Most of the hyper scaler actually do not store container images as tarballs at scale. They usually flatten the layers and either cache the entire file system merkle tree, or breaking it down to even smaller blocks to cache them efficiently. See Alibaba Firefly Nydus, AWS Firecracker, etc… There is also various different forms of snapshotters that can lazily materialize the layers like estargz, soci, nix, etc… but none of them are widely adopted.
I'm the author of one of those off the shelf tools, and the rules_oci decision here always struck me as a bit unusual. OCI is a relatively easy spec with a number of libraries that implement it. Instead of creating a custom build command that leveraged those libraries to be an efficient build tool, they found commands that could be leveraged even if image building wasn't their design.
It looks like rules_img is taking that other path with their own build command based on the go-containerregistry library. I wish them all the best with their effort.
That said, if all you need to do is add a layer to an existing base, there are tools like crane [0] and regctl [1] that do that today.
The reason other build tools typically pull the base image first is to support "RUN" build steps that execute a command inside of a container and store the filesystem changes in a new layer. If that functionality is ever added to rules_img, I expect it to have the same performance as other build tools.
[0]: https://github.com/google/go-containerregistry/blob/main/cmd...
I'm not too surprised that out of the box docker images exhibit more of this. While it's good they're fixing it, it feels like maybe some of the core concepts cause pretty systematic issues anytime you try to do anything beyond the basic feature set...
Unfortunately, the amount of work you need to just maintain the build across language and bazel version upgrades is incredibly high. Let alone adding new build steps, or going even slightly off the well-trodded path.
I feel like Bazel would need at least 5 more full-time engineers to eventually turn it into an actually usable build tool outside Big Tech. Right now many critical open source Bazel rules get a random PR every now and then from people who don't actually (have time to) care about the open source community.
My go-to now is to use mise + just to glue together build artifacts from every language's standard build tools. It's not great but at least I get to spend time on programming instead of fixing the build.
Genuine question - also find Bazel frustrating at times.
Bazel is a general purpose tool like Make. But with caching and sandboxing and different syntax.
Make is no less focused on Docker than Bazel is.
Unlike Make however, Bazel does make it easy to share rule sets.
But you don’t need to use other people’s Bazel rule sets any more than you need to use other people’s Make recipes.
This author has a clever way to minimize needing to touch layers at all.
I don't agree with your parent comment about Bazel, but your comment is not fair too. Bazel tries to be better build tool so it took on responsibility on registry / rules_* and get critics for it is a fair game.
The "bloated Bazel" blame is not fair too, but I think somewhat understandable. If you ever going to only do JavaScript, bun or other package manager is enough and "lighter-weight". Same goes to uv + Python bundle. Bazel only shines if you are dealing with your C++ mess and even there, people prefer CMake for reasons beyond me.
Ok, wait, why?
Buildkit from Docker is just a pure bullshit design. Instead of the elegant layer-based system, there's now two daemons that fling around TAR files. And for no real reason that I can discern. But the worst thing is that the caching is just plain broken.
I'd also avoid loading the result back into the docker daemon unless you really need it there. Buildkit can output directly to a registry, or an OCI Layout, each of which will maintain the image digest and support multi-platform images (admittedly, those problems go away with the containerd storage changes happening, but it's still an additional export/import that can be skipped).
All that said, I think caching is often the wrong goal. Personally, I want reproducible builds, and those should bypass any cache to verify each step always has the same output. Also, when saving the cache, every build caches every step, even if they aren't used in future builds. As a result, for my own projects, the net result of adding a cache could be slower builds.
Instead of catching the image build steps, I think where we should be spending a lot more effort is in creating local proxies of upstream dependencies, removing the network overhead of pulling dependencies on every build. Compute intensive build steps would still be slow, but a significant number of image builds could be sped up with a proxy at the CI server level without tuning builds individually.
Well, that's what I've been trying to do. And failing, because it simply doesn't work.
> I'd also avoid loading the result back into the docker daemon unless you really need it there.
I need Docker to provide me a reproducible environment to run lints, inspections, UI tests and so on. These images are quite massive. And because caching in Docker is broken, they were getting rebuilt every time we did a push.
Well. I switched to Podman and podman-compose. Now they do get cached, and the build time is within ~1 min with the help of the GHA cache.
And yes, my deployment builds are produced without any caching.
More importantly, the layers were represented as directories on the host system. So when you wanted to run something in the final container, Docker just needed to reassemble it.
Buildkit has broken all of it. Now building is done, essentially, in a separate system, the "docker buildx" command talks with it over a socket. It transmits the context, and gets the result back as an OCI image that it then needs to unpack.
This is an entirely useless step. It also breaks caching all the time. If you build two images that differ only slightly, the host still gets two full OCI artifacts, even if two containers share most of the layers.
It looks like their Bazel infrastructure optimized it by moving caching down to the file level.
Buldkit is far more efficient than the old model.
And since it's a separate system, there are also these strange limitations. For example, I can't just cache pre-built images in an NFS directory and then just push them into the Buildkit context. There's simply no command for it. Buildkit can only pull them from a registry.
> Buldkit is far more efficient than the old model.
I've yet to see it work faster than podman+buildah. And it's also just plain buggy. Caching for multi-stage and/or parallel builds has been broken since the beginning. The Docker team just ignores it and closes the bugs: https://github.com/moby/buildkit/issues/1981 https://github.com/moby/buildkit/issues/2274 https://github.com/moby/buildkit/issues/2279
I understand why. I tried to debug it, and simply getting it running under a debugger is an adventure.
So far, I found that switching to podman+podman-compose is a better solution. At least my brain is good enough to understand them completely, and contribute fixes if needed.
I'm not quite sure I understand what you are trying to do with nfs there. But you can definitely export the cache to a local filesystem and import it with cache-from. You can also provide named contexts.
"Buildkit can only pull them from a registry" is just plain false.
Each layer is a tarball.
So build your tarballs (concurrently!), and then add some metadata to make an image.
From your comment elsewhere it seems maybe you are expecting the docker build paradigm of running a container and snapshotting it at various stages.
That is messy and has a number of limitations — not the least of which is cross-compilation. Reproducibility being another. But in any case, that definitely not what these rules are trying to do.
I guess the answer for Bazel is "don't do this"? Docker handles cross-compilation by using emulators, btw.
Yes. The Bazel way use to produce binaries, files, directories, and then create an image “directly” from these.
Much as you would create a JAR or ZIP or DEB.
This is (1) fast (2) small and (3) more importantly reproducible. Bazel users want their builds to produce artifacts that are exactly the same, for a number of reasons. Size is also nice…do you really need ls and dozens of other executables in your containerized service?
Most Docker users don’t care about reproducibility. They’ll apt-get install and get one version today and another version tomorrow.
Good? Bad? That’s a value judgement. But Bazel users have fundamentally different objectives.
> emulators
Yeah emulators is the Docker solution for producing images of different architectures.
Since Bazel doesn’t run commands as a running container, it skips that consideration entirely.
Yeah, I do. For debugging mostly :(
> Most Docker users don’t care about reproducibility. They’ll apt-get install and get one version today and another version tomorrow.
Ubuntu has daily snapshots. Not great, but works reasonably well. I tried going down the Nix route, but my team (well, and also myself) struggled with it.
I'd love to have fully bit-for-bit reproducible builds, but it's too complicated with the current tooling. Especially for something like mobile iOS apps (blergh).
Container layers are so large that moving them around is heavy.
So defer that part for the non-hermetic push/load parts of the process, while retaining heremticity/reproducibility.
You can sort of think of it like the IO monad in Haskell…defer it all until the impure end.
It drives me absolute batshit insane that modern systems are incapable of either building or running computer programs without docker. Everyone should profoundly embarrassed and ashamed by this.
I’m a charlatan VR and gamedev that primarily uses Windows. But my deeply unpopular opinion is that windows is a significantly better dev environment and runtime environment because it doesn’t require all this Docker garbage. I swear that building and running programs does not actually have to be that complicated!! Linux userspace got pretty much everything related to dependencies and packages very very very wrong.
I am greatly pleased and amused that the most reliable API for gaming in Linux is Win32 via Proton. That should be a clear signal that Linux userspace has gone off the rails.
On Linux vs Win32 flame warring: can you be more specific? What specifically is very very wrong with Linux packaging and dependency resolution?
Fair. Docker does trigger my predator drive.
I’m pretty shocked that the Bazel workflow involves downloading Docker base images from external URLs. That seems very unbazel like! That belongs in the monorepo for sure.
> What specifically is very very wrong with Linux packaging and dependency resolution?
Linux userspace for the most part is built on a pool of global shared libraries and package managers. The theory is that this is good because you can upgrade libfoo.so just once for all programs on the system.
In practice this turns into pure dependency hell. The total work around is to use Docker which completely nullifies the entire theoretic benefit.
Linux toolchains and build systems are particularly egregious at just assuming a bunch of crap is magically available in the global search path.
Docker is roughly correct in that computer programs should include their gosh darn dependencies. But it introduces so many layers of complexity that are solved by adding yet another layer. Why do I need estargz??
If you’re going to deploy with Docker then you might as well just statically link everything. You can’t always get down to a single exe. But you can typically get pretty close!
Not every dependency in Bazel requires you to "first invent the universe" locally. Lots of examples of this like toolchains, git_repository, http_archive rules and on and on. As long as they are checksum'ed (as they are in this case) so that you can still output a reproducible artifact, I don't see the problem
I suppose a URL with checksum is kinda sorta equivalent. But the article adds a bunch of new layers and complexity to avoid “downloading Cuda for the 4th time this week”. A whole lot of problems don’t exist if they binary blobs exist directly in the monorepo and local blob store.
It’s hard to describe the magic of a version control system that actually controls the version of all your dependencies.
Webdev is notorious for old projects being hard to compile. It should be trivial to build and run a 10+ year old project.
I don’t think they’re opposites. It seems orthogonal to me.
If you have a bunch of remote execution workers then ideally they sit idle on a full (shallow) clone of the repo. There should be no reason to reset between jobs. And definitely no reason to constantly refetch content.
Like, using the words "leverage", "matters for...", "as for", and so on. And you could almost hear him doing the bullet points.
When you work with AI a lot, it changes your vocabulary.