containerd/design/snapshots.md

Snapshots

Docker containers, from the beginning, have long been built on a snapshotting methodology known as layers. Layers provide the ability to fork a filesystem, make changes then save the changeset back to a new layer.

Historically, these have been tightly integrated into the Docker daemon as a component called the graphdriver. The graphdriver allows one to run the docker daemon on several different operating systems while still maintaining roughly similar snapshot semantics for committing and distributing changes to images.

The graphdriver is deeply integrated with the import and export of images, including managing layer relationships and container runtime filesystems. The behavior of the graphdriver informs the transport of image formats.

In this document, we propose a more flexible model for managing layers. It focuses on providing an API for the base snapshotting functionality without coupling so tightly to the structure of images and their identification. The minimal API simplifies behavior without sacrificing power. This makes the surface area for driver implementations smaller, ensuring that behavior is more consistent between implementations.

These differ from the concept of the graphdriver in that the LayerManipulator has no knowledge of images or containers. Users simply prepare and commit directories. We also avoid the integration between graph drivers and the tar format used to represent the changesets.

The best aspect is that we can get to this model by refactoring the existing graphdrivers, minimizing the need to new code and sprawling tests.

Scope

In the past, the graphdriver component has provided quite a lot of funcionality in Docker. This includes serialization, hashing, unpacking, packing, mounting.

This snapshot manager will only provide mount-oriented snapshot access with minimal metadata. Serialization, hashing, unpacking, packing and mounting are not included in this design, opting for common implementations between graphdrivers, rather than specialized ones. This is less of a problem for performance, since direct access to changesets is provided in the interface.

Architecture

The Snapshot Manager provides an API for allocating, snapshotting and mounting abstract, layer-based filesystems. The model works by building up sets of directories with parent-child relationships, known as Snapshots.

Every snapshot is represented by an opaque diff directory, which acts as a handle to the snapshot. It may contain driver specifc data, including changeset data, parent information and arbitrary metadata.

The diff directory for a snapshot is created with a transactional operation. Each snapshot may have one parent snapshot. When one starts a transaction on an existing snapshot, the result may only be used as a parent after being committed. The empty string diff directory is a handle to the empty snapshot, which is the ancestor of all snapshots.

The target directory represents the active snapshot location. The driver may maintain internal metadata associated with the target but the contents is generally manipulated by the client.

Operations

The manifestation of snapshots is facilitated by the mount object and user-defined directories used for opaque data storage. When creating a new snapshot, the caller provides a directory where they would like the snapshot to be mounted, called the target. This operation returns a list of mounts that, if mounted, will have the fully prepared snapshot at the requested path. We call this the prepare operation.

Once a path is prepared and mounted, the caller may write new data to the snapshot. Depending on application, a user may want to capture these changes or not.

If the user wants to keep the changes, the commit operation is employed. The commit operation takes the target directory, which represents an open transaction, and a diff directory. A successful result will end up with the difference between the parent and snapshot in the diff directory, which should be treated as opaque by the caller. This new diff directory can then be used as the parent in calls to future prepare operations.

If the user wants to discard the changes, the rollback operation will release any resources associated with the snapshot. While rollback may a rare operation in other transactional systems, this is a common operation for containers. After removal, most containers will have rollback called.

For both rollback and commit the mounts provided by prepare should be unmounted before calling these methods.

Graph metadata

As snapshots are imported into the container system, a "graph" of snapshots and their parents will form. Queries over this graph must be a supported operation. Subsequently, each snapshot ends up representing

Path Management

No path layout for snapshot locations is imposed on the caller. The paths used by the snapshot drivers are largely under control of the caller. This provides the most flexibility in using the snapshot system but requires discipline when deciding which paths to use and which ones to avoid.

We may provide a helper component to manage diff path layout when working with OCI and docker images.

How snapshots work

To bring the terminology of snapshots, we are going to demonstrate the use of the snapshot manager from perspective of importing layers. We'll use a Go API to represent the process.

Importing a Layer

To import a layer, we simply have the Snapshot Manager provide a list of mounts to be applied such that our dst will capture a changeset. We start out by getting a path to the layer tar file and creating a temp location to unpack it to:

layerPath, tmpLocation := getLayerPath(), mkTmpDir() // just a path to layer tar file.

Per the terminology above, tmpLocation is known as the target. layerPath is simply a tar file, representing a changset. We start by using SnapshotManager to prepare the temporary location as a snapshot point:

lm := SnapshotManager()
mounts, err := lm.Prepare(tmpLocation, "")
if err != nil { ... }

Note that we provide "" as the parent, since we are applying the diff to an empty directory. We get back a list of mounts from SnapshotManager.Prepare. Before proceeding, we perform all these mounts:

if err := MountAll(mounts); err != nil { ... }

Once the mounts are performed, our temporary location is ready to capture a diff. In practice, this works similar to a filesystem transaction. The next step is to unpack the layer. We have a special function, unpackLayer that applies the contents of the layer to target location and calculates the DiffID of the unpacked layer (this is a requirement for docker implementation):

digest, err := unpackLayer(tmpLocation, layer) // unpack into layer location
if err != nil { ... }

When the above completes, we should have a filesystem the represents the contents of the layer. Careful implementations should verify that digest matches the expected DiffID. When completed, we unmount the mounts:

unmount(mounts) // optional, for now

Now that we've verified and unpacked our layer, we create a location to commit the actual diff. For this example, we are just going to use the layer digest, but in practice, this will probably be the ChainID:

diffPath := filepath.Join("/layers", digest) // name location for the uncompressed layer digest
if err := lm.Commit(diffPath, tmpLocation); err != nil { ... }

The new layer has been imported as a snapshot into the SnapshotManager under the name diffPath. diffPath, which is a user opaque directory location, can then be used as a parent in later snapshots.

Importing the Next Layer

Making a layer depend on the above is identical to the process described above except that the parent is provided as diffPath when calling Snapshot.Prepare:

mounts, err := lm.Prepare(tmpLocation, parentDiffPath)

Because have a provided a parent, the resulting tmpLocation, after mounting, will have the changes from above. Any new changes will be isolated to the snapshot target.

We run the same unpacking process and commit as above to get the new diff.

Running a Container

To run a container, we simply provide SnapshotManager.Prepare the diff of the image we want to start the container from. After mounting, the prepared path can be used directly as the container's filesystem:

mounts, err := lm.Prepare(containerRootFS, imageDiffPath)

The returned mounts can then be passed directly to the container runtime. If one would like to create a new image from the filesystem, SnapshotManipulator.Commit is called:

if err := lm.Commit(newImageDiff, containerRootFS); err != nil { ... }

Alternatively, for most container runs, Snapshot.Rollback will be called to signal SnapshotManager to abandon the changes.