In Golang, how can a consumer define an interface for a function that accepts an interface?

Question

If I understand Go practices correctly, callers (aka consumers) are supposed to define interfaces of what they want to use from their dependencies (aka producers).

However, if the producer has a function that accepts a custom type, then it's better to make it accept an interface, right? This way a consumer could just pass some value that complies with producer's interface, without knowing the exact type. Because an input value into a producer's function makes the producer become the "consumer" of that input value.

Okay, fair enough.

The question is, how can consumer define an interface, which contains a function, whose parameter is an interface defined in the producer?

Trying to make the question clearer

Let's say I have a package called chef which has a struct Chef. It has a method Cut(fruit) error and fruit is an interface defined in my chef package.

Now let's say I am in the calling code, and I import package chef. I want to give it a fruit to cut, but in my case, I implemented a specific fruit called Apple. Naturally, I will try to build this interface for myself:

type myRequirements interface {
  Cut(Apple) error
}

Because I have the specific implementation of fruit interface called Apple, I want to indicate that my interface just works with apple.

However, if I try to use Chef{} against my interface, Go will throw a compile error, because my interface wants to Cut(Apple) and the Chef{} wants to Cut(Fruit). This is despite the fact that Apple implements fruit.

The only way to avoid this, it seems, is to make chef.Fruit a public interface, and use that in my own interface.

type myRequirements interface {
  Cut(chef.Fruit) error
}

But this completely ruins my ability to plug a different implementation (instead of chef) under my interface, because now I'm tightly coupled to chef.

So Chef has an internal interface fruit, but caller only knows about Apple. How can I indicate in the caller's interface what input should go into Cut without referencing chef?

Answering a comment "Why do you need myRequirements?"

I was surprised that this isn't a more agreed upon concept in the Go community.

The reason I need a myRequirements interface is because I’m a consumer of chef package. Besides Cut, chef may have 100 more methods. But I only use Cut. I want to indicate to other developers, that in my situation I’m only using Cut. I also want to allow tests to only mock Cut for my code to work. Additionally, I need to be able to plug a different implementation of Cut (from a different chef). This is a golang best practice as alluded to in the beginning of my post.

Some quotes as evidence:

Golang Wiki says: "Go interfaces generally belong in the package that uses values of the interface type, not the package that implements those values."

Dave Cheney's blog explains: "Interfaces declare the behaviour the caller requires not the behaviour the type will provide. Let callers define an interface that describes the behaviour they expect. The interface belongs to them, the consumer, not you."

Jason Moiron's tweet points out a common misunderstanding: "people have it backwards: #golang interfaces exist for the functions that use them, not to describe the types that implement them"

Update

The best advice I got so far is to move the interface into a 3rd package, independent of caller and producer. For example, make a kitchen package, define Fruit interface in it, and use it in both chefs and callers. Kind of like everyone uses time.Time. Perhaps that's the best advice. That said, I would still like to get an authoritative perspective from someone who tried to deal with this problem in their real work.

Answer 1

I would say that it comes down to what you have control over. In your example, it appears that you've described two separate packages. There are a number of ways to handle this issue:

Accept a Function

You could modify ApiFunction to accept a function that handles the cases you want:

type consumerDeps interface {
    ApiFunction(func() string) string
}

This would allow you to inject the exact functionality you desire into the consumer. However, the downside here is that this can quickly become messy and it can obfuscate the intent of the defined function and lead to unintended consequences when the interface is implemented.

Accept an interface{}

You could modify ApiFunction to accept an interface{} object that is handled by whoever implements the interface:

type consumerDeps interface {
    ApiFunction(interface{}) string
}

type producer struct{}

type apiFunctionInput interface {
    hello() string
}

func (producer) ApiFunction(i interface{}) string {
    return i.(apiFunctionInput).hello()
}

This is a little better but now you're depending on the producer-side to interpret the data correctly, and if it doesn't have all the context necessary to do that, you might wind up with unexpected behavior or panics if it casts to the wrong type.

Accept a Third-Party Interface

You could also create a third-party interface, call it Adapter here, that will define functions both the producer-side and consumer-side can agree to:

type Adapter interface {
    hello() string
}

type consumerDeps interface {
    ApiFunction(Adapter) string
}

Now, you have a data contract that can be used to send by the consumer and to receive by the producer. This may be as simple as defining a separate package, or as complex as an entire repository.

Redesign

Finally, you could redesign your codebase so the producer and consumer are not coupled together like this. Although I don't know your specific usecase, the fact that you're having this particular problem implies that your code is coupled too tightly, and should probably be redesigned. There's probably an element split between both the consumer-side and producer-side package that could be extracted to a third package.

Answer 2

After discussing a parallel case in "Problem with "define interface where it is used" principle in Golang", I believe that, in your case, the essence of the problem seems to be this: when a function/method in a producer package expects an interface as an argument, that creates a conundrum when we try to create an interface in the consumer package to abstract away the producer.

That is because we either have to reference the producer's interface in our consumer interface (violating the principle of defining interfaces where they are used), or we have to create a new interface in the consumer package that might not be fully compatible with the producer's interface (resulting in compile errors).

That is a situation where the "define interfaces where they are used" guideline (as discussed in "Effective Go / Interfaces and other types / Generality") comes into tension with the goal of reducing coupling.

In such cases, it might be useful to think about why we are trying to reduce coupling. In general, we want to reduce coupling to increase flexibility: the less our consumer package knows about the producer package, the easier it is to swap out the producer package for a different implementation.
But at the same time, our consumer package needs to know enough about the producer package to use it correctly.

See also, as illustration (in different languages, but with a similar idea) "Why coupling is always bad / Cohesion vs. coupling" from Vidar Hokstad.

---

One possible solution, as you have noted, is to define a shared interface in a third package. That reduces the coupling between the consumer and producer packages, since they both depend on the third package, but not on each other. However, this solution can feel like overkill, especially for small interfaces. Plus, it introduces a new dependency, which can increase complexity.

Given the chef and fruit package structure, one approach could involve creating a common interface in a shared package. Let's name it kitchen.

package kitchen

// Define the Fruit interface in a shared package.
type Fruit interface {
    Color() string
    Taste() string
}

Then in your chef package, you reference this Fruit interface:

package chef

import "yourproject/kitchen"

type Chef struct{}

func (c Chef) Cut(f kitchen.Fruit) error {
    // Do the cutting...
    return nil
}

Now, in your consumer package:

package consumer

import "yourproject/kitchen"

// Define an interface specific to the needs of your package.
type myRequirements interface {
  Cut(kitchen.Fruit) error
}

That approach allows the consumer package to define an interface to match its needs without directly depending on the chef package. However, it does rely on a common understanding of what a Fruit is, which is defined in the shared kitchen package.

Keep in mind, this approach introduces a dependency on the kitchen package, but if Fruit is a fundamental concept to your application that many packages will need to understand, it might be reasonable to define it in a common place.

---

In practice, the decision might come down to practical considerations. If the producer's interface is stable and unlikely to change, and if the consumer package is already closely tied to the producer package in other ways, then it might make sense to simply reference the producer's interface in the consumer package, even though this increases coupling.

On the other hand, if the producer's interface is likely to change, or if you want to preserve the flexibility to swap out the producer package for a different implementation, then defining a shared interface in a third package might be the better choice, despite the added complexity.

In that last case, you would still have the chef package, but you would also have to deal with another package, butler, which does similar tasks but might not comply with the same interface as chef.

First, you have the Fruit interface in the kitchen package like before:

package kitchen

// Define the Fruit interface in a shared package.
type Fruit interface {
    Color() string
    Taste() string
}

And in the chef package:

package chef

import "yourproject/kitchen"

type Chef struct{}

func (c Chef) Cut(f kitchen.Fruit) error {
    // Do the cutting...
    return nil
}

Now, let's imagine a new package butler which has a similar Prepare method:

package butler

import "yourproject/kitchen"

type Butler struct{}

func (b Butler) Prepare(f kitchen.Fruit) error {
    // Do some preparing...
    return nil
}

In your consumer code, you can now define two separate interfaces, one for each package's requirements:

package consumer

import "yourproject/kitchen"

// Define an interface for the chef package
type chefRequirements interface {
  Cut(kitchen.Fruit) error
}

// And another for the butler package
type butlerRequirements interface {
  Prepare(kitchen.Fruit) error
}

That means you have defined separate interfaces for the chef and butler, each tailored to the consumer's use case for each package.
By doing this, you retain the flexibility to use either the chef or butler packages (or even both) without tightly coupling your code to either one. You have moved the shared interface (Fruit) to a third package, kitchen, and you are using that in your consumer's interfaces.

That example demonstrates the flexibility of Go's interface system and how it can be used to decouple packages and support multiple implementations of a concept. But that does add some complexity, in exchange for a good deal of flexibility in return.

---

In other words, there is no hard and fast resolution to this dilemma: the best solution depends on the specific context and requirements of your project. Both approaches have trade-offs, and the right choice depends on which trade-offs are more acceptable in your situation.

It is also worth noting that this is a fairly niche problem that you are unlikely to encounter often in day-to-day Go programming. In many cases, you can structure your code in such a way that this issue does not arise. But when it does, it is a reminder that guidelines are just that—guides, not rigid rules, and it is up to us as developers to make the final judgment based on our understanding of the specific problem and the broader context in which it exists.

Answer 3

I'm not quite sure why you would introduce the myRequirements interface. If Chef requires a Fruit to Cut and you want to define a specific fruit called Apple - all you need to do is to define an Apple struct which implements the Fruit inteface.

type Chef struct {
}

type fruit interface {
    Cut() error
}

func (c Chef) Cut(fruit fruit) error {
    return fruit.Cut()
}

All you need to do is then to define Apple which implements the Fruit interface based on your requirements:

package kitchen

import chef "goplayground/interfaces/fruits/chef"

type Apple struct {
}

func (a Apple) Cut() error {
    // lets cut
    return nil
}

type myRequirements interface {
    Cut(Apple) error
}

type myChef struct {
    chef chef.Chef
}

func (m myChef) Cut(apple Apple) error {
    // since Apple implements the chef`s fruit interface this is possible
    return m.chef.Cut(apple)
}

func cook() {
    remy := myChef{}
    apple := Apple{}
    _ = remy.Cut(apple)
}

Answer 4

There are nuances to the correct use of duck-typing, which is what Go type system is when it comes to interfaces. It is usually a good practice to define the interfaces where you use them, but io.Reader interface is defined in the standard library. So there are limits to the applicability of that advice.

In your case, the package chef has two interfaces, Chef and Fruit. These two interfaces are closely coupled, because Chef has a method that uses Fruit. With the current Go type system, you cannot use Chef without exporting Fruit from that package. So:

type myRequirements interface {
  Cut(chef.Fruit) error
}

is the only way you can use the chef.Chef with an Apple implementation from your package.

But what you want to do is:

type myRequirements interface {
  Cut(Apple) error
}

and that you want to be able to convey that this is a subset of Chef, that is, the semantics of Cut is the same as the semantics of Chef. Well, the semantics are different. It is unsafe otherwise.

Say, you implemented Apple as:

type Apple struct {}

func (a Apple) SomeFunc()
func (a Apple) FruitFunc()

whereas chef.Fruit is:

type Fruit interface {
   FruitFunc()
}

Clearly, Apple implements chef.Fruit, so you can pass Apple to whereever a chef.Fruit is required. But you cannot pass chef.Fruit to myRequirements.Cut(Apple) func. Because in myRequirements.Cut you also implied that you may use Apple.SomeFunc, which is not defined in chef.Fruit.

So, if you really would like to define an interface like myRequirements, then you have to define it using chef.Fruit. If you define is using Apple, the myRequirements.Cut method is different from chef.Cut.