280
votes

Original question: How to use the VOLUME instruction in Dockerfile?

The actual question I want to solve is -- how to mount host volumes into docker containers in Dockerfile during build, i.e., having the docker run -v /export:/export capability during docker build.

The reason behind it, for me, is when building things in Docker, I don't want those (apt-get install) caches locked in a single docker, but to share/reuse them. That's the main reason I'm asking about this question.

Latest Update:

Before docker v18.09, the correct answer should be the one that starts with:

There is a way to mount a volume during a build, but it doesn't involve Dockerfiles.

However, that was a poorly stated, organized and supported answer. When I was reinstalling my docker contains, I happened to stumble upon the following article:

Dockerize an apt-cacher-ng service
https://docs.docker.com/engine/examples/apt-cacher-ng/

That's the docker's solution to this/my question, not directly but indirectly. It's the orthodox way docker suggests us to do. And I admit it is better than the one I was trying to ask here.

Another way is, the newly accepted answer, e.g., the Buildkit in v18.09.

Pick whichever suits you.


Was: There had been a solution -- rocker, which was not from Docker, but now that rocker is discontinued, I revert the answer back to "Not possible" again.


Old Update: So the answer is "Not possible". I can accept it as an answer as I know the issue has been extensively discussed at https://github.com/docker/docker/issues/3156. I can understand that portability is a paramount issue for docker developer; but as a docker user, I have to say I'm very disappointed about this missing feature. Let me close my argument with a quote from aforementioned discussion: "I would like to use Gentoo as a base image but definitely don't want > 1GB of Portage tree data to be in any of the layers once the image has been built. You could have some nice a compact containers if it wasn't for the gigantic portage tree having to appear in the image during the install." Yes, I can use wget or curl to download whatever I need, but the fact that merely a portability consideration is now forcing me to download > 1GB of Portage tree each time I build a Gentoo base image is neither efficient nor user friendly. Further more, the package repository WILL ALWAYS be under /usr/portage, thus ALWAYS PORTABLE under Gentoo. Again, I respect the decision, but please allow me expressing my disappointment as well in the mean time. Thanks.


Original question in details:

From

Share Directories via Volumes
http://docker.readthedocs.org/en/v0.7.3/use/working_with_volumes/

it says that Data volumes feature "have been available since version 1 of the Docker Remote API". My docker is of version 1.2.0, but I found the example given in above article not working:

# BUILD-USING:        docker build -t data .
# RUN-USING:          docker run -name DATA data
FROM          busybox
VOLUME        ["/var/volume1", "/var/volume2"]
CMD           ["/usr/bin/true"]

What's the proper way in Dockerfile to mount host-mounted volumes into docker containers, via the VOLUME command?

$ apt-cache policy lxc-docker
lxc-docker:
  Installed: 1.2.0
  Candidate: 1.2.0
  Version table:
 *** 1.2.0 0
        500 https://get.docker.io/ubuntu/ docker/main amd64 Packages
        100 /var/lib/dpkg/status

$ cat Dockerfile 
FROM          debian:sid

VOLUME        ["/export"]
RUN ls -l /export
CMD ls -l /export

$ docker build -t data .
Sending build context to Docker daemon  2.56 kB
Sending build context to Docker daemon 
Step 0 : FROM          debian:sid
 ---> 77e97a48ce6a
Step 1 : VOLUME        ["/export"]
 ---> Using cache
 ---> 59b69b65a074
Step 2 : RUN ls -l /export
 ---> Running in df43c78d74be
total 0
 ---> 9d29a6eb263f
Removing intermediate container df43c78d74be
Step 3 : CMD ls -l /export
 ---> Running in 8e4916d3e390
 ---> d6e7e1c52551
Removing intermediate container 8e4916d3e390
Successfully built d6e7e1c52551

$ docker run data
total 0

$ ls -l /export | wc 
     20     162    1131

$ docker -v
Docker version 1.2.0, build fa7b24f
10
Apparently more current feature request (not that I expect it to be implemented, but just in case): docker/docker #14080Jesse Glick
indeed there is an extensive discussion that it should not be allowed to link a host directory and container directory during build i.e. something like VOLUME ~/host_dir ~/container_dir. The discussion is quite extensive, si there a short way to summarize what is the reason?Charlie Parker
Unfortunately, the link docker.readthedocs.org/en/v0.7.3/use/working_with_volumes is broken. Appreciate the thorough Q&A though.ramhiser
Thanks for the notice @ramhiser, the underline link to docker.readthedocs.org/en/v0.7.3/use/working_with_volumes is now updated, and clickable.xpt

10 Answers

62
votes

First, to answer "why doesn't VOLUME work?" When you define a VOLUME in the Dockerfile, you can only define the target, not the source of the volume. During the build, you will only get an anonymous volume from this. That anonymous volume will be mounted at every RUN command, prepopulated with the contents of the image, and then discarded at the end of the RUN command. Only changes to the container are saved, not changes to the volume.


Since this question has been asked, a few features have been released that may help. First is multistage builds allowing you to build a disk space inefficient first stage, and copy just the needed output to the final stage that you ship. And the second feature is Buildkit which is dramatically changing how images are built and new capabilities are being added to the build.

For a multi-stage build, you would have multiple FROM lines, each one starting the creation of a separate image. Only the last image is tagged by default, but you can copy files from previous stages. The standard use is to have a compiler environment to build a binary or other application artifact, and a runtime environment as the second stage that copies over that artifact. You could have:

FROM debian:sid as builder
COPY export /export
RUN compile command here >/result.bin

FROM debian:sid
COPY --from=builder /result.bin /result.bin
CMD ["/result.bin"]

That would result in a build that only contains the resulting binary, and not the full /export directory.


Buildkit is coming out of experimental in 18.09. It's a complete redesign of the build process, including the ability to change the frontend parser. One of those parser changes has has implemented the RUN --mount option which lets you mount a cache directory for your run commands. E.g. here's one that mounts some of the debian directories (with a reconfigure of the debian image, this could speed up reinstalls of packages):

# syntax = docker/dockerfile:experimental
FROM debian:latest
RUN --mount=target=/var/lib/apt/lists,type=cache \
    --mount=target=/var/cache/apt,type=cache \
    apt-get update \
 && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
      git

You would adjust the cache directory for whatever application cache you have, e.g. $HOME/.m2 for maven, or /root/.cache for golang.


TL;DR: Answer is here: With that RUN --mount syntax, you can also bind mount read-only directories from the build-context. The folder must exist in the build context, and it is not mapped back to the host or the build client:

# syntax = docker/dockerfile:experimental
FROM debian:latest
RUN --mount=target=/export,type=bind,source=export \
    process export directory here...

Note that because the directory is mounted from the context, it's also mounted read-only, and you cannot push changes back to the host or client. When you build, you'll want an 18.09 or newer install and enable buildkit with export DOCKER_BUILDKIT=1.

If you get an error that the mount flag isn't supported, that indicates that you either didn't enable buildkit with the above variable, or that you didn't enable the experimental syntax with the syntax line at the top of the Dockerfile before any other lines, including comments. Note that the variable to toggle buildkit will only work if your docker install has buildkit support built in, which requires version 18.09 or newer from Docker, both on the client and server.

123
votes

It is not possible to use the VOLUME instruction to tell docker what to mount. That would seriously break portability. This instruction tells docker that content in those directories does not go in images and can be accessed from other containers using the --volumes-from command line parameter. You have to run the container using -v /path/on/host:/path/in/container to access directories from the host.

Mounting host volumes during build is not possible. There is no privileged build and mounting the host would also seriously degrade portability. You might want to try using wget or curl to download whatever you need for the build and put it in place.

64
votes

UPDATE: Somebody just won't take no as the answer, and I like it, very much, especially to this particular question.

GOOD NEWS, There is a way now --

The solution is Rocker: https://github.com/grammarly/rocker

John Yani said, "IMO, it solves all the weak points of Dockerfile, making it suitable for development."

Rocker

https://github.com/grammarly/rocker

By introducing new commands, Rocker aims to solve the following use cases, which are painful with plain Docker:

  1. Mount reusable volumes on build stage, so dependency management tools may use cache between builds.
  2. Share ssh keys with build (for pulling private repos, etc.), while not leaving them in the resulting image.
  3. Build and run application in different images, be able to easily pass an artifact from one image to another, ideally have this logic in a single Dockerfile.
  4. Tag/Push images right from Dockerfiles.
  5. Pass variables from shell build command so they can be substituted to a Dockerfile.

And more. These are the most critical issues that were blocking our adoption of Docker at Grammarly.

Update: Rocker has been discontinued, per the official project repo on Github

As of early 2018, the container ecosystem is much more mature than it was three years ago when this project was initiated. Now, some of the critical and outstanding features of rocker can be easily covered by docker build or other well-supported tools, though some features do remain unique to rocker. See https://github.com/grammarly/rocker/issues/199 for more details.

15
votes

There is a way to mount a volume during a build, but it doesn't involve Dockerfiles.

The technique would be to create a container from whatever base you wanted to use (mounting your volume(s) in the container with the -v option), run a shell script to do your image building work, then commit the container as an image when done.

Not only will this leave out the excess files you don't want (this is good for secure files as well, like SSH files), it also creates a single image. It has downsides: the commit command doesn't support all of the Dockerfile instructions, and it doesn't let you pick up when you left off if you need to edit your build script.

UPDATE:

For example,

CONTAINER_ID=$(docker run -dit ubuntu:16.04)
docker cp build.sh $CONTAINER_ID:/build.sh
docker exec -t $CONTAINER_ID /bin/sh -c '/bin/sh /build.sh'
docker commit $CONTAINER_ID $REPO:$TAG
docker stop $CONTAINER_ID
8
votes

As you run the container, a directory on your host is created and mounted into the container. You can find out what directory this is with

$ docker inspect --format "{{ .Volumes }}" <ID>
map[/export:/var/lib/docker/vfs/dir/<VOLUME ID...>]

If you want to mount a directory from your host inside your container, you have to use the -v parameter and specify the directory. In your case this would be:

docker run -v /export:/export data

SO you would use the hosts folder inside your container.

5
votes

It's ugly, but I achieved a semblance of this like so:

Dockerfile:

FROM foo
COPY ./m2/ /root/.m2
RUN stuff

imageBuild.sh:

docker build . -t barImage
container="$(docker run -d barImage)"
rm -rf ./m2
docker cp "$container:/root/.m2" ./m2
docker rm -f "$container"

I have a java build that downloads the universe into /root/.m2, and did so every single time. imageBuild.sh copies the contents of that folder onto the host after the build, and Dockerfile copies them back into the image for the next build.

This is something like how a volume would work (i.e. it persists between builds).

4
votes

I think you can do what you want to do by running the build via a docker command which itself is run inside a docker container. See Docker can now run within Docker | Docker Blog. A technique like this, but which actually accessed the outer docker from with a container, was used, e.g., while exploring how to Create the smallest possible Docker container | Xebia Blog.

Another relevant article is Optimizing Docker Images | CenturyLink Labs, which explains that if you do end up downloading stuff during a build, you can avoid having space wasted by it in the final image by downloading, building and deleting the download all in one RUN step.

3
votes

As many have already answered, mounting host volumes during the build is not possible. I just would like to add docker-compose way, I think it'll be nice to have, mostly for development/testing usage

Dockerfile

FROM node:10
WORKDIR /app
COPY . .
RUN npm ci
CMD sleep 999999999

docker-compose.yml

version: '3'
services:
  test-service:
    image: test/image
    build:
      context: .
      dockerfile: Dockerfile
    container_name: test
    volumes:
      - ./export:/app/export
      - ./build:/app/build

And run your container by docker-compose up -d --build

2
votes

Here is a simplified version of the 2-step approach using build and commit, without shell scripts. It involves:

  1. Building the image partially, without volumes
  2. Running a container with volumes, making changes, then committing the result, replacing the original image name.

With relatively minor changes the additional step adds only a few seconds to the build time.

Basically:

docker build -t image-name . # your normal docker build

# Now run a command in a throwaway container that uses volumes and makes changes:
docker run -v /some:/volume --name temp-container image-name /some/post-configure/command

# Replace the original image with the result:
# (reverting CMD to whatever it was, otherwise it will be set to /some/post-configure/command)   
docker commit --change="CMD bash" temp-container image-name 

# Delete the temporary container:
docker rm temp-container

In my use case I want to pre-generate a maven toolchains.xml file, but my many JDK installations are on a volume that isn't available until runtime. Some of my images are not compatible with all the JDKS, so I need to test compatibility at build time and populate toolchains.xml conditionally. Note that I don't need the image to be portable, I'm not publishing it to Docker Hub.

0
votes

If you are looking for a way to "mount" files, like -v for docker run, you can now use the --secret flag for docker build

echo 'WARMACHINEROX' > mysecret.txt
docker build --secret id=mysecret,src=mysecret.txt .

And inside your Dockerfile you can now access this secret

# syntax = docker/dockerfile:1.0-experimental
FROM alpine

# shows secret from default secret location:
RUN --mount=type=secret,id=mysecret cat /run/secrets/mysecret

# shows secret from custom secret location:
RUN --mount=type=secret,id=mysecret,dst=/foobar cat /foobar

More in-depth information about --secret available on Docker Docs