Docker Volume Plugin

Introduction

Docker 1.9 has added support for creating named volumes via command-line interface and mounting them in containers as a way to share data between them. Since Docker 1.10 you can create named volumes with Docker Compose by descriptions in docker-compose.yml files for use by container groups on a single host. As of Docker 1.12 volumes are supported by Docker Swarm included with Docker Engine and created from descriptions in swarm compose v3 files for use with swarm stacks across multiple cluster nodes.

Docker Volume Plugins augment the default local volume driver included in Docker with stateful volumes shared across containers and hosts. Unlike local volumes, your data will not be deleted when such volume is removed. Plugins can run managed by the docker daemon, as a native system service (under systemd, sysv or upstart) or as a standalone executable. Rclone can run as docker volume plugin in all these modes. It interacts with the local docker daemon via plugin API and handles mounting of remote file systems into docker containers so it must run on the same host as the docker daemon or on every Swarm node.

Getting started

In the first example we will use the SFTP rclone volume with Docker engine on a standalone Ubuntu machine.

Start from installing Docker on the host.

The FUSE driver is a prerequisite for rclone mounting and should be installed on host:

sudo apt-get -y install fuse

Create two directories required by rclone docker plugin:

sudo mkdir -p /var/lib/docker-plugins/rclone/config
sudo mkdir -p /var/lib/docker-plugins/rclone/cache

Install the managed rclone docker plugin for your architecture (here amd64):

docker plugin install rclone/docker-volume-rclone:amd64 args="-v" --alias rclone --grant-all-permissions
docker plugin list

Create your SFTP volume:

docker volume create firstvolume -d rclone -o type=sftp -o sftp-host=_hostname_ -o sftp-user=_username_ -o sftp-pass=_password_ -o allow-other=true

Note that since all options are static, you don't even have to run rclone config or create the rclone.conf file (but the config directory should still be present). In the simplest case you can use localhost as hostname and your SSH credentials as username and password. You can also change the remote path to your home directory on the host, for example -o path=/home/username.

Time to create a test container and mount the volume into it:

docker run --rm -it -v firstvolume:/mnt --workdir /mnt ubuntu:latest bash

If all goes well, you will enter the new container and change right to the mounted SFTP remote. You can type ls to list the mounted directory or otherwise play with it. Type exit when you are done. The container will stop but the volume will stay, ready to be reused. When it's not needed anymore, remove it:

docker volume list
docker volume remove firstvolume

Now let us try something more elaborate: Google Drive volume on multi-node Docker Swarm.

You should start from installing Docker and FUSE, creating plugin directories and installing rclone plugin on every swarm node. Then setup the Swarm.

Google Drive volumes need an access token which can be setup via web browser and will be periodically renewed by rclone. The managed plugin cannot run a browser so we will use a technique similar to the rclone setup on a headless box.

Run rclone config on another machine equipped with web browser and graphical user interface. Create the Google Drive remote. When done, transfer the resulting rclone.conf to the Swarm cluster and save as /var/lib/docker-plugins/rclone/config/rclone.conf on every node. By default this location is accessible only to the root user so you will need appropriate privileges. The resulting config will look like this:

[gdrive]
type = drive
scope = drive
drive_id = 1234567...
root_folder_id = 0Abcd...
token = {"access_token":...}

Now create the file named example.yml with a swarm stack description like this:

version: '3'
services:
  heimdall:
    image: linuxserver/heimdall:latest
    ports: [8080:80]
    volumes: [configdata:/config]
volumes:
  configdata:
    driver: rclone
    driver_opts:
      remote: 'gdrive:heimdall'
      allow_other: 'true'
      vfs_cache_mode: full
      poll_interval: 0

and run the stack:

docker stack deploy example -c ./example.yml

After a few seconds docker will spread the parsed stack description over cluster, create the example_heimdall service on port 8080, run service containers on one or more cluster nodes and request the example_configdata volume from rclone plugins on the node hosts. You can use the following commands to confirm results:

docker service ls
docker service ps example_heimdall
docker volume ls

Point your browser to http://cluster.host.address:8080 and play with the service. Stop it with docker stack remove example when you are done. Note that the example_configdata volume(s) created on demand at the cluster nodes will not be automatically removed together with the stack but stay for future reuse. You can remove them manually by invoking the docker volume remove example_configdata command on every node.

Creating Volumes via CLI

Volumes can be created with docker volume create. Here are a few examples:

docker volume create vol1 -d rclone -o remote=storj: -o vfs-cache-mode=full
docker volume create vol2 -d rclone -o remote=:tardigrade,access_grant=xxx:heimdall
docker volume create vol3 -d rclone -o type=tardigrade -o path=heimdall -o tardigrade-access-grant=xxx -o poll-interval=0

Note the -d rclone flag that tells docker to request volume from the rclone driver. This works even if you installed managed driver by its full name rclone/docker-volume-rclone because you provided the --alias rclone option.

Volumes can be inspected as follows:

docker volume list
docker volume inspect vol1

Volume Configuration

Rclone flags and volume options are set via the -o flag to the docker volume create command. They include backend-specific parameters as well as mount and VFS options. Also there are a few special -o options: remote, fs, type, path, mount-type and persist.

remote determines an existing remote name from the config file, with trailing colon and optionally with a remote path. See the full syntax in the rclone documentation. This option can be aliased as fs to prevent confusion with the remote parameter of such backends as crypt or alias.

The remote=:backend:dir/subdir syntax can be used to create on-the-fly (config-less) remotes, while the type and path options provide a simpler alternative for this. Using two split options

-o type=backend -o path=dir/subdir

is equivalent to the combined syntax

-o remote=:backend:dir/subdir

but is arguably easier to parameterize in scripts. The path part is optional.

Mount and VFS options as well as backend parameters are named like their twin command-line flags without the -- CLI prefix. Optionally you can use underscores instead of dashes in option names. For example, --vfs-cache-mode full becomes -o vfs-cache-mode=full or -o vfs_cache_mode=full. Boolean CLI flags without value will gain the true value, e.g. --allow-other becomes -o allow-other=true or -o allow_other=true.

Please note that you can provide parameters only for the backend immediately referenced by the backend type of mounted remote. If this is a wrapping backend like alias, chunker or crypt, you cannot provide options for the referred to remote or backend. This limitation is imposed by the rclone connection string parser. The only workaround is to feed plugin with rclone.conf or configure plugin arguments (see below).

Special Volume Options

mount-type determines the mount method and in general can be one of: mount, cmount, or mount2. This can be aliased as mount_type. It should be noted that the managed rclone docker plugin currently does not support the cmount method and mount2 is rarely needed. This option defaults to the first found method, which is usually mount so you generally won't need it.

persist is a reserved boolean (true/false) option. In future it will allow to persist on-the-fly remotes in the plugin rclone.conf file.

Connection Strings

The remote value can be extended with connection strings as an alternative way to supply backend parameters. This is equivalent to the -o backend options with one syntactic difference. Inside connection string the backend prefix must be dropped from parameter names but in the -o param=value array it must be present. For instance, compare the following option array

-o remote=:sftp:/home -o sftp-host=localhost

with equivalent connection string:

-o remote=:sftp,host=localhost:/home

This difference exists because flag options -o key=val include not only backend parameters but also mount/VFS flags and possibly other settings. Also it allows to discriminate the remote option from the crypt-remote (or similarly named backend parameters) and arguably simplifies scripting due to clearer value substitution.

Using with Swarm or Compose

Both Docker Swarm and Docker Compose use YAML-formatted text files to describe groups (stacks) of containers, their properties, networks and volumes. Compose uses the compose v2 format, Swarm uses the compose v3 format. They are mostly similar, differences are explained in the docker documentation.

Volumes are described by the children of the top-level volumes: node. Each of them should be named after its volume and have at least two elements, the self-explanatory driver: rclone value and the driver_opts: structure playing the same role as -o key=val CLI flags:

volumes:
  volume_name_1:
    driver: rclone
    driver_opts:
      remote: 'gdrive:'
      allow_other: 'true'
      vfs_cache_mode: full
      token: '{"type": "borrower", "expires": "2021-12-31"}'
      poll_interval: 0

Notice a few important details:

  • YAML prefers _ in option names instead of -.
  • YAML treats single and double quotes interchangeably. Simple strings and integers can be left unquoted.
  • Boolean values must be quoted like 'true' or "false" because these two words are reserved by YAML.
  • The filesystem string is keyed with remote (or with fs). Normally you can omit quotes here, but if the string ends with colon, you must quote it like remote: "storage_box:".
  • YAML is picky about surrounding braces in values as this is in fact another syntax for key/value mappings. For example, JSON access tokens usually contain double quotes and surrounding braces, so you must put them in single quotes.

Installing as Managed Plugin

Docker daemon can install plugins from an image registry and run them managed. We maintain the docker-volume-rclone plugin image on Docker Hub.

Rclone volume plugin requires Docker Engine >= 19.03.15

The plugin requires presence of two directories on the host before it can be installed. Note that plugin will not create them automatically. By default they must exist on host at the following locations (though you can tweak the paths):

  • /var/lib/docker-plugins/rclone/config is reserved for the rclone.conf config file and must exist even if it's empty and the config file is not present.
  • /var/lib/docker-plugins/rclone/cache holds the plugin state file as well as optional VFS caches.

You can install managed plugin with default settings as follows:

docker plugin install rclone/docker-volume-rclone:amd64 --grant-all-permissions --alias rclone

The :amd64 part of the image specification after colon is called a tag. Usually you will want to install the latest plugin for your architecture. In this case the tag will just name it, like amd64 above. The following plugin architectures are currently available:

  • amd64
  • arm64
  • arm-v7

Sometimes you might want a concrete plugin version, not the latest one. Then you should use image tag in the form :ARCHITECTURE-VERSION. For example, to install plugin version v1.56.2 on architecture arm64 you will use tag arm64-1.56.2 (note the removed v) so the full image specification becomes rclone/docker-volume-rclone:arm64-1.56.2.

We also provide the latest plugin tag, but since docker does not support multi-architecture plugins as of the time of this writing, this tag is currently an alias for amd64. By convention the latest tag is the default one and can be omitted, thus both rclone/docker-volume-rclone:latest and just rclone/docker-volume-rclone will refer to the latest plugin release for the amd64 platform.

Also the amd64 part can be omitted from the versioned rclone plugin tags. For example, rclone image reference rclone/docker-volume-rclone:amd64-1.56.2 can be abbreviated as rclone/docker-volume-rclone:1.56.2 for convenience. However, for non-intel architectures you still have to use the full tag as amd64 or latest will fail to start.

Managed plugin is in fact a special container running in a namespace separate from normal docker containers. Inside it runs the rclone serve docker command. The config and cache directories are bind-mounted into the container at start. The docker daemon connects to a unix socket created by the command inside the container. The command creates on-demand remote mounts right inside, then docker machinery propagates them through kernel mount namespaces and bind-mounts into requesting user containers.

You can tweak a few plugin settings after installation when it's disabled (not in use), for instance:

docker plugin disable rclone
docker plugin set rclone RCLONE_VERBOSE=2 config=/etc/rclone args="--vfs-cache-mode=writes --allow-other"
docker plugin enable rclone
docker plugin inspect rclone

Note that if docker refuses to disable the plugin, you should find and remove all active volumes connected with it as well as containers and swarm services that use them. This is rather tedious so please carefully plan in advance.

You can tweak the following settings: args, config, cache, HTTP_PROXY, HTTPS_PROXY, NO_PROXY and RCLONE_VERBOSE. It's your task to keep plugin settings in sync across swarm cluster nodes.

args sets command-line arguments for the rclone serve docker command (none by default). Arguments should be separated by space so you will normally want to put them in quotes on the docker plugin set command line. Both serve docker flags and generic rclone flags are supported, including backend parameters that will be used as defaults for volume creation. Note that plugin will fail (due to this docker bug) if the args value is empty. Use e.g. args="-v" as a workaround.

config=/host/dir sets alternative host location for the config directory. Plugin will look for rclone.conf here. It's not an error if the config file is not present but the directory must exist. Please note that plugin can periodically rewrite the config file, for example when it renews storage access tokens. Keep this in mind and try to avoid races between the plugin and other instances of rclone on the host that might try to change the config simultaneously resulting in corrupted rclone.conf. You can also put stuff like private key files for SFTP remotes in this directory. Just note that it's bind-mounted inside the plugin container at the predefined path /data/config. For example, if your key file is named sftp-box1.key on the host, the corresponding volume config option should read -o sftp-key-file=/data/config/sftp-box1.key.

cache=/host/dir sets alternative host location for the cache directory. The plugin will keep VFS caches here. Also it will create and maintain the docker-plugin.state file in this directory. When the plugin is restarted or reinstalled, it will look in this file to recreate any volumes that existed previously. However, they will not be re-mounted into consuming containers after restart. Usually this is not a problem as the docker daemon normally will restart affected user containers after failures, daemon restarts or host reboots.

RCLONE_VERBOSE sets plugin verbosity from 0 (errors only, by default) to 2 (debugging). Verbosity can be also tweaked via args="-v [-v] ...". Since arguments are more generic, you will rarely need this setting. The plugin output by default feeds the docker daemon log on local host. Log entries are reflected as errors in the docker log but retain their actual level assigned by rclone in the encapsulated message string.

HTTP_PROXY, HTTPS_PROXY, NO_PROXY customize the plugin proxy settings.

You can set custom plugin options right when you install it, in one go:

docker plugin remove rclone
docker plugin install rclone/docker-volume-rclone:amd64 \
       --alias rclone --grant-all-permissions \
       args="-v --allow-other" config=/etc/rclone
docker plugin inspect rclone

Healthchecks

The docker plugin volume protocol doesn't provide a way for plugins to inform the docker daemon that a volume is (un-)available. As a workaround you can setup a healthcheck to verify that the mount is responding, for example:

services:
  my_service:
    image: my_image
    healthcheck:
      test: ls /path/to/rclone/mount || exit 1
      interval: 1m
      timeout: 15s
      retries: 3
      start_period: 15s

Running Plugin under Systemd

In most cases you should prefer managed mode. Moreover, MacOS and Windows do not support native Docker plugins. Please use managed mode on these systems. Proceed further only if you are on Linux.

First, install rclone. You can just run it (type rclone serve docker and hit enter) for the test.

Install FUSE:

sudo apt-get -y install fuse

Download two systemd configuration files: docker-volume-rclone.service and docker-volume-rclone.socket.

Put them to the /etc/systemd/system/ directory:

cp docker-volume-plugin.service /etc/systemd/system/
cp docker-volume-plugin.socket  /etc/systemd/system/

Please note that all commands in this section must be run as root but we omit sudo prefix for brevity. Now create directories required by the service:

mkdir -p /var/lib/docker-volumes/rclone
mkdir -p /var/lib/docker-plugins/rclone/config
mkdir -p /var/lib/docker-plugins/rclone/cache

Run the docker plugin service in the socket activated mode:

systemctl daemon-reload
systemctl start docker-volume-rclone.service
systemctl enable docker-volume-rclone.socket
systemctl start docker-volume-rclone.socket
systemctl restart docker

Or run the service directly:

  • run systemctl daemon-reload to let systemd pick up new config
  • run systemctl enable docker-volume-rclone.service to make the new service start automatically when you power on your machine.
  • run systemctl start docker-volume-rclone.service to start the service now.
  • run systemctl restart docker to restart docker daemon and let it detect the new plugin socket. Note that this step is not needed in managed mode where docker knows about plugin state changes.

The two methods are equivalent from the user perspective, but I personally prefer socket activation.

Troubleshooting

You can see managed plugin settings with

docker plugin list
docker plugin inspect rclone

Note that docker (including latest 20.10.7) will not show actual values of args, just the defaults.

Use journalctl --unit docker to see managed plugin output as part of the docker daemon log. Note that docker reflects plugin lines as errors but their actual level can be seen from encapsulated message string.

You will usually install the latest version of managed plugin for your platform. Use the following commands to print the actual installed version:

PLUGID=$(docker plugin list --no-trunc | awk '/rclone/{print$1}')
sudo runc --root /run/docker/runtime-runc/plugins.moby exec $PLUGID rclone version

You can even use runc to run shell inside the plugin container:

sudo runc --root /run/docker/runtime-runc/plugins.moby exec --tty $PLUGID bash

Also you can use curl to check the plugin socket connectivity:

docker plugin list --no-trunc
PLUGID=123abc...
sudo curl -H Content-Type:application/json -XPOST -d {} --unix-socket /run/docker/plugins/$PLUGID/rclone.sock http://localhost/Plugin.Activate

though this is rarely needed.

Caveats

Finally I'd like to mention a caveat with updating volume settings. Docker CLI does not have a dedicated command like docker volume update. It may be tempting to invoke docker volume create with updated options on existing volume, but there is a gotcha. The command will do nothing, it won't even return an error. I hope that docker maintainers will fix this some day. In the meantime be aware that you must remove your volume before recreating it with new settings:

docker volume remove my_vol
docker volume create my_vol -d rclone -o opt1=new_val1 ...

and verify that settings did update:

docker volume list
docker volume inspect my_vol

If docker refuses to remove the volume, you should find containers or swarm services that use it and stop them first.