Run Nix Based Environments in Kubernetes
flox.dev
flox.dev
We did just launch this last week after a good bit of work from the team. Steve wrote up a deeper technical dive here if anyone is interested - https://flox.dev/blog/kubernetes-uncontained-explained-unloc...
The solution is to decompose the docker images and make sure that every layer is hash equivalent. So if people update their Cuda version, it result in a change within the Python layers.
But it looks like Flox now simplifies this via Nix. Every Nix package already has a hash and you can combine packages however you would like.
After spending a few years using nix, the docker image situation looks pretty bonkers. If two files end up in separate layers, the system assumes dependency so if the lower file changes you need to build a separate copy of the higher one just in case there's actual dependency there.
Within nix you can be more precise about what depends on what, which is nice, but you do have to be thoughtful about it or you can summon the same footgun that got you with docker, just in smaller form. Because a nix derivation, while a box with nicely labeled inputs and output, is still a black box. If you insert a readme as an input to a derivation that does a build, nix will assume that the compiled binary depends on it and when you fix a typo in the readme and rebuild you'll end up with a duplicate binary build in the nix store despite the contents of the binary not actually depending on the text of the readme.
> you can combine packages however you would like
So this is true, more or less, but be aware that while nix lets you do this in ways that don't force needless duplication, it doesn't force you to avoid that duplication. Things carelessly packaged with nix can easily recreate the problem you mentioned with docker.
One my issues with Nix is the black box that is the store, and maybe it's just my system, but over time I find it full of redundant files / orphans and no obvious way to flatten it or clean it safely without breaking something.
I wonder how flox solves this.
On store bloat: Flox makes it clearer what's in use (explicit environments vs. implicit dependencies), but you still need nix-collect-garbage.
The store accumulates cruft, that's Nix reality, we haven't changed it.
We're also adding "stabilities" (downsample from daily to weekly/monthly snapshots) to reduce churn even more. Still need GC, but a lot fewer bags on trash day. .
When I have space problems, I run that and they're gone, then later I do it again. It could be that I'm just avoiding functionality that it breaks though.
Instead of running a tight ship I spend a lot of time dreaming about alternate computational universes, and one I particularly like is where new hard disks come pre-loaded with a fragment of bits deemed culturally valuable. Wikipedia and a well curated subset of nixpkgs would be a fine start to such an archive. In this world your files don't grow/shrink to consume/yield empty space, but rather the boundary between your data and that drive's shard of the public archive shifts in one direction or another. This way you or somebody in your neighborhood is likely to already have the file you need, so you can get it from them instead of the internet. Better for being able to roll with the punches if the internet is partitioned.
I don't worry about the size of the nix store because according to this weird fantasy of mine it's on the side of my disk that shrinks when I add files to it; it's not the contained object, but the gas that expands to fill the rest of the container. Not by accident, but as part of a redundant worldwide distributed cache that we put together after deciding that servers controlled by people we don't know are not to be relied on.
The wiki rightfully points towards "roots", i.e. references produced by nix-build or similar. Additionally, there are other places that will keep references and hence block garbage collection though:
1. Your nix profile (`nix profile list` / `nix profile remove`) and its old generations (`nix profile (wipe-)history`)
2. Your NixOS configuration (configuration.nix) and its old generations (`nixos-rebuild list-generations`)
It doesn't help that there's no discoverable way to tell why a particular nix store path is not being garbage collected either.
With the Kubernetes shim, you can run the hash-pinned environments without building or pulling an image at all. It starts the pod with a stub, then activates the exact runtime from a node-local store.
Sure, it's easy to stand up a mail server in NixOS, or to just use docker/kubernetes to deploy stuff. But after a few years it felt like I don't have a single understanding of the stack. When shit hits the fan, it makes it very difficult to troubleshoot.
I am now back on running my servers on FreeBSD/OpenBSD and jails or VMM respectively. And also dumbing the stack down to just "run it in a jail, but set it up manually".
The only outlier is Immich. For some reason they only officially support the docker images but not a single clear instruction on how to set it up manually. Sure, I could look at the Dockerfiles, but many of the scripts also expect docker to be present.
And now that FreeBSD also has reproducible builds, it took one more stone away from Nix.
Also, think it's a huge ecosystem win for FreeBSD pushing on reproducibility too. I think we are trending in a direction where this just becomes a critical principle for certain stacks. (also needed when you dive into AI stacks/infra...)
I like it when my system comes with a complete set of manpages and good docs.
But you mentioned Flox, which I didn't even know about. First I thought that's what they renamed the Nix fork to after the schism, but now I see it's a paid product and yuck...just further deepens my believe in going more bare bones manual control, even if sometimes bothersome.
We have six dev teams and are just about done with migrating to k8s. It's an immense improvement over what we had before.
It's a version of Greenspun's tenth rule: "Any sufficiently complicated distributed system contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of Kubernetes."
at orgs significantly larger than that, the kube team has to aggressively spin out platform functions that enable further layering or risk getting overwhelmed trying to support and configure kube features to cover diverse team needs (that is, storage software doesn’t have the same needs or concerns as middleware or the frontend). this incubator model isn’t easy in practice. trying to adopt kube at this scale is very challenging because it requires the kube team to spin up and out sub-teams at a very high rate or risk slowing the migration down to a crawl or outright failure and purchasing e.g. off the shelf AWS because teams need to offboard their previous platform.
I don't doubt it. By "larger" I just meant larger than something like "running my servers on FreeBSD/OpenBSD and jails or VMM respectively" above, which sounds like a one-person operation.
> I wouldn’t be surprised if that’s close to the perfect size to move onto kube and create and adopt a standard set of platform idioms.
My previous position was at a company about 5x the size, with many loosely related enterprise and government products that they sold into markets in at least 20 countries. They also used k8s quite effectively.
But, I think the key is that you mentioned "the kube team". Having a single team responsible for everything k8s-related at a large org is likely to make it difficult to be effective.
For supporting individual dev teams I think you need people on those teams who have at least some of the necessary knowledge, so they're not entirely dependent on a central team. Even a watered-down version of what devops was supposed to be about is better than nothing.
Besides, this particular comment would need to explain why it’s likely that its unsupported opinion is correct, rather than all the counterexamples that exist in the actual highly competitive industry that’s heavily using this system.
There’s a fundamental conflict there that doesn’t work in favor of your inchoate opinion. The most likely conclusion is that there are factors at work here that you don’t understand.
https://github.com/pdtpartners/nix-snapshotter
so kind of allowing pull images from nix store, mounting shared host nix store per node into each container, incremental fast rebuilds, generating basic pod configs are good things.
and local, ci and remote runs same flows and envs.
I'm curious! and ignorant! help!
Is that via (centrally?) cached eval? or what? there's only so much room for magic in this arena.
When you create a Flox environment, we evaluate the Nix expressions once and store the concrete result (ie exact store paths) in FloxHub. The k8s node just fetches that pre-rendered manifest and bind-mounts the packages with no evaluation step at pod startup.
It's like the difference between giving the node a recipe to interpret vs. giving it a shopping list of exact items. Faster, safer, and the node doesn't need to know how to cook (evaluate Nix). I don't know, there's a metaphor here somewhere, I'll find it.
Only so much room for magic, for sure, but tons of room for efficiency and optimization.
Our catalog continuously pre-evaluates nixpkgs in the background. 'flox install' just selects from pre-evaluated packages -- no eval needed, ever. The k8s node fetches the manifest and mounts the packages.
Eval is done once, centrally, continuously. So... even more pre-val'd, so to speak.
The license file in their github seems to indicate that it is. https://github.com/flox/flox?tab=GPL-2.0-1-ov-file
It's hard for me to understand if I should be excited about this. I think companies do themselves such huge disservices from not being transparent to the nerds that WILL be the ones helping choose/implement these things. Instead of the current feeling I have, there could be three sentences that explains what Flox is offering here beyond what *anyone* can go do right now with nix-snapshotter.
If it's ecosystem stuff (you get Flox's CI, or CLI, or whatever else), that's not very well sold to me on the landing page. Otherwise I'm feeling left empty-handed.
Not nix-snapshotter because we skip Nix eval entirely and get way better cache sharing across unrelated workloads (quantized catalog means everything shares base deps). On "environments": these aren't devshells-as-prod, they're the actual runtime; same as 'flox activate' works everywhere. You're shipping a declarative, hash-pinned runtime that happens to also work great in dev/CI
And yeah, we should have been upfront that this is alpha and we're planning to open source it after vetting at KubeCon.
You're right that we're doing ourselves a disservice not being transparent with the technical crowd. What specific technical details would help you evaluate this?
> Kubernetes, Uncontained
> Are you replacing containers?
> No. Kubernetes still runs containers.
Re: Relationship to nix-snapshotter and prior art This is original work, though very much built on prior innovations. Our approach hooks into the upstream containerd runc shim to pull the FloxHub-managed environment and bind-mount the closure at startup. The key distinction is that we use how Flox environments are rendered to avoid Nix evaluation entirely, making it safe and fast for a k8s node to realize packages directly on the node. Less about images and containers, per se, and more out bringing the power of Flox and Nix at the buildtime end to the runtime end of SDLC.
The cache story is surprisingly strong: nix store paths effectively behave like layers in the node’s registry, but with dramatically higher hit rates -- often across entirely unrelated pod deployments. Because all pods rely on the same underlying system libraries drawn from the “quantized” Flox catalog, different environments naturally share glibc, core utilities, and common dependencies, where traditional containers typically share nothing.
Tools like nix-snapshotter, Nixery, and others have pioneered this space and we're grateful for that work. This rising "post-Docker" tide raises all ships.
Re: Open Source The software is brand new -- only slightly older than Ron’s baby -- and currently in alpha. KubeCon was our first opportunity for broad feedback, and we uncovered a few issues we’re still addressing. Our intent is to open-source the project once we’ve fully vetted the approach, ideally in the coming weeks.
Yes, we launched early and the product is imperfect, but we’re doing so transparently and with a commitment to getting it right and releasing it to the community, we will continue to release early and often.
Re: Abstraction depth concerns I appreciate @rootnod3’s point about deeper abstractions complicating debugging. We’re thinking hard about how to keep things simple for people who need to run and fix systems quickly. It’s encouraging to see the broader ecosystem—like FreeBSD—lean further into reproducibility, especially as AI-centric stacks make this increasingly important.
Re: Nix vs traditional approaches Skilled Dockerfile authors can achieve great caching results -- and you can pin and you can prune registries, etc -- but our goal is to make these best practices the default. Nix enables finer-grained caching and a universal packaging format for building and consuming open source software.
We see intrinsic value in Flox environments -- whether on the CLI, k8s, Nomad down the road, or other platforms. Our aim is for Flox environments to be as universal and natural as Nix packages themselves -- essentially extending “flox activate” into the k8s world.
We likewise got a ton of valuable feedback at KubeCon, most of which was validating, all of which was very inline with this conversation.
A dilemma I’m facing is that the win from nix in terms of faster builds and smaller images would be largely from python and R images (where the average size is often 1Gi or larger). However, the developers that use Python or R are less likely to “get” the point of Nix and might have a steeper learning curve than F# developers (where the builds are quite efficient).
That was the context, my question is, how’s the integration with Flox and R/RStudio? I know there’s Rix[1] for managing R packages with Nix.
Driver version is host-managed (stable), toolkit is hash-pinned (stable). No drift on either side.
Initial selection matters (pick compatible versions, nothing automatic will stop you from initially installing the wrong thing), but once pinned, and unless/until you change it, you get the same toolkit forever regardless of catalog updates or time passing.