generic implementations with the underlying fs does not provide the method.

Has any research gone into whether this can be solved at compile time, not needing to add extra run time checks which Go programs would need to add? I know it's a hard problem to solve, but I assume it's worth a try before falling back to run time checks.

[rsc]

An early version of the FS interface draft had such an error, which is maybe part of what inspired this proposal.
I removed it from the FS interface draft before publishing, because in almost all cases I could think of, it was better for a method to either (1) report a definitive failure, or (2) fall back to code that does implement the behavior.

The fact that io.Copy checks for two optional methods complicates a lot and may well have been a mistake. I would rather not make general conclusions about the awkwardness of io.Copy.

Let's look instead at fs.ReadFile, which you raised as an example. Suppose I have an fs wrapper type that adds a prefix to all the underlying calls:

type Subdir struct {
    prefix string
    fsys fs.FS
}

If that type wants to make the ReadFile method on fs available, it can already do:

func (s *Subdir) ReadFile(file string) ([]byte, error) {
    file = path.Join(s.prefix, file)
    if fsys, ok := s.fsys.(fs.ReadFileFS); ok {
        return fsys.ReadFile(file)
    }
    return fs.ReadFile(fsys, file)
}

With this proposal, the code could instead do:

func (s *Subdir) ReadFile(file string) ([]byte, error) {
    file = path.Join(s.prefix, file)
    if fsys, ok := s.fsys.(fs.ReadFileFS); ok {
        return fsys.ReadFile(file)
    }
    return errors.ErrUnimplemented
}

The last line is the only one that changed.

Compared to the first version, being able to write the second version is only slightly less work for the wrapper author, but far more work for the call sites. Now every time a method like this gets called, the caller must check for ErrUnimplemented and do something else to retry the operation a different way. That is, ErrUnimplemented is a new kind of error for Go programs, a "it didn't quite fail, you have to do more work!" error.

And it's not the case that you only have to worry about this if you've tested for an optional interface right before the call. Suppose you have code that takes a value of the concrete type *Subdir as an argument. You can see from godoc etc that there's a ReadFile method, 100% guaranteed. But now every time you call ReadFile you have to check for ErrUnimplemented.

The pattern of "handle the call one way or another" seems much better for more code than the pattern of "refuse to handle the call". It preserves the property that when an error happens, it's a real failure and not something that needs retrying.

In that sense, ErrUnimplemented is a bit like EINTR. I'm wary of introducing that as a new pattern.

[rsc]

On a much more minor point, given that package errors today exports four symbols, none of which has type error, I don't believe that "this is an error" is implied by errors.. (For example, errors.Unwrap is not an error about unwrapping.)

If we are at some point to add one or more standard error values to package errors, the names should probably continue the convention of using an Err prefix. That will be avoid readers needing to memory which symbols in package errors are and are not errors.

[rsc]

For the record, although it's mostly unrelated to this discussion, io.CopyBuffer was a mistake and should not have been added. It introduced new API for a performance optimization that could have been achieved without the new API.

The goal was to reuse a buffer across multiple Copy operations, as in:

buf := make([]byte, size)
for _, op := range ops {
    io.CopyBuffer(op.dst, op.src, buf)
}

But this could instead be done using:

w := bufio.NewWriterSize(nil, size)
for _, op := range ops {
    w.Reset(op.dst)
    io.Copy(w, op.src)
    w.Flush()
}

There was no need to add CopyBuffer to get a buffer reused across Copy operations. Nothing to be done about it now, but given that the entire API was a mistake I am not too worried about the leakiness of the abstraction.

Credit to @bcmills for helping me understand this.

[rhysh]

The result of an optional interface check is always the same for a particular value. Code can branch based off of that and know that the value won't suddenly implement or un-implement the optional method. (Consider an http.Handler that uses http.Flusher several times per response.)

What are the rules for methods that return ErrUnimplemented? If a method call on a value returns it, does that method need to return it on every subsequent call? If a method call on a value doesn't return it (maybe does a successful operation, maybe returns a different error), is the method allowed to return it on a future call?

If there were a way to construct types with methods at runtime (possibly by trimming the method set of an existing type), with static answers for "does it support this optional feature", would that address the need?

[tooolbox]

Has any research gone into whether this can be solved at compile time, not needing to add extra run time checks which Go programs would need to add? I know it's a hard problem to solve, but I assume it's worth a try before falling back to run time checks.

This is where my mind goes on this subject. It seems like the conclusion has been made that this isn't possible, but I think the space is worth exploring.

If there were a way to construct types with methods at runtime (possibly by trimming the method set of an existing type), with static answers for "does it support this optional feature", would that address the need?

I suppose if you had interface A with method set X, and wanted to wrap it with B with method set Y (superset of X) you could re-wrap it with A afterwards to ensure that the final result had only the method set of X. Then at compile time you're assured to not call methods of B which are not actually supported by the underlying A.

[tv42]

Can we please name it ErrNotImplemented? We have os.ErrNotExist not os.ErrNonexistent,

What are the rules for methods that return ErrUnimplemented? If a method call on a value returns it, does that method need to return it on every subsequent call?

Example where guaranteeing that would be tough: A filesystem that delegates to other filesystems based on a prefix of the path. Any FS-level optional interface that takes path names can lead to different implementations. Forcing the multiplexer to cope with "one answer must stick", either way, could get messy.

[tv42]

@rsc

The pattern of "handle the call one way or another" seems much better for more code than the pattern of "refuse to handle the call". It preserves the property that when an error happens, it's a real failure and not something that needs retrying.

That's fine for cases where the optional interface is an optimization that can be safely ignored. That's not always true, though. Consider cp --reflink=always on a file tree larger than my free disk; if I can't copy-on-write it, I want to abort.

[jimmyfrasche]

Another pattern could be a HasX() bool method that returns true if the wrapped object actually implements method X.

If there's no HasX method but there is a method X, you assume that X is supported.

This is a bit heavy but not much more than a sentinel error check.

It can be ignored when it doesn't matter and old code will continue to function. If it matters, new code can query it and make a decision about how to proceed. This also let's multiple methods on multiple objects be queried before anything happens.

It also has the nice property that you can see in the documentation that an optimization may or may not be available.

[jimmyfrasche]

I wrote out a simple example for io.StringWriter even though it's probably overkill for something that simple.

It's a bit wordy to have an optional interface for an optional interface but seems to work fine and is easy enough to use.

https://play.golang.org/p/WCnA9tCk189

[tv42]

@jimmyfrasche

Another pattern could be a HasX() bool method that returns true if the wrapped object actually implements method X.

An FS delegating to other FS'es based on path prefix can't answer that yes-or-no without knowing the arguments to the call (the path to delegate based on).

[jimmyfrasche]

Would it work if the HasX methods for the FS took the path as an argument?

[tv42]

Maybe, but you're still left with a TOCTOU race. Consider the delegation mapping changing on the fly.

[tooolbox]

@tv42 sounds like you're arguing in favor of ErrNotImplemented, correct?

[tv42]

@tooolbox I'm trying to discover requirements and make sure people realize their consequences. So far, I haven't seen anything else cover everything. That's not quite the same as having fixed my take on a winner, new ideas welcome!

[bcmills]

See previously #39436, but that is proposed for operations that are not supported at all, not as an indication to fall back to an alternate implementation.

[rsc]

@tv42

That's fine for cases where the optional interface is an optimization that can be safely ignored. That's not always true, though. Consider cp --reflink=always on a file tree larger than my free disk; if I can't copy-on-write it, I want to abort.

You lost me a bit here.

If the method in question is defined to do exactly X, it should definitely not do almost-X instead. I think that's true in general, with or without ErrUnimplemented.

Translating to your example (I hope!), suppose there's a Copy method and an optional CopyReflink method, and CopyReflink is considered a reasonable optimized implementation of Copy, but Copy is not a valid implementation of CopyReflink.

Then I would expect that func Copy might look for a CopyReflink method, use it if possible, and otherwise fall back to Copy.
But of course a CopyReflink implementation would never settle for calling Copy instead. In this case, I think you'd write:

func Copy(x Copier, src, dst string) error {
    if x, ok := x.(CopyReflinker); ok {
        if err := x.CopyReflink(src, dst); err == nil {
            return nil
        }
    }
    return x.Copy(src, dst)
}

As written, if x.CopyReflink fails for any reason, Copy falls back to plain x.Copy. In this case, that seems fine: if it fails, presumably no state has changed so doing x.Copy is OK.

But regardless of whether I got the above right, having ErrUnimplemented available doesn't seem to help any. If x.CopyReflink returns ErrUnimplemented, then we agree that the code would fall back to x.Copy. But what if it returns a different error, like "cannot reflink across file systems"? Shouldn't that fall back to x.Copy too? And what if it returns "permission denied"? Maybe that shouldn't fall back to x.Copy, but presumably x.Copy will get the same answer.

In this case there are other errors that should be treated the same way as ErrUnimplemented, but not all. So testing for ErrUnimplemented introduces a special path that is either too special or not special enough.

[tv42]

@rsc You've successfully debated against my point wrt cp --reflink=always. I think part of the confusion comes from GNU making the various modes of cp --reflink= all be called cp. CopyReflink makes sense, and behaves as cp --reflink=always.

I think this sort of "try an optimization" codepaths should only fall back to the unoptimized case on specific errors explicitly talking about the optimization (here, e.g. EXDEV), not just any error. And in that world, ErrUnimplemented is one of those specific errors. I will happily admit this is more of a philosophical stance than something I can vigorously defend. I don't like programs "hitting the same failure multiple times". I don't like seeing e.g. multiple consecutive attempts to open the same file, with the same error, just because the code tries all possible alternatives on errors that aren't specific to the alternate codepath.

[ianlancetaylor]

Consider a file system operation like os.Link, which creates a new hard link to an existing file. Not all file systems support hard links. A middleware file system might want to provide the Link method. But it might turn out that the target file system does not provide the Link method. There is no fallback operation for the middleware file system: if hard links aren't supported, there is no substitute. What error should the middleware Link method return if the target file system does not define Link?

Of course we can define a particular error result for Link that means "hard links not supported," but it seems like a moderately general concept. (I suppose one could also make an argument for returning syscall.EMLINK, which means "file already has maximum number of hard links", but that doesn't seem entirely satisfactory.)

[bcmills]

@ianlancetaylor, the Link case is exactly why I proposed os.ErrNotSupported (#39436). (One of my motivating examples there was file locking, which is similar: either the filesystem supports locking or it doesn't, and no semantically-equivalent fallback is possible.)

I still prefer the ErrNotSupported naming over ErrUnimplemented. To me, “unimplemented” suggests that the functionality can and should exist and the author just hasn't gotten around to it, while “not supported” does not carry any implication about whether the unsupported operation could, in principle, become supported.

[dsnet]

my main use case for an error of this form is to skip tests of functionality that is not present on the current platform.

Thanks for clarifying. That's an reasonable use-case, but I doubt that's what most people are thinking of when they see this.

The difference between that and "unimplemented" is that no semantically-equivalent fallback is possible

I'm going to make the case that most people will not understand this, leading to this error being widely misused and therefore the meaning of it being increasingly muddled to the point where the documentation mismatches what is done in practice.

(If a fallback implementation is possible, than the function being called should just do that in the first place.)

I'm not a fan of this recommendation since there is often information that the caller has where they can do a better job at doing a fallback than what the function can do locally.

[jimmyfrasche]

Clearer documentation that defines when this should and should not be used would certainly be helpful should this be accepted.

[ianlancetaylor]

// ErrUnsupported indicates that a requested operation cannot be performed,
// because it is unsupported. This error indicates that there is no alternative
// way to perform the operation. For example, a call to os.Link when using a
// file system that does not support hard links.
// Functions and methods should not return this error but should instead
// return an error including appropriate context that satisfies
//     errors.Is(err, errors.ErrUnsupported)
// either by directly wrapping ErrUnsupported or by implementing an Is method.

[darkfeline]

This error indicates that there is no alternative way to perform the operation.

That sounds like it contradicts many of the given examples and the original proposal:

Those optional methods are used if available, and otherwise a generic mechanism is used.

[ianlancetaylor]

@darkfeline Yes. The meaning of the suggested error has shifted from the original proposal. I'll edit the initial comment.

[tooolbox]

It seems like a key aspect of the error is that it's for generic code which does not accurately model the underlying system or the thing it's wrapping, and so you have a method call that's actually invalid. It's incorrect that the method exists at all on the interface or struct, but there's just no better option given the desired abstraction, the language, and whatever other constraints.

At first brush it seems like this would most come up in an operating system or filesystem, which is why I am/was in favor of it being in one of those packages. But I suppose there are other areas where there are similar-but-different items, and which therefore don't perfectly fit into abstractions, and so we get invalid methods on an interface, etc.

If we are going to add this, I feel like it's worth documenting that point: if you're returning this error, it's because the abstraction has failed, and what you have in that interface or struct doesn't actually quite fulfill it, and it can't fake fulfillment with a semantically equivalent operation.

To me, that gives me a much better conceptual understanding of when I should return that error, while informing me that I should strive to avoid doing so if at all possible.

(I'm sure there's a more prosaic or nicer way to put it in the actual docs, while still getting this point across.)

[rsc]

There was some confusion over the past week about what was the likely accept, but @ianlancetaylor has edited the top comment to make that clearer. I don't see any change in consensus about accepting the text from #41198 (comment):

package errors

// ErrUnsupported indicates that a request operation cannot be performed,
// because it is unsupported.
// Functions and methods should not return this error but should instead
// return an error including appropriate context that satisfies
//     errors.Is(err, errors.ErrUnsupported)
// either by directly wrapping ErrUnsupported or by implementing an Is method.
var ErrUnsupported = New("unsupported operation")

To respond to @darkfeline's comment above #41198 (comment), it's true for every generic error "class" that different packages may have slightly different meanings, but that still doesn't make them not useful. For example os.ErrPermission, io.ErrUnexpectedEOF, and so on can all have slightly different interpretations based on what the actual operation was. It can still be worth having a general instance.

As a concrete example, http.Push's ErrNotSupported could declare itself to be Is(ErrUnsupported).

Will leave open for another week but this still seems like a likely accept.

[rogpeppe]

A couple of thoughts:

Firstly, by making this a centrally defined error value, there's an increased risk of confusion when errors are passed up unchanged. For example, if I call F which calls G which returns ErrUnsupported, then the caller of F might think that F itself is unsupported, when it's actually G. I've seen this kind of error-conflation issue in the past with shared ErrNotFound errors, and it can result in rare, hard-to-find bugs. I don't really see what's wrong with having domain-specific ErrUnsupported error values targetted to the specific interface that's returning the error. What is it about ErrUnsupported that means that it's worthy being the only centrally defined error in the errors package?

Secondly, unlike with type assertions, it's not possible to find out whether the operation is supported without attempting to actually perform the operation. For example, in this comment @bcmills suggests:

If symlinks are not supported on the current filesystem, we want to skip tests for symlink-specific behavior.

That means that you'll need to actually create a symlink (perhaps with a bogus name) to see whether symlinks are supported.
Symlink creation is a lightweight operation, so it probably doesn't matter in this case, but I could imagine other situations where there might be more cost involved. That said, I'm not keen on having additional HasX methods either, so perhaps this is fine - maybe it should just be a "thing" that when documenting a given interface, we say that calling a method with zero values will do nothing except return an error if the method isn't supported.

In general, I'm -1 on this proposal. Although ErrUnsupported is a useful concept, I don't think that it in a central place is particularly helpful, and it could be harmful in some circumstances.

[ianlancetaylor]

I do think that any method that might return this has to document it. So if F calls G which returns ErrUnsupported, then F needs to decide at that point whether it should return ErrUnsupported, meaning that F is not supported, or whether it should take some other action. For cases where it escapes up the stack, the docs are clear that it should always be wrapper so if nothing else the Error text should report which method is unavailable.

It's true that there will be cases where a HasM method is appropriate, or where there should be some special way of calling the method that just reports whether it is supported. That's OK.

[rogpeppe]

if nothing else the Error text should report which method is unavailable.

Unfortunately in this kind of situation, the program tends to think that it knows exactly what the error is and thus discards the error rather than logging it.

If the result of the method is documented, I still don't really see why there's an advantage in having this value in a central place rather than associated more closely with the API that contains the documentation.

[carlmjohnson]

Although ErrUnsupported is a useful concept, I don't think that it in a central place is particularly helpful, and it could be harmful in some circumstances.

I can see cases where a caller might notice that they get ErrUnsupported and then stop attempting a sequences of operations—say I have a bunch of files I want to link, stopping if any return ErrUnsupported—but I'm unclear what the benefits are of having a central definition for ErrUnsupported. Why not have fs.ErrUnsupported distinct from http.ErrUnsupported? Is it just an error hierarchy for the sake of ontological tidiness?

[rsc]

No change in consensus, so accepted:

package errors

// ErrUnsupported indicates that a request operation cannot be performed,
// because it is unsupported.
// Functions and methods should not return this error but should instead
// return an error including appropriate context that satisfies
//     errors.Is(err, errors.ErrUnsupported)
// either by directly wrapping ErrUnsupported or by implementing an Is method.
var ErrUnsupported = New("unsupported operation")

(And again, people writing APIs that don't find this error appropriate are in no way obligated to use or return it.)

[sagikazarmark]

Sorry, I might be a little late to the party, but have you considered using a constant instead of a variable (aka. constant errors):

package errors

type constantError string

func (e constantError) Error() string { return string(e) }

// ErrUnsupported indicates that a request operation cannot be performed,
// because it is unsupported.
// Functions and methods should not return this error but should instead
// return an error including appropriate context that satisfies
//     errors.Is(err, errors.ErrUnsupported)
// either by directly wrapping ErrUnsupported or by implementing an Is method.
const ErrUnsupported = constantError("unsupported operation")

Admittedly, I haven't read the entire issue, but I couldn't find any mention of constants.

Also, this doesn't follow the current pattern for global error variables, so might be out of scope here (although the Go 1 BC promise would not allow to change this later).

Sorry for the noise if this is too late or out of scope for this proposal.

[carlmjohnson]

They specifically don’t want anyone to return ErrUnsupported directly, so none of the advantages of a constant would apply, no?

[sagikazarmark]

Not quite sure, I see mentions of "returning" it. Not being able to change it to something that implements As or Is differently, or setting it to nil for that matter (although that would be a typed nil at that point) is enough to make it a constant in my opinion.

[tv42]

@sagikazarmark errors.New works the way it does by design, using type constantError string instead would allow bad things like err == "EOF". We probably can't get safe constant errors without constant composite literals.

[sagikazarmark]

would allow bad things like err == "EOF".

@tv42 That's not entirely true. That code won't even compile, because err is an error at that point (returned by a function like func do() error), not a string.

errors.ErrUnsupported == "unsupported operation" would be true though, but that feels less dangerous to me as I don't really see it in an expression like that.

We probably can't get safe constant errors without constant composite literals.

Could be, but this issue seemed to be a good place to start a conversation. 🙂

[tv42]

@sagikazarmark #39493 (comment)

[ianlancetaylor]

@sagikazarmark Perhaps this issue is a good place to start the conversation, but the conversation should definitely not occur on this issue. It should occur elsewhere so that this issue remains focused on the specific proposal. Thanks.

[rogpeppe]

For the record, despite the apparent consensus, I'm -1 on this proposal: I think it's unnecessary and sets a bad precedent.

[davecheney]

Sorry, the "close" and "unsubscribe" buttons in the GitHub app are very close