It's a shame to me how docker has come to be seen as "the usual software", as it's really just another layer of duck tape which uses OS features to pretend that arbitrary build processes are "reproducible" (when actually the way most Dockerfiles are written, they are most certainly not).

As a result, working with docker images often needs special privileges which leads to complex situations involving docker-in-docker etc., which just shouldn't be necessary.

Docker’s build process sucks, but that’s not the point of Docker—the point is that it’s a standard image format for containers which allows for orchestrators like Kubernetes to exist without caring about the internal details of any given container (e.g., what language it uses). It does pretty well in that regard, and you can even use Nix to build Docker images. Nix’s value proposition is reproducible builds, and while it does reproducibly build software, writing those definitions is incredibly hard (though the difficulty varies widely from package to package).

No, I absolutely acknowledge the value of a "docker image" as, essentially, an interchange format.

But it's also reached for as a solution for a whole heap of other things which are much better done in other ways.

Unfortunately today there aren’t many better options. Building images with Nix is great in theory (as Nix in general is great in theory), but it implies packaging one’s own app with Nix which is an enormous burden in most cases. There are other projects emerging that give us better image build systems than Docker, but none are well-tested at this point such that their tradeoffs are understood. This is a real pain point in the ecosystem, and it will be interesting to see how it resolves.

> packaging one’s own app with Nix which is an enormous burden in most cases

I have not found this at all tbh.

I’ve tried this personally, and we attempted it as a team professionally and in both cases the juice was deemed not worth the squeeze. In particular, we were constantly packaging low-level transitive dependencies for C projects, and ultimately you had to not only know how to build your own package, but also any package for your entire dependency tree. If your dependency tree doesn’t involve any C dependencies (or if Nix definitions already exist for them) then you probably won’t run into so many issues (which is likely to be the case for Go, Haskell, etc—languages that don’t have a heavy emphasis on C dependencies and probably those that can be statically linked) but for Python, Node, etc you will probably have problems. None of this is to speak about the dearth of documentation or the difficulty using nixpkgs as a reference.

I tend to find unmet dependencies to be an opportunity to package that dependency and upstream it. Even keeping the package private in situations where that's not appropriate feels like less work than the alternative of manual management.

And I've actually found the discoverability of the tricks used in nixpkgs to be pretty amazing. Sure the manual could always do with being more thorough (indeed it's always getting better), but again, in comparison, the quantity of stuff understandable from just reading the contents of nixpkgs and knowing basically one language is huge.

I compare it to a lot of the other systems I work with where just to decipher relatively simple setups people have created I have to understand cryptic combinations of terraform, Dockerfiles, bash, helm, github workflows, concourse pipelines, puppet/chef/ansible/..., travis.yml, Brewfiles... each of which being weirdly arbitrarily restricted and inevitably needing to call in yet another tool to fill in the gaps in their capabilities.