Docker and securing passwords

I've been experimenting with Docker recently on building some services to play around with and one thing that keeps nagging me has been putting passwords in a Dockerfile. I'm a developer so storing passwords in source feels like a punch in the face. Should this even be a concern? Are there any good conventions on how to handle passwords in Dockerfiles?


Definitely it is a concern. Dockerfiles are commonly checked in to repositories and shared with other people. An alternative is to provide any credentials (usernames, passwords, tokens, anything sensitive) as environment variables at runtime. This is possible via the -e argument (for individual vars on the CLI) or --env-file argument (for multiple variables in a file) to docker run. Read this for using environmental with docker-compose.

Using --env-file is definitely a safer option since this protects against the secrets showing up in ps or in logs if one uses set -x.

However, env vars are not particularly secure either. They are visible via docker inspect, and hence they are available to any user that can run docker commands. (Of course, any user that has access to docker on the host also has root anyway.)

My preferred pattern is to use a wrapper script as the ENTRYPOINT or CMD. The wrapper script can first import secrets from an outside location in to the container at run time, then execute the application, providing the secrets. The exact mechanics of this vary based on your run time environment. In AWS, you can use a combination of IAM roles, the Key Management Service, and S3 to store encrypted secrets in an S3 bucket. Something like HashiCorp Vault or credstash is another option.

AFAIK there is no optimal pattern for using sensitive data as part of the build process. In fact, I have an SO question on this topic. You can use docker-squash to remove layers from an image. But there's no native functionality in Docker for this purpose.

You may find shykes comments on config in containers useful.

Our team avoids putting credentials in repositories, so that means they're not allowed in Dockerfile. Our best practice within applications is to use creds from environment variables.

We solve for this using docker-compose.

Within docker-compose.yml, you can specify a file that contains the environment variables for the container:

- .env

Make sure to add .env to .gitignore, then set the credentials within the .env file like:


Store the .env file locally or in a secure location where the rest of the team can grab it.


Docker now (version 1.13 or 17.06 and higher) has support for managing secret information. Here's an overview and more detailed documentation

Similar feature exists in kubernetes and DCOS

You should never add credentials to a container unless you're OK broadcasting the creds to whomever can download the image. In particular, doing and ADD creds and later RUN rm creds is not secure because the creds file remains in the final image in an intermediate filesystem layer. It's easy for anyone with access to the image to extract it.

The typical solution I've seen when you need creds to checkout dependencies and such is to use one container to build another. I.e., typically you have some build environment in your base container and you need to invoke that to build your app container. So the simple solution is to add your app source and then RUN the build commands. This is insecure if you need creds in that RUN. Instead what you do is put your source into a local directory, run (as in docker run) the container to perform the build step with the local source directory mounted as volume and the creds either injected or mounted as another volume. Once the build step is complete you build your final container by simply ADDing the local source directory which now contains the built artifacts.

I'm hoping Docker adds some features to simplify all this!

Update: looks like the method going forward will be to have nested builds. In short, the dockerfile would describe a first container that is used to build the run-time environment and then a second nested container build that can assemble all the pieces into the final container. This way the build-time stuff isn't in the second container. This of a Java app where you need the JDK for building the app but only the JRE for running it. There are a number of proposals being discussed, best to start from and follow some of the links for alternate proposals.

With Docker v1.9 you can use the ARG instruction to fetch arguments passed by command line to the image on build action. Simply use the --build-arg flag. So you can avoid to keep explicit password (or other sensible information) on the Dockerfile and pass them on the fly.




FROM busybox
ARG user
RUN echo "user is $user"

build image command

docker build --build-arg user=capuccino -t test_arguments -f path/to/dockerfile .

during the build it print

$ docker build --build-arg user=capuccino -t test_arguments -f ./test_args.Dockerfile .

Sending build context to Docker daemon 2.048 kB
Step 1 : FROM busybox
 ---> c51f86c28340
Step 2 : ARG user
 ---> Running in 43a4aa0e421d
 ---> f0359070fc8f
Removing intermediate container 43a4aa0e421d
Step 3 : RUN echo "user is $user"
 ---> Running in 4360fb10d46a
**user is capuccino**
 ---> 1408147c1cb9
Removing intermediate container 4360fb10d46a
Successfully built 1408147c1cb9

Hope it helps! Bye.

An alternative to using environment variables, which can get messy if you have a lot of them, is to use volumes to make a directory on the host accessible in the container.

If you put all your credentials as files in that folder, then the container can read the files and use them as it pleases.

For example:

$ echo "secret" > /root/configs/password.txt
$ docker run -v /root/configs:/cfg ...

In the Docker container:

# echo Password is `cat /cfg/password.txt`
Password is secret

Many programs can read their credentials from a separate file, so this way you can just point the program to one of the files.

My approach seems to work, but is probably naive. Tell me why it is wrong.

ARGs set during docker build are exposed by the history subcommand, so no go there. However, when running a container, environment variables given in the run command are available to the container, but are not part of the image.

So, in the Dockerfile, do setup that does not involve secret data. Set a CMD of something like /root/ In the run command, use environmental variables to send secret data into the container. uses the variables essentially to finish build tasks.

To make managing the secret data easier, put it into a file that is loaded by docker run with the --env-file switch. Of course, keep the file secret. .gitignore and such.

For me, runs a Python program. It checks to make sure it hasn't run before, then finishes the setup (e.g., copies the database name into Django's

run-time only solution

docker-compose also provides a non-swarm mode solution (since v1.11: Secrets using bind mounts).

The secrets are mounted as files below /run/secrets/ by docker-compose. This solves the problem at run-time (running the container), but not at build-time (building the image), because /run/secrets/ is not mounted at build-time. Furthermore this behavior depends on running the container with docker-compose.



FROM alpine
RUN cat /run/secrets/password
CMD sleep inifinity


version: '3.1'
    build: .
      - password

    file: password.txt

To build, execute:

docker-compose up -d

Further reading:

While I totally agree there is no simple solution. There continues to be a single point of failure. Either the dockerfile, etcd, and so on. Apcera has a plan that looks like sidekick - dual authentication. In other words two container cannot talk unless there is a Apcera configuration rule. In their demo the uid/pwd was in the clear and could not be reused until the admin configured the linkage. For this to work, however, it probably meant patching Docker or at least the network plugin (if there is such a thing).

There is a new docker command[1] for "secrets" management, but that only works for swarm clusters.

docker service create
--name my-iis
--publish target=8000,port=8000
--secret src=homepage,target="\inetpub\wwwroot\index.html"


The 12-Factor app methodology tells, that any configuration should be stored in environment variables.

Docker compose could do variable substitution in configuration, so that could be used to pass passwords from host to docker.

Need Your Help

What mode for MySQL WEEK() complies with ISO 8601

mysql iso8601

What mode for MySQL's WEEK() function yields the ISO 8601 week of the year? Argument 2 of WEEK() sets the mode according to this chart:

How to build sources jar with gradle

java jar gradle

I am working with an open source project that is built with gradle. I would like to generate a (project)-sources.jar file that I can load into my IDE (IntelliJ IDEA) and debug through the project....