Gleicher Tag, anderes Image: Azure Container Registry erklärt

Your app scales out. A new server starts and pulls the same image as the others: inference-api, tag latest. But the new server runs different code than the rest. Nobody deployed anything. Someone just pushed.

In this video, you will see how Azure Container Registry stores images, why a tag is not the same as a version, how to build images in the cloud and keep them patched, and a tagging scheme that keeps production predictable.

Start with the structure. Azure Container Registry is a private registry for container images, run by Azure. At the top is the registry itself. It has one address, called the login server: your registry name, followed by azurecr.io. Inside the registry are repositories. A repository holds all versions of one image, for example inference-api. Repository names can contain slashes, like production/inference-api, or ml-team/model-server. These namespaces group images by team or environment. And access can be granted per repository, so a team only reaches its own images. Inside a repository are the artifacts: container images, but also Helm charts and other OCI artifacts. Now look inside one image. It is a stack of layers. Each instruction in your Dockerfile that changes files adds a layer. Layers are stored by their content. So when ten images share the same Python base layer, the registry keeps that layer only once. And a server that already has the layer doesn't download it again. On top of the layers sits the manifest. It lists every layer of the image, and it has a digest: a SHA-256 hash that identifies it. Registries come in three tiers: Basic, Standard and Premium. Pushing and pulling work the same in all of them. Premium adds features for larger setups, like geo-replication to other regions and private endpoints.

Now the most important idea in this video. A tag is a sticky note. A digest is a fingerprint. A tag, like version 1.2.0 or latest, is a readable name for an image. One image can carry several tags at once, for example 1.2.0 and stable. But tags can move. When you push a new image with a tag that already exists, the tag jumps to the new image. The old image stays in the registry, but now no tag points to it. It is untagged. That is exactly what happened at the start of this video. Someone pushed a new image as latest. The new server pulled latest a few minutes later, and got the new image. There is one more trap. If you leave out the tag completely, Docker uses latest automatically. A digest works differently. It is calculated from the content, so it can never point to another image. Pull by digest, with an @ sign and the SHA-256 hash, and you get exactly the same image every time. So: tags are for people. Digests are for guarantees.

So how do images get into the registry? The classic way is to build on your own machine with Docker, and push. That works, until five developers build the same image on five different laptops. ACR Tasks moves the build into Azure. The simplest form is a quick task, with one command: az acr build. You give it the registry, an image name with a tag, and a build context. The context is where your source files are: a local folder, a Git repository, or a remote tarball. The command uploads the context, runs the Docker build in Azure, and pushes the result to your registry. Like docker build plus docker push, but you don't need Docker on your machine at all. Every build runs in the same environment, no matter who starts it. And ACR Tasks can build images for Linux, Windows and ARM. You can also run a container from your registry in the cloud. az acr run takes an image and a command, for example python --version, and shows you the output. A quick smoke test before anything is deployed.

Quick tasks run when you call them. A task created with az acr task create can also start on its own. There are three kinds of triggers. First, a source code trigger. Point the task at a Git repository in GitHub or Azure DevOps, and every commit starts a build. Builds for pull requests are possible too, but they are off by default. Second, a timer. A cron expression in UTC, for example every night at midnight. And third, the trigger that saves the most work: the base image trigger. Your Dockerfile starts with a FROM line, for example python, tag 3.12-slim. When that base image gets a security patch, every image built on top of it is out of date. ACR Tasks learns the base image of your app the first time the task runs. From then on, when the base image is updated, it rebuilds your app automatically. This works when the base image is in the same registry, in another Azure container registry, or in a public repository on Docker Hub or the Microsoft Container Registry. A base image in an Azure registry triggers the rebuild right away. Public base images are checked every ten to sixty minutes. One condition matters: the base image must keep a stable tag, like 3.12-slim. If the patch arrives under a new tag, nothing is triggered. For more than one step, you describe the task in YAML. A multi-step task can build the image, run the tests inside it, and push it as the last step. A failed step fails the whole run, so an image with failing tests is never pushed. And the run ID gives every build its own unique tag.

That brings us back to tags, because there are two kinds, and each one has its job. Stable tags are reused. Think of them as version channels. Tag 1 always points to the newest version 1. Tag 1.1 points to the newest patch of 1.1. Stable tags are right for base images, because they let patches flow in. That is exactly what the base image trigger needs. Unique tags are never reused. Every push gets a new one: a build ID, a timestamp, or a commit hash. The build ID is usually the best choice, because it leads straight back to the pipeline run, with its logs and test results. A commit hash can repeat, because a base image rebuild uses the same commit. Unique tags are right for deployments. Every server pulls exactly the same image, and a rollback is just the previous tag. So the rule is simple: stable tags for base images, unique tags for deployments. And latest stays out of production. One more step protects what is running. Lock the deployed image by setting write-enabled to false. Now nobody can overwrite or delete that tag by accident, until you unlock it.

Every moved tag leaves something behind: an untagged image. Over time, these orphans fill your registry and your bill. The cleanup tool is acr purge. It runs as a task inside your registry, on demand or on a schedule. But read the filter carefully, because this is a common mistake. The filter selects tags. Purge deletes every matching tag that is older than the age you set. The untagged option removes untagged images in addition to those tags. So a filter for all tags in inference-api, with an age of thirty days, deletes every tag in that repository that is older than thirty days. Including the one in production, unless it is locked. If you only want the orphans, use the --untagged-only option. And run every new purge command with --dry-run first. It shows what would be deleted, without deleting anything. On the Premium tier, a retention policy can do this for you. It deletes untagged images automatically after a number of days you choose. It is still in preview. One warning: if a system pulls images by digest, the image it needs might have no tag at all, and cleanup deletes untagged images. Unique tags avoid that problem.

To sum it up. A registry holds repositories, and repositories hold images built from shared layers. A tag is a sticky note that can move. A digest is a fingerprint that never changes. ACR Tasks builds in the cloud, and rebuilds on a commit, on a schedule, or when the base image is patched. Stable tags for base images, unique tags for deployments, and a lock on what runs in production. And clean up with care, because purge deletes the tags your filter selects.

Now you know where your images live, and how to keep them under control. If this was useful, subscribe for more practical .NET and Azure.