Cyclic dependencies and interfaces

I am a long time python developer. I was trying out Go, converting an existing python app to Go. It is modular and works really well for me.

Upon creating the same structure in Go, I seem to land in cyclic import errors, a lot more than I want to. Never had any import problems in python. I never even had to use import aliases. So I may have had some cyclic imports which were not evident in python. I actually find that strange.

Anyways, I am lost, trying to fix these in Go. I have read that interfaces can be used to avoid cyclic dependencies. But I don't understand how. I didn't find any examples on this either. Can somebody help me on this?

The current python application structure is as follows:

/main.py

/settings/routes.py      contains main routes depends on app1/routes.py, app2/routes.py etc
/settings/database.py    function like connect() which opens db session
/settings/constants.py   general constants

/apps/app1/views.py      url handler functions
/apps/app1/models.py     app specific database functions depends on settings/database.py
/apps/app1/routes.py     app specific routes

/apps/app2/views.py      url handler functions
/apps/app2/models.py     app specific database functions depends on settings/database.py
/apps/app2/routes.py     app specific routes

settings/database.py has generic functions like connect() which opens a db session. So an app in the apps package calls database.connect() and a db session is opened.

The same is the case with settings/routes.py it has functions that allow apps to add their sub-routes to the main route object.

The settings package is more about functions than data/constants. This contains code that is used by apps in the apps package, that would otherwise have to be duplicated in all the apps. So if I need to change the router class, for instance, I just have to change settings/router.py and the apps will continue to work with no modifications.

There're two high-level pieces to this: figuring out which code goes in which package, and tweaking your APIs to reduce the need for packages to take on as many dependencies.

On designing APIs that avoid the need for some imports:

• Write config functions for hooking packages up to each other at run time rather than compile time. Instead of routes importing all the packages that define routes, it can export routes.Register, which main (or code in each app) can call. In general, configuration info probably flows through main or a dedicated package; scattering it around too much can make it hard to manage.

• Pass around basic types and interface values. If you're depending on a package for just a type name, maybe you can avoid that. Maybe some code handling a []Page can get instead use a []string of filenames or a []int of IDs or some more general interface (sql.Rows) instead. <!-- There's something to extract about layering here. -->

• Consider having 'schema' packages with just pure data types and interfaces, so User is separate from code that might load users from the database. It doesn't have to depend on much (maybe on anything), so you can include it from anywhere. Ben Johnson gave a lightning talk at GopherCon 2016 suggesting that and organizing packages by dependencies.

On organizing code into packages:

• As a rule, split a package up when each piece could be useful on its own. If two pieces of functionality are really intimately related, you don't have to split them into packages at all; you can organize with multiple files or types instead. Big packages can be OK; Go's net/http is one, for instance.

• Break up grab-bag packages (utils, tools) by topic or dependency. Otherwise you can end up importing a huge utils package (and taking on all its dependencies) for one or two pieces of functionality (that wouldn't have so many dependencies if separated out).

• Consider pushing reusable code 'down' into lower-level packages untangled from your particular use case. If you have a package page containing both logic for your content management system and all-purpose HTML-manipulation code, consider moving the HTML stuff "down" to a package html so you can use it without importing unrelated content management stuff.

Here, I'd rearrange things so the router doesn't need to include the routes: instead, each app package calls a router.Register() method. This is what the Gorilla web toolkit's mux package does. Your routes, database, and constants packages sound like low-level pieces that should be imported by your app code and not import it.

Generally, try to build your app in layers. Your higher-layer, use-case-specific app code should import lower-layer, more fundamental tools, and never the other way around. Here are some more thoughts:

• Packages are good for separating independently usable bits of functionality from the caller's perspective. For your internal code organization, you can easily shuffle code between source files in the package. The initial namespace for symbols you define in x/foo.go or x/bar.go is just package x, and it's not that hard to split/join files as needed, especially with the help of a utility like goimports.

The standard library's net/http is about 7k lines (counting comments/blanks but not tests). Internally, it's split into many smaller files and types. But it's one package, I think 'cause there was no reason users would want, say, just cookie handling on its own. On the other hand, net and net/url are separate because they have uses outside HTTP.

It's great if you can push "down" utilities into libraries that are independent and feel like their own polished products, or cleanly layer your application itself (e.g., UI sits atop an API sits atop some core libraries and data models). Likewise "horizontal" separation may help you hold the app in your head (e.g., the UI layer breaks up into user account management, the application core, and administrative tools, or something finer-grained than that). But, the core point is, you're free to split or not as works for you.

• Set up APIs to configure behavior at run-time so you don't have to import it at compile time. So, for example, your URL router can expose a Register method instead of importing appA, appB, etc. and reading a var Routes from each. You could make a myapp/routes package that imports router and all your views and calls router.Register. The fundamental idea is that the router is all-purpose code that needn't import your application's views.

Some ways to put together config APIs:

• Pass app behavior via interfaces or funcs: http can be passed custom implementations of Handler (of course) but also CookieJar or File. text/template and html/template can accept functions to be accessible from templates (in a FuncMap).

• Export shortcut functions from your package if appropriate: In http, callers can either make and separately configure some http.Server objects, or call http.ListenAndServe(...) that uses a global Server. That gives you a nice design--everything's in an object and callers can create multiple Servers in a process and such--but it also offers a lazy way to configure in the simple single-server case.

• If you have to, just duct-tape it: You don't have to limit yourself to super-elegant config systems if you can't fit one to your app: maybe for some stuff a package "myapp/conf" with a global var Conf map[string]interface{} is useful.
  But be aware of downsides to global conf. If you want to write reusable libraries, they can't import myapp/conf; they need to accept all the info they need in constructors, etc. Globals also risk hard-wiring in an assumption something will always have a single value app-wide when it eventually won't; maybe today you have a single database config or HTTP server config or such, but someday you don't.

Some more specific ways to move code or change definitions to reduce dependency issues:

• Separate fundamental tasks from app-dependent ones. One app I work on in another language has a "utils" module mixing general tasks (e.g., formatting datetimes or working with HTML) with app-specific stuff (that depends on the user schema, etc.). But the users package imports the utils, creating a cycle. If I were porting to Go, I'd move the user-dependent utils "up" out of the utils module, maybe to live with the user code or even above it.

• Consider breaking up grab-bag packages. Slightly enlarging on the last point: if two pieces of functionality are independent (that is, things still work if you move some code to another package) and unrelated from the user's perspective, they're candidates to be separated into two packages. Sometimes the bundling is harmless, but other times it leads to extra dependencies, or a less generic package name would just make clearer code. So my utils above might be broken up by topic or dependency (e.g., strutil, dbutil, etc.). If you wind up with lots of packages this way, we've got goimports to help manage them.

• Replace import-requiring object types in APIs with basic types and interfaces. Say two entities in your app have a many-to-many relationship like Users and Groups. If they live in different packages (a big 'if'), you can't have both u.Groups() returning a []group.Group and g.Users() returning []user.User because that requires the packages to import each other.

  However, you could change one or both of those return, say, a []uint of IDs or a sql.Rows or some other interface you can get to without importing a specific object type. Depending on your use case, types like User and Group might be so intimately related that it's better just to put them in one package, but if you decide they should be distinct, this is a way.

<!-- note for later: can think of at least one time in my short experience that it helped to push a "dumb" data type down into a lower-level package separate from much of the functionality using it. that package didn't need to import much (anything in my case), so other packages could import it and use the datatype it defined to communicate. io and errors are arguably examples, with the import-heavier functionality moved off to io/ioutuil, os, etc. however, i'd like to research up more cases and get a sense of how important/preferred this technique is in prod software out there. -->

Thanks for the detailed question and followup.

Possible partial, but ugly answer:
Have struggled with the import cyclic dependency problem for a year. For a while, was able to decouple enough so that there wasn't an import cycle. My application uses plugins heavily. At the same time, it uses encode/decode libraries (json and gob). For these, I have custom marshall and unmarshall methods, and equivalent for json.
For these to work, the full type name including the package name must be identical on data structures that are passed to the codecs. The creation of the codecs must be in a package. This package is called from both other packages as well as from plugins.
Everything works as long as the codec package doesn't need to call out to any package calling it, or use the methods or interfaces to the methods. In order to be able to use the types from the package in the plugins, the plugins have to be compiled with the package. Since I don't want to have to include the main program in the builds for the plugins, which would break the point of the plugins, only the codec package is included in both the plugins and the main program. Everything works up until I need to call from the codec package in to the main program, after the main program has called in to the codec package. This will cause an import cycle. To get rid of this, I can put the codec in the main program instead of its own package. But, because the specific datatypes being used in the marshalling/unmarshalling methods must be the same in the main program and the plugins, I would need to compile with the main program package for each of the plugins. Further, because I need to the main program to call out to the plugins I need the interface types for the plugins in the main program. Having never found a way to get this to work, I did think of a possible solution:
First, separate the codec in to a plugin, instead of just a package
Then, load it as the first plugin from the main program.
Create a registration function to exchange interfaces with underlying methods.
All encoders and decoders are created by calls in to this plugin.
The plugin calls back to the main program through the registered interface.
The main program and all the plugins use the same interface type package for this.
However, the datatypes for the actual encoded data are referenced in the main program
with a different name, but same underlying type than in the plugins, otherwise the same import cycle exists. to do this part requires doing an unsafe cast. Wrote
a little function that does a forced cast so that the syntax is clean:
(<cast pointer type*>Cast(<pointer to structure, or interface to pointer to structure>).

The only other issue for the codecs is to make sure that when the data is sent to the encoder, it is cast so that the marshall/unmarshall methods recognize the datatype names. To make that easier, can import both the main program types from one package, and the plugin types from another package since they don't reference each other.

Very complex workaround, but don't see how else to make this work.
Have not tried this yet. May still end up with an import cycle when everything is done.

[more on this]

To avoid the import cycle problem, I use an unsafe type approach using pointers. First, here is a package with a little function Cast() to do the unsafe typecasting, to make the code easier to read:

package ForcedCast

import (
    "unsafe"
    "reflect"
)

// cast function to do casts with to hide the ugly syntax
// used as the following:
// <var> = (cast type)(cast(input var))
func Cast(i interface{})(unsafe.Pointer) {
    return (unsafe.Pointer(reflect.ValueOf(i).Pointer()))
}

Next I use the "interface{}" as the equivalent of a void pointer:

package firstpackage
type realstruct struct {
     ...
}   

var Data realstruct

// setup a function to call in to a loaded plugin
var calledfuncptr func(interface)

func callingfunc() {

        pluginpath := path.Join(<pathname>, "calledfuncplugin")
        plug, err := plugin.Open(pluginpath)

        rFunc, err := plug.Lookup("calledfunc")
        calledfuncptr = rFunc.(interface{})

        calledfuncptr (&Data)
}


//in a plugin
//plugins don't use packages for the main code, are build with -buildmode=plugin
package main

// identical definition of structure
type realstruct struct {
     ...
}   

var localdataptr *realstruct

func calledfunc(needcast interface{}) {

    localdataptr = (*realstruct)(Cast(needcast))

}

For cross type dependencies to any other packages, use the "interface{}" as a void pointer and cast appropriately as needed.

This only works if the underlying type that is pointed to by the interface{} is identical wherever it is cast. To make this easier, I put the types in a separate file. In the calling package, they start with the package name. I then make a copy of the type file, change the package to "package main", and put it in the plugin directory so that the types are built, but not the package name.

There is probably a way to do this for the actual data values, not just pointers, but I haven't gotten that to work right.

One of the things I have done is to cast to an interface instead of a datatype pointer. This allows you to send interfaces to packages using the plugin approach, where there is an import cycle. The interface has a pointer to the datatype, and then you can use it for calling the methods on the datatype from the caller from the package that called in to the plugin.

The reason why this works is that the datatypes are not visible outside of the plugin. That is, if I load to plugins, which are both package main, and the types are defined in the package main for both, but are different types with the same names, the types do not conflict.

However, if I put a common package in to both plugins, that package must be identical and have the exact full pathname for where it was compiled from. To accommodate this, I use a docker container to do my builds so that I can force the pathnames to always be correct for any common containers across my plugins.

I did say this was ugly, but it does work. If there is an import cycle because a type in one package uses a type in another package that then tries to use a type from the first package, the approach is to do a plugin that erases both types with interface{}. You can then make method and function calls back and forth doing the casting on the receiving side as needed.

In summary:
Use interface{} to make void pointers (that is, untyped).
Use the Cast() to force them to a pointer type that matches the underlying pointer. Use the plugin type localization so that types in the package main in separate plugins, and in the main program do not conflict If you use a common package between plugins, the path must be identical for all built plugins and the main program. Use the plug package to load the plugins, and exchange function pointers

For one of my issues I'm actually calling from a package in the main program out to a plugin, just to be able to call back to another package in the main program, avoiding the import cycle between the two packages. I ran in to this problem using the json and gob packages with custom marshaller methods. I use the types that are custom marshalled both in my main program, and in other plugins, while at the same time, I want the plugins to be built independent of the main program. I accomplish this by using a package for json and gob encode/decode custom methods that is included both in the main program and the plugins. However, I needed to be able to call back to the main program from the encoder methods, which gave me the import cycle type conflict. The above solution with another plugin specifically to solve the import cycle works. It does create an extra function call, but I have yet to see any other solution to this.

Hope this helps with this issue.

A shorter answer to your question (using interface), that does not take away the correctness and completeness of the other answers, is this example:

UserService is causing cyclic import, where it should not really be called from AuthorizationService. It's just there to be able to extract the user details, so we can declare only the desired functionality in a separated receiver-side interface UserProvider:

https://github.com/tzvatot/cyclic-import-solving-exaple/commit/bc60d7cfcbd4c3b6540bdb4117ab95c3f2987389

Basically, extracting an interface that contains only the required functionality on the receiver side, and use it instead of declaring a dependency on something external.