A core part of the Docker model is the efficient use of layered images and containers based on images. To implement this Docker relies heavily on various filesystem features in the kernel. This document will explain how this works and give some advice in how to best use it.
The basis of the filesystem use in Docker is the storage backend abstraction. A storage backend allows you to store a set of layers each addressed by a unique name. Each layer is a filesystem tree that can be mounted when needed and modified. New layers can be started from scratch, but they can also be created with a specified parent. This means that they start out with the same content as the parent layer, but over time they may diverge. This is typically implemented in the backend by some form of copy-on-write in order to allow the create operation to be fast.
Some backends also have extra operations that allow efficient computation of the differences between layers. These operations all have fallback implementations, but they can be slower as they have to compare all the files in the layers.
Based on the layer abstraction Docker implements the highlevel concepts of images and containers.
Each Docker image on the system is stored as a layer, with the parent being the layer of the parent image. To create such an image a new layer is created (based on the right parent) and then the changes in that image are applied to the newly mounted filesystem.
Containers are a bit more complicated. Each containers has two layers,
one (called the
init layer), which is based on an image layer and a
child of that which contains the actual container content. The init
layer contains a few files that must always exist in Docker containers
/.dockerinit). Committing a container (and thus creating an
image) involves finding all the changes from the init layer to the
container layer and applying those to a new layer based on the same
image the container used.
The vfs Backend
The vfs backend is a very simple fallback that has no copy-on-write support. Each layer is just a separate directory. Creating a new layer based on another layer is done by making a deep copy of the base layer into a new directory.
Since this backend doesn’t share diskspace use between layers, and since creating a new layer is a slow operation this is not a very practical backend. However, it still has its uses, for instance to verify other backends against, or if you need a super robust (if slow) backend that works everywhere.
The devicemapper Backend
The devicemapper backend uses the device-mapper thin provisioning module (dm-thinp) to implement layers. Device-mapper is the kernel part of the LVM2 logical volumes system, so this is is a block-level copy-on-write system.
The thin provisioning module takes two block devices (the data and the
metadata devices) and creates a
pool of space that can be used to
create other block devices from. Such block devices are
provisioned meaning they start out empty, and the parts that are
unused are not allocated. Additionally it is possible to take a
copy-on-write snapshot of a device, producing a new device.
On first startup Docker creates a base device on the thin pool, containing a empty ext4 filesystem. All other layers are (directly or indirectly) snapshots of this base layer. The filesystem has a fixed size, which means that all the images and containers have a maximum size. By default this size is 10GB, although due to the thin provisioning each device tends to use only less space in the pool.
In order to set up a thin pool you need two block devices, which is
not always something users want to deal with. So, by default Docker
creates two regular files inside
metadata and uses loopback to create block devices of these for the
thin pool. These files are by default 100GB (data) and 2GB (metadata),
but they are
sparse meaning that unused blocks are not mapped, and
thus does not take space on the host filesystem. Additionally there
is and external file
/var/lib/docker/devicemapper/devicemapper/json) that contains the
mapping from Docker layer names to thin pool ids.
The loopback setup makes it very easy to start using docker on any machine, but it is not the most efficient way to use devicemapper. On production servers it is recommended that you set up the docker thin pool to use real block devices. For best performance the metadata device should be on a SSD driver, or at least on a different spindle from the data device.
In order to support multiple Docker instances on a system the thin
pool will be named something like
0:30 part is the minor/major device nr and
19478248 is the
inode number of the /var/lib/docker/devicemapper directory. The same
prefix is used for the actual thin devices.
The btrfs Backend
The brtfs backend requires
/var/lib/docker to be on a btrfs filesystem
and uses the filesystem level snapshotting to implement layers.
Each layer is stored as a btrfs subvolume inside
/var/lib/docker/btrfs/subvolumes and start out as a snapshot of the
parent subvolume (if any).
This backend is pretty fast, however btrfs is still maturing and is not
considered production ready for heavy write loads. Mounting /var/lib/docker
on a different filesystem than the rest of your system is recommended in
order to limit the impact of filesystem corruption. You would also want to
mount the volume directory
/var/lib/docker/vfs/ on a standard XFS or EXT4
filesystem to ensure container data is protected.
The aufs Backend
The aufs backend uses the aufs union filesystem. This is not supported on the upstream kernel and most distributions (including any from Red Hat), and thus is not recommended as a production filesystem. It is the original backend for Docker and commonly used on Ubuntu based distributions.
The backend stores each layer as an regular directory, containing regular files and special aufs metadata. This makes up for all the files unique to that layer, as well as information about which files are removed from the previous layer. It then relies on the aufs filesystem to combine all the layers into a single mountpoint. Any changes to this mountpoint goes into the topmost layer.
Comparison of the Backends
All backends except the vfs one shares diskspace between base images. However, they work on different levels, so the behaviour is somewhat different. Both devicemapper and btrfs share data on the block level, so a single change in a file will cause just the block containing that byte being duplicated. However the aufs backend works on the file level, so any change to a file means the entire file will be copied to the top layer and then changed there. The exact behaviour here therefore depends on what kind of write behaviour an application does.
However, any kind of write-heavy load inside a container (such as databases or large logs) should generally be done to a volume. A volume is a plain directory from the host mounted into the container, which means it has none of the overhead that the storage backends may have. It also means you can easily access the data from a new container if you update the image, or if you want to access the same data from multiple concurrent containers.