Dependency injection on Android with dagger-android and Kotlin | Hacker News Reader

Dependency injection on Android with dagger-android and Kotlin

albertgao.xyz

32 pointsalbertgao17 comments

Dagger2 is great when you have it setup, generating the dependency graph during compilation makes it lightning fast during execution. I also like the error messages, as long as you can read Java stack traces all the info for solving errors are in there - I've never seen an unsolvable situation this far.

Kapt does bump the clean build time to about 2min, incremental builds take between 10-60s - although actually launching the app on an emulator/device adds another 20s.

The cons of Dagger2 that I've experienced since it's launch are; The documentation and support is useless. You're on your own of you don't use a 3rd party sample project as a template. No one understands scopes and subscopes, subcomponents etc. The new Android dagger api is arcane and weird, no one wants to use it.

The Dagger2 team should (if they aren't already) create a Kotlin extension for it, I believe there are some syntactic optimizations to be offered.

I wrote a semantics doc to try to help on the documentation front:

https://google.github.io/dagger/semantics/

It's not intended as a tutorial, so perhaps it won't quite meet your needs, but if you're curious how all the pieces fit together, it may help.

Interesting, will check this and see what i could add to my blog. But yeah, if there is an official example of how to use dagger-android, it will be much helpful. Took me a day to figure it out the whole things and 2 different types of dagger...

> No one understands scopes and subscopes, subcomponents etc

I think there are two reasons for this. First, is how scopes interact with Android component lifecycles, which make everything harder (like RxJava, etc) since they are a complexity multiplier.

Second, there are seemingly bizarre design decisions. For example, if you use dependent components (aka component dependencies), the syntax is pretty straightforward:

  @Component { }
  interface AppComponent { X,Y,Z }

  @Component { dependencies = AppComponent.class }
  interface LoginComponent { }

So the natural order of things is, components use modules, and components can depend on other components. Pretty simple.

Subcomponents turns this relatively simple schema around. The documentation for subcomponents say:

> "To add a subcomponent to a parent component, add the subcomponent class to the subcomponents attribute of a @Module that the parent component installs."

I find is truly strange, now we have modules depending on (sub)components which is the inverse relationship as above.

You get it totally right, I just knew that there are 2 ways to implment DI with dagger, dagger 2 or dagger-android.... And the official docs are indeed useless... And before you learn some dagger basic, some online tutorial just can't be understandable... because either they mixed up the 2 plugin or they just all have their different setup.

Unfortunately, using Dagger for DI with Kotlin requires using the Kotlin annotation processor, kapt, which is still incredibly slow, and works non-incrementally.

If you value your build times or your sanity in the slightest, I'd recommend either severely limiting use of Dagger, e.g. by only using it in certain, small Gradle modules, or using a different DI solution.

I have only used Kotlin in the context of Android so far, and like the language — I wish Java had many of these features — but almost find it more of a pain in the ass than its worth for reasons like this (kapt). At some point, the build times get so bad I asked myself, am I being more productive having neat language features X,Y,Z or having 15sec build times (already too long) vs 2 minute build times (unbearable). Its like going from REPL driven development to something else.

Thanks for this tip. Don't know this.

I tried to learn Dagger for a recent project in a similar setting (Dagger+Android+Kotlin). Spent some 8 hours, couldn't make it work and gave up. I suppose, as with most systems, one needs to build up a list of recipes for solving common problems. But with Dagger I found that notoriously hard for some reason. Perhaps because it tends to fail silently at runtime or with non-descriptive errors.

It really helps to adopt a pattern from another project rather than trying to build one from scratch if you're new to Dagger. But it's definitely worth it, I wouldn't do an Android project nowadays without dependency injections.

Yeah, I tried that, but the simple examples didn't do enough for my purposes, I had to take it further and was never able to get too far.

No need to use kapt to kill compilation times. Data Binding is good enough. We started using it a year ago and since then instant run does not work and the debugger takes three minutes to start. Sometimes it mixes generated classes and we have to do a full clean / rebuild.

Since then we have introduced Kotlin and Dagger and we are super happy with it.

As with Dagger, the learning curve is pretty steep. Having it working with modules + activities + fragments + architecture components has been a real challenge but we love it.

For those who are struggling with Dagger: version 2.10 introduced major changes for Android, for good reasons, but it made much harder to use and understand. Many tutorials focus on Dagger < 2.10. You probably should ignore those.

And for those who are going crazy with compilation times, Jrebel for Android is a huge time saver. It has an incremental compiler which works with annotation processors. They will stop supporting it in one year, but even then I think it is worth using it. Who knows: if enough people buy licenses maybe they reconsider killing it.

I've never quite understood the appeal of Dependency Injection. What's wrong with just using constructors?

Constructors in the middle of your code make it harder to substitute mock objects for testing purposes.

For simple scenarios DI is obviously an overkill, but as soon as you're dealing with nested dependencies, you'd end up with either:

* monstrously big constructors (for carrying transitive dependencies)

* lots of @VisibleForTesting code to handle manually injecting various dependencies only for the sake of testing (poor man's DI and generally bad practice)

* a lot of factories (service locator or poor man's DI, essentially)

* code that's hard to unit test due to dependencies being hardcoded.

In other words, you'll either reinvent DI poorly, or give up on testability.

The problem is not the testing. The problem is that if you have classes that depend on other classes that depend on other classes that depend on specific configuration.

Then you have a mess without DI.

You should look at kodein, a DI library for kotlin. Works brilliantly.

Reply on news.ycombinator.com