goose: dependency injection proof of concept

See documentation and demo for usage and known limitations.

Reviewed-by: Herbie Ong <herbie@google.com>
This commit is contained in:
Ross Light
2018-03-26 07:39:00 -07:00
commit 26c8618466
4 changed files with 1376 additions and 0 deletions

203
README.md Normal file
View File

@@ -0,0 +1,203 @@
# goose: Compile-Time Dependency Injection for Go
goose is a compile-time [dependency injection][] framework for Go, inspired by
[Dagger][]. It works by using Go code to specify dependencies, then
generating code to create those structures, mimicking the code that a user
might have hand-written.
[dependency injection]: https://en.wikipedia.org/wiki/Dependency_injection
[Dagger]: https://google.github.io/dagger/
## Usage Guide
### Defining Providers
The primary mechanism in goose is the **provider**: a function that can
produce a value, annotated with the special `goose:provide` directive. These
functions are ordinary Go code and live in packages.
```go
package module
type Foo int
// goose:provide
// ProvideFoo returns a Foo.
func ProvideFoo() Foo {
return 42
}
```
Providers are always part of a **module**: if there is no module name specified
on the `//goose:provide` line, then `Module` is used.
Providers can specify dependencies with parameters:
```go
package module
// goose:provide SuperModule
type Bar int
// ProvideBar returns a Bar: a negative Foo.
func ProvideBar(foo Foo) Bar {
return Bar(-foo)
}
```
Providers can also return errors:
```go
package module
import (
"context"
"errors"
)
type Baz int
// goose:provide SuperModule
// ProvideBaz returns a value if Bar is not zero.
func ProvideBaz(ctx context.Context, bar Bar) (Baz, error) {
if bar == 0 {
return 0, errors.New("cannot provide baz when bar is zero")
}
return Baz(bar), nil
}
```
Modules can import other modules. To import `Module` in `SuperModule`:
```go
// goose:import SuperModule Module
```
### Injectors
An application can use these providers by declaring an **injector**: a
generated function that calls providers in dependency order.
An injector is declared by writing a function declaration without a body in a
file guarded by a `gooseinject` build tag. Let's say that the above providers
were defined in a package called `example.com/module`. The following would
declare an injector to obtain a `Baz`:
```go
//+build gooseinject
package main
import (
"context"
"example.com/module"
)
// goose:use module.SuperModule
func initializeApp(ctx context.Context) (module.Baz, error)
```
Like providers, injectors can be parameterized on inputs (which then get sent to
providers) and can return errors. The `goose:use` directive specifies the
modules to use in the injection. Both `goose:use` and `goose:import` use the
same syntax for referencing modules: an optional import qualifier (either a
package name or a quoted import path) with a dot, followed by the module name.
For example: `SamePackageModule`, `foo.Bar`, or `"example.com/foo".Bar`.
You can generate the injector using goose:
```
goose
```
Or you can add the line `//go:generate goose` to another file in your package to
use [`go generate`]:
```
go generate
```
(Adding the line to the injection declaration file will be silently ignored by
`go generate`.)
goose will produce an implementation of the injector that looks something like
this:
```go
// Code generated by goose. DO NOT EDIT.
//+build !gooseinject
package main
import (
"example.com/module"
)
func initializeApp(ctx context.Context) (module.Baz, error) {
foo := module.ProvideFoo()
bar := module.ProvideBar(foo)
baz, err := module.ProvideBaz(ctx, bar)
if err != nil {
return 0, err
}
return baz, nil
}
```
As you can see, the output is very close to what a developer would write
themselves. Further, there is no dependency on goose at runtime: all of the
written code is just normal Go code, and can be used without goose.
[`go generate`]: https://blog.golang.org/generate
## Best Practices
goose is still not mature yet, but guidance that applies to Dagger generally
applies to goose as well. In particular, when thinking about how to group a
package of providers, follow the same [guidance](https://google.github.io/dagger/testing.html#organize-modules-for-testability) as Dagger:
> Some [...] bindings will have reasonable alternatives, especially for
> testing, and others will not. For example, there are likely to be
> alternative bindings for a type like `AuthManager`: one for testing, others
> for different authentication/authorization protocols.
>
> But on the other hand, if the `AuthManager` interface has a method that
> returns the currently logged-in user, you might want to [export a provider of
> `User` that simply calls `CurrentUser()`] on the `AuthManager`. That
> published binding is unlikely to ever need an alternative.
>
> Once youve classified your bindings into [...] bindings with reasonable
> alternatives [and] bindings without reasonable alternatives, consider
> arranging them into packages like this:
>
> - One [package] for each [...] binding with a reasonable alternative. (If
> you are also writing the alternatives, each one gets its own [package].) That
> [package] contains exactly one provider.
> - All [...] bindings with no reasonable alternatives go into [packages]
> organized along functional lines.
> - The [packages] should each include the no-reasonable-alternative [packages] that
> require the [...] bindings each provides.
One goose-specific practice though: create one-off types where in Java you
would use a binding annotation.
## Future Work
- The names of imports and provider results in the generated code are not
actually as nice as shown above. I'd like to make them nicer in the
common cases while ensuring uniqueness.
- I'd like to support optional and multiple bindings.
- At the moment, the entire transitive closure of all dependencies are read
for providers. It might be better to have provider imports be opt-in, but
that seems like too many levels of magic.
- Currently, all dependency satisfaction is done using identity. I'd like to
use a limited form of assignability for interface types, but I'm unsure
how well this implicit satisfaction will work in practice.
- Errors emitted by goose are not very good, but it has all the information
it needs to emit better ones.