Using S3 as a Container Registry
ochagavia.nl
ochagavia.nl
> According to the specification, a layer push must happen sequentially: even if you upload the layer in chunks, each chunk needs to finish uploading before you can move on to the next one.
As far as I've tested with DockerHub and GHCR, chunked upload is broken anyways, and clients upload each blob/layer as a whole. The spec also promotes `Content-Range` value formats that do not match the RFC7233 format.
(That said, there's parallelism on the level of blobs, just not per blob)
Another gripe of mine is that they missed the opportunity to standardize pagination of listing tags, because they accidentally deleted some text from the standard [1]. Now different registries roll their own.
[1] https://github.com/opencontainers/distribution-spec/issues/4...
Yes you can theoretically create a repository object in ECR in Terraform to mimic that behavior, but it sucks in pipelines where the result image path is dynamic - you need to give more privileges to the IAM role of the CI pipeline than I'm comfortable with, not to mention that I don't like any AWS resources managed outside of the central Terraform repository.
[1] https://stackoverflow.com/questions/64232268/storing-images-...
Regardless it's a huge pain that ECR doesn't support this. Everybody I know of who has used ECR has run into this.
There's a long standing issue open which I've been subscribed to for years now: https://github.com/aws/containers-roadmap/issues/853
That’s par for everything around Docker and containers. As a user experience Docker is amazing, but as technology it is hot garbage. That’s not as much of a dig on it as it might sound: it really was revolutionary; it really did make using Linux namespaces radically easier than they had ever been; it really did change the world for the better. But it has always prioritised experience over technology. That’s not even really a bad thing! Just as there are tons of boring companies solving expensive problems with Perl or with CSVs being FTPed around, there is a lot of value in delivering boring or even bad tech in a good package.
It’s just sometimes it gets sad thinking how much better things could be.
Looking at it now, it won't even run in the latest systemd, which now refuses to boot with cgroups v1. Good luck even accessing /dev/null under cgroups v2 with systemd.
Docker is important because:
1) it made a convenient process to build a “system” image of sorts, upload it, download it, and run it.
2) (the important bit!) Enough people adopted this process for it to become basically a standard
Before Docker, it wasnt uncommon to ship some complicated apps in VMs. Packaging those was downright awful with all of the bespoke scripting needed for the various steps of distribution. And then you get a new job? Time to learn a brand new process.
And for people that disagree, please write a library like TestContainers using cobbled together bash scripts, that can download and cleanly execute and then clean up almost any common use backend dependency.
I mean, Podman exists, as do lots of custom build tools and other useful options. Personally, I mostly just stick with vanilla Docker (and Compose/Swarm), because it's pretty coherent and everything just fits together, even if it isn't always perfect.
Either way, agreed about the concepts behind the technology making things better for a lot of folks out there, myself included (haven't had prod issues with mismatched packages or inconsistent environments in years at this point, most of my personal stuff also runs on containers).
I don’t know about that (hyperbole aside). I’ve been in IT for more than 25 years now. I can’t see that Docker container actually delivered any tangible benefits in terms of end-product reliability or velocity of development to be honest. This might not necessarily be Dockers fault though, maybe it’s just that all the potential benefits get eaten up by things like web development frameworks and Kubernetes.
But at the end of the day, todays Docker-based web app development delivers less than fat-client desktop app development delivered 20 years ago, as sad as that is.
You mean, aside from not having to handle installation of your software on your users' machines?
Also I'm not sure this is related to docker at all.
I know, in theory this stuff all sounds very nice. With web apps, you can "deploy" within seconds ideally, compared to say at least a couple of minutes or maybe hours with desktop software distribution.
But all of that doesn't really matter if the endusers now actually have to wait weeks or months to get the features they want, because all that new stuff added so much complexity that the devs have to handle.
And that was my point. In terms of enduser quality, I don't think we have gained much, if anything at all.
The fact that I don’t have to install dependencies on a server, or set up third-party applications like PHP, Apache, Redis, and the myriad of other packages anymore, or manage config files in /etc, or handle upgrades of libc gracefully, or worry about rolling restarts and maintenance downtime… all of this was solvable before, but has become radically easier with containers.
Packaging an application and its dependencies into a single, distributable artifact that can be passed around and used on all kinds of machines was a glorious success.
(And this enables some increasingly wondrous acts of software archaeology, due to people building OCI images not for preservation, but just for "use at the time" — and then just never purging them from whatever repository they've pushed them to. People are preserving historical software builds in a runnable state, completely by accident!)
Before Docker, the nearest thing you could do to this was to package software as a VM image — and there was no standard for what "a VM image" was, so this wasn't a particularly portable/long-term solution. Often VM-image formats became unsupported faster than the software held in them did!
But now, with OCI images, we're nearly to the point where we've e.g. convinced academic science to publish a paper's computational apparatus as an OCI image, so that it can be pulled 10 years later when attempting to replicate the paper.
I think you’re onto part of the problem here. The thing is that you have to snapshot a lot of nowadays software together with its runtime environment.
I mean, I can still run Windows software (for example) that is 10 years or older without that requirement.
Anyone who worked with Perl CGI and CPAN would tell you managing dependencies across environments has always been an issue. Regarding desktop software; the phrase "DLL hell" precedes NPM and pip by decades and is fundamentally the same dependency management challenge that docker mostly solves.
Additionally, I think lots of projects became able to adopt better tooling faster, since the barrier to use container-based tools is lower. Just think of GitHub Actions, which suddenly enabled everyone and their mother to adopt CI pipelines. That simply wasn’t possible before, and has led to more software adopting static analysis and automated testing, I think.
For self-hosted software, be it for business or personal use, it immensely simplified how a software package can be pulled, and run in isolated environment.
Dependency hell is avoided, and you can easily create/start/stop/delete a specific software, without affecting the rest of the host machine.
I had scripts that would automatically generate the Apache configuration to deploy a new site in less than 30 seconds.
At that time I found that most web sites have just a few things to configure: often a database connection, the path to where files are, and maybe a cryptographic secret. If you are systematic about where you put your files and how you do your configuration running servers with a lot of sites is about as easy as falling off a log, not to mention running development, test, staging, prod and any other sites you need.
I have a Python system now with gunicorn servers and celery workers that exists in three instances on my PC, because I am disciplined and everything is documented I could bring it up on another machine manually pretty quickly, probably more quickly than I could download 3GB worth of docker images over my ADSL connection. With a script it would be no contest.
There also was a time I was building AMIs and even selling them on the AMZN marketplace and the formula was write a Java program that writes a shell script that an EC2 instance runs on boot, when it is done it sends a message through SQS to tell the Java program to shut down and image the new machine.
If Docker is anything it is a system that turns 1 MB worth of I/O into 1 GB of I/O. I found Docker was slowing me down when I was using a gigabit connection, I found it basically impossible to do anything with it (like boot up an image) on a 2MB/sec ADSL connection, with my current pair of 20MB/s connections it is still horrifyingly slow.
I like how the OP is concerned about I/O speed and bringing it up and I think it could be improved if there was a better cache system (e.g. Docker might even work on slow ADSL if it properly recovered from failed downloads)
However I think Docker has a conflict between “dev” (where I’d say your build is slow if you ever perceive yourself to be waiting) and “ops” (where a 20 minute build is “internet time”)
I think ops is often happy with Docker, some devs really seem to like it, but for some of us it is a way to make a 20 sec task a 20 minute task.
As for docker image size, really just depends on dev discipline for better or for worse. The nginx image, for example, adds about 1MB of data on top of the whatever you did with your website.
By virtue of shipping around an entire system's worth of libraries as a deployment artifact, you are indeed drastically increasing the payload size. It's easy to question whether payload efficiency is worthwhile when the advent of >100, and even >1000 Mbit internet connections available to the home, but that is certainly not the case everywhere. That said, assuming smart squashing of image deltas and basing off of a sane upstream image, much of that pain is felt only once.
You bring up that you built a system that helped you quickly and efficiently configure systems, and that discipline and good systems design can bring many of the same benefits that containerized workloads do. No argument! What the Docker ecosystem provided however was a standard implemented in practice that became ubiquitous. It became less important to need to build one's own system, because the container image vendor could define that, using a collection of environment variables or config files being placed in a standardized location.
You built up a great environment, and one that works well for you. The containerization convention replicates much of what you developed, with the benefit that it grabbed a majority mindshare, so now many more folks are building with things like standardization of config, storage, data, and environment in mind. It's certainly not the only way to do things, and much as you described, it's not great in your case. But if something solves a significant amount of cases well, then it's doing something right and well. For a non inconsequential amount of people, trading bandwidth and storage for operational knowledge and complexity are a more than equitable trade
Now I don't have to care what OS the host is running, or what dependencies are installed, and adding replicas is either automatic or editing a number in a config file.
Containerization and orchestration tools like k8s have made life so much easier.
A lot of us were just forced to "switch" from VMs to Docker; Docker that still got deployed to a VM.
And then we got forced to switch to podman as they didn't want to pay for Docker.
Washing clothes was possible before people had a washing machine, too; I’m not sure they would want to go back to that, though.
I was there in the VM time, and I had to set up appliances shipped as a VM instance. It was awful. The complexity around updates and hypervisors, and all that OS adjustment work just to get a runtime environment going, that just disappeared with Docker (if done right, I’ll give you that).
Organisations manage to abuse technology all the time. Remember when Roy Fielding wrote about using HTTP sensibly to transfer state from one system to another? Suddenly everything had to be „RESTful“, which for most people just meant that you tried to use as many HTTP verbs as possible and performed awkward URL gymnastics to get speaking resource identifiers. Horrible. But all of this doesn’t mean REST is a bad idea of itself - it’s a wonderful one, in fact, and can make an API substantially easier to reason about.
I think it actually winds up speeding up ci/cd docker builds, too.
It's somewhat surprising that the results of that process are looking to be nearly as fly-by-the-seat-of-your-pants as Docker itself is.
Anyone tried it?
It does mention that it's limited to 500MB per layer.
For some people's use case that limitation might not be a big deal, but for others that's a dealbreaker.
* Pushing with docker is limited to images that have layers of maximum size 500MB. Refer to maximum request body sizes in your Workers plan.
* To circumvent that limitation, you can manually add the layer and the manifest into the R2 bucket or use a client that is able to chunk uploads in sizes less than 500MB (or the limit that you have in your Workers plan).
Edit: to clarify, I'm talking about sequentially pushing a _single_ layer's contents. You can, of course, push multiple layers in parallel.
~~I added parallel pushes to docker I think, unless I'm mixing up pulls & pushes, it was a while ago.~~ My stuff was around parallelising the checks not the final pushes.
Edit - does a layer say which layer it goes "on top" of? If so perhaps that's the reason, so the IDs of what's being pointed to exist.
It's a mystery... Here are the bits of the OCI spec about multipart pushes (https://github.com/opencontainers/distribution-spec/blob/58d...). In short, you can only upload the next chunk after the previous one finishes, because you need to use information from the response's headers.
That's chunks of a single layer though, not multiple layers right?
(I just updated my original comment to make clear I'm talking about single-layer pushes here)
I think it was a mistake to make layers as a storage model visible in to the end user. This should just have been an internal implementation detail, perhaps similar to how Git handles delta compression and makes it independent of branching structure. We also should have delta pushes and pulls, using global caches (for public content), and the ability to start containers while their image is still in transfer.
Theoretically there's nothing stopping you from building the docker image and "re-layering it", as they're "just" bundles of tar files at the end of the day.
eg: `RUN ... ; LAYER /usr ; LAYER /var ; LAYER /etc ; LAYER [discard|remainder]`
BEGIN
RUN foo
RUN bar
COMMITIn true "when all you have is a hammer" fashion, as very best I can tell that syntax= directive is pointing to a separate docker image whose job it is to read the file and translate it into builtkit api calls, e.g. https://github.com/moby/buildkit/blob/v0.15.0/frontend/docke...
But, again for clarity: I've never tried such a stunt, that's just the impression I get from having done mortal kombat with builtkit's other silly parts
RUN <<EOF
foo
bar
EOF
This might work with pip but I absolutely hate pip and using poetry with great success. I will investigate how to do this with poetry.
> For the last four months I’ve been developing a custom container image builder, collaborating with Outerbounds
I know you said this was something for another blog post but could you already provide some details? Maybe a link to a GitHub repo?
Background: I'm looking for (or might implement myself) a way to programmatically build OCI images from within $PROGRAMMING_LANGUAGE. Think Buildah, but as an API for an actual programming language instead of a command line interface. I could of course just invoke Buildah as a subprocess but that seems a bit unwieldy (and I would have to worry about interacting with & cleaning up Buildah's internal state), plus Buildah currently doesn't support Mac.
Thanks for the link! Though I'm less worried about the tarball / OCI spec part, more about platform compatibility. I tried running runc/crun by hand at some point and let's just say I've done things before that were more fun. :)
Buildah uses it: https://github.com/containers/buildah/blob/main/go.mod#L27C2...
Edit: buildkit seems to be the same, used by docker, but needs a daemon?
But that's an implementation detail, and I suspect isn't one that's meaningful or intentional. Your S3 approach should work fine btw, I've done it before in a prior life when I was at a company shipping huge images and $.10/gb/month _really_ added up.
You lose the 'bells and whistles' of ECR, but those are pretty limited (imho)
At this point you’d validate that the layers exist and have been uploaded, otherwise you’d just bail out?
And those missing chunks would be handled by the normal registry GC, which evicts unreferenced layers?
I think OCI ordered manifests first to "open the flow", but then close is only when the manifests last entry was completed - which led to this ordered upload problem.
If your uploader knows where the chunks are going to live (OCI is more or less CAS, so it's predictable), it can just put them there in any order as long as it's all readable before something tries to pull it.
When the client finalizes a blob upload, they need to supply the digest of the full blob. This requirement evidently serves to enable the server side to validate the integrity of the supplied bytes. If the server only started checking the digest as part of the finalize HTTP request, it would have to read back all the blob contents that had already been written into storage in previous HTTP requests. For large layers, this can introduce an unreasonable delay. (Because of specific client requirements, I have verified my implementation to work with blobs as large as 150 GiB.)
Instead, my implementation runs the digest computation throughout the entire sequence of requests. As blob data is taken in chunk by chunk, it is simultaneously streamed into the digest computation and into blob storage. Between each request, the state of the digest computation is serialized in the upload URL that is passed back to the client in the Location header. This is roughly the part where it happens in my code: https://github.com/sapcc/keppel/blob/7e43d1f6e77ca72f0020645...
I believe that this is the same approach that the reference implementation uses. Because digest computation can only work sequentially, therefore the upload has to proceed sequentially.
[1] https://github.com/sapcc/keppel [2] https://github.com/distribution/distribution
Personally, I just use Nexus because it works well enough (and supports everything from OCI images to apt packages and stuff like a custom Maven, NuGet, npm repo etc.), however the configuration and resource usage both are a bit annoying, especially when it comes to cleanup policies: https://www.sonatype.com/products/sonatype-nexus-repository
That said:
> More specifically, I logged the requests issued by docker pull and saw that they are “just” a bunch of HEAD and GET requests.
this is immensely nice and I wish more tech out there made common sense decisions like this, just using what has worked for a long time and not overcomplicating.
I am a bit surprised that there aren't more simple container repositories out there (especially with auth and cleanup support), since Nexus and Harbor are both a bit complex in practice.
https://distribution.github.io/distribution/storage-drivers/...
Of course, if you use images for dev only, all of that are worthless and you just store your images on your dev machine
docker save alpine:3.19 > alpine.tar
docker load < alpine.tar
It's a difficult uphill battle to get people interested in how things work under the hood, which is what you need in order to know you can do things like easily host your own package repositories.
Yes I can host my packages and write tooling around it to make it easy. But JFrog already has all the tooling around it, and it integrates with current tooling. Why would I write the whole thing again?
> I don't do a ton with Docker outside dev tooling, but I have never understood why private container registries even exist?
You know the options and have made a conscious choice:
> Yes I can host my packages and write tooling around it to make it easy. But JFrog already has all the tooling around it, and it integrates with current tooling. Why would I write the whole thing again?
So presumably you are not the kind of people I was talking about.
EDIT: I'm also assuming by the rent seeking part that the parent is referring to paid hosted services like ECR etc.
Because you want a central store someplace with all the previous versions that is easily accessible to lots of consumers.
I don't want to build my app and then have to push it to every single place that might run it. Instead, I'll build it and push it to a central repo and have everything reference that repo.
> It just smells like rent seeking.
You don't need to pay someone to host a private repo for you. There are lots of tools out there so you can self-host.
Related ECR APIs:
- InitiateLayerUpload API: called at the beginning of upload of each image layer
- UploadLayerPart API: called for each layer chunk (up to 20 MB)
- PutImage API: called after layers are uploaded, to push image manifest, containing references to all image layers
The only weird thing seems to be that you have to upload layer chunks in base64 encoding, which increases data for ~33%.
I do wonder though how you would deal with the Docker-Content-Digest header. While not required it is suggested that responses should include it as many clients expect it and will reject layers without the header.
Another thing to consider is that you will miss out on some feature from the OCI 1.1 spec like the referrers API as that would be a bit tricky to implement.
Awesome. Developer experience is so much better when CI doesn't take ages. Every little bit counts.
> S3 115 MiB/s (1.7 s)
It's great that it's faster but absolutely, it's only an improvement of 6.5s observed, as you said, on the CI server. And it means using something for a purpose that it's not intended for. I'd hate to have to spend time debugging this if it breaks for whatever reason.
The default storage backend does not keep compressed layers, so those need to be recreated and digested on push.
With the new store all that stuff is kept and reused.
A hash tree would have been nicer, and parallel uploads would have been an extra bonus.
blake3 support has been proposed both in the OCI spec and in the runtimes, which at least for runtimes I expect to happen soon.
I tend to think gzip is the bigger problem, though.
Sure, and one can construct a perfectly nice tree hash from SHA256. (AWS Glacier did this, but their construction should not be emulated.)
Adding a rule that you hash the files in question in, say, 1 MiB chunks and hash the resulting hashes (and maybe that’s it, or maybe you add another level) is maybe 10 lines of code in any high level language.
FWIW, using a (root hash, data length) pair hides many sins, although I haven’t formally proven this. And I don’t think that extension attacks are very relevant to the OCI use case.
I don't see any reason why ECR couldn't support parallel uploads as an optimization. Provide an alternative to `docker push` for those who care about speed that doesn't conform to the spec.
My interest was mainly around a hardening stand point. The base idea was the release system through IAM permissions would be the only system with any write access to the underlying S3 bucket. All the public / internet facing components could then be limited to read only access as part of the hardening.
This would of course be in addition to signing the images, but I don't think many of the customers at the time knew anything about or configured any of the signature verification mechanisms.
There is a real usecase for this in some high security sectors. I can't put complete info here for the security reasons, let me know if you are interested.
[Chorus: Chilli & T-Boz]
No, I don't want no scrub
A scrub is a guy that can't get no love from me
Hangin' out the passenger side of his best friend's ride
Trying to holla at me
I don't want no scrub
A scrub is a guy that can't get no love from me
Hangin' out the passenger side of his best friend's ride
Trying to holla at meEdit: CF already did it https://github.com/cloudflare/serverless-registry
A Docker registry backed by Workers and R2.
the author is missing something huge - ECR does a security scan on upload, too.
It would be nice if a Kubernetes distro took a page out of the "serverless" playbook and just embedded a registry. Or maybe I should just use GHCR
If you want some lightweight registry, use "official" docker registry. I'm running it inside Kubernetes and it consumes it just fine.
In my case it's still faster than doing the builds would be -- but I'm definitely gonna have to get machines with persistent local cache in the mix at some point so that these operations will finish within a few seconds instead of a few minutes ...
Does this not do what you want? https://docs.aws.amazon.com/AmazonECR/latest/userguide/lifec...
I can't speak to the other "registry cloud products" except for GitLab, which is its own special UX nonsense, but they also support expiry after enough whisky consumption
In theory it wouldn't be hard to just take docker-distribution and run it as a pod in the cluster with an attached volume if you wanted a registry in the cluster. So it's probably somewhere between trivial and takes a bit of effort if you're really motivated to have something in cluster.
Here are all the projects already using zot in some form or another.