[erpellan]

Calling a constructor is DI. ‘new’ is the only DI framework required.

Benefits:

Unbelievably fast startup time

No magic

No annotations, XML or YAML required

No classpath scanning related security vulnerabilities

Conpletely deterministic

In no other language is ‘calling a constructor’ considered so complicated a framework is required.

Try it!

[therealdrag0]

It took me a hot moment to adjust after leaving classic spring Java land to a Scala shop that doesn’t use DI framework, but now I love it. I agree with what you said.

It does help to keep services small so the manual DI arg passing doesn’t get too spread out and boilerplate-y. But I have worked on a dozen services here and don’t miss an automatic DI framework.

[KronisLV]

> Calling a constructor is DI. ‘new’ is the only DI framework required.

This feels like the sane thing to do, but sadly for historical reasons this won't work with many (if not the most) Java frameworks out there, or even certain options in the .NET ecosystem.

When you get non-trivial frameworks that are deeply entrenched in having DI and using it internally for a bunch of stuff, you'll find yourself out of options, e.g. if you use Spring/Spring Boot.

And in many places something like Spring Boot might be considered an "industry standard" and you might get looked at funny if you suggested using another framework, even more so if they have their own configuration/plugin/utility solutions for it already built within the org.

Personally, however, I rather enjoy alternative takes on how web frameworks could look and even something like Dropwizard/Vert.X/Quarkus have nice quality of life aspects to them.

In a sense, it's nice that the industry is moving towards smaller service based architectures, where you might spin up a new service with whatever tech fits the problem that you're trying to solve, as opposed to being locked into adding code on top of some bloated monolith (though that comes with certain tradeoffs and drawbacks, in regards to complexity and maintenance).

[coverj]

I have a few questions about this approach, cards on the table I mainly deal in C# where DI (through the framework) is the default. How do you manage the initialization of all of the dependencies? Do you need some sort of "root" where everything is initialized and passed in? That was all I could come up with while keeping it testable and that doesn't sound maintainable once you build up the number of dependencies.

[rockwotj]

I'm also a fan of the manual DI approach - I strongly dislike the magic.

I think it helps if you are building test/fakes of your deps that don't really need as many sub-dependencies. Generally you in a test `@BeforeAll` construct a bunch of objects or for a program you do it in `main`.

I did work on a project while I was at Google that was really big and a decade old that instead had a class with a bunch of lazy getters for each item in the dependency graph, then tests would override or reset the getters with testing versions as needed. It was a little clunky but I preferred that approach over the projects I worked on that used dagger.

If you've read through the Dagger dev-guide [0] there is *a lot* in there, while the manual approach it's usually just `new` which is a really simple concept on it's own :) I think this is the right tradeoff because reading code is harder than writing code, so it's worth the extra setup. In practice I haven't seen it to amount to be an overwhelming amount.

[0]: https://dagger.dev/dev-guide/

[yCombLinks]

What dependencies are multiplying? I organize my code so that the very top level of the api has all external dependencies injected there, there is no public classes below a single exposed class that would need to be directly injected into. The only thing I need to actually inject are things that are outside of my memory space, ie network calls, event buses, dbs. If it's not using something outside of my memory space I am just using new inside my classes, it's far easier to test. For example, if I'm adding a new feature to allow a user to change their address, I would have a single class at the top level "CustomerAddressChange" or whatever. I would inject an interface that wraps everything external, and that is the only thing that anything would be injected into. My tests would all stub only that interface. Everything else is done without any sort of DI container.

[thiht]

When I switched from Java to Go and started doing DI like this, I was amazed at how simple and efficient DI could be. I was taught DI with Spring and never fully grasped why and how it worked. With Go it just clicked and to this day I still have no idea why DI libs are so complicated. If I ever come back to Java, I’ll make sure to do DI manually.