gitlab-runner passes wrong arguments to custom image

Question

I have written a small CLI using Java, Argparse4j, and packaged it in docker using this Dockerfile:

FROM openjdk:18

ENV JAR_NAME "my-jar-with-dependencies.jar"

ENV PROJECT_HOME /opt/app
RUN mkdir -p $PROJECT_HOME

WORKDIR $PROJECT_HOME
COPY run.sh $PROJECT_HOME/run.sh
RUN chmod +x $PROJECT_HOME/run.sh

COPY target/$JAR_NAME $PROJECT_HOME/cli.jar

ENTRYPOINT ["./run.sh"]

The last line of the Dockerfile then invokes a simple bash script:

#!/bin/bash

java $JAVA_OPTS -Djava.security.egd=file:/dev/./urandom -jar ./cli.jar "$@"

The CLI I wrote has three main actions: upload, download and apply. Therefore argparse4j expects one of these actions to be passed as the first parameter, i.e.

java -jar cli.jar download #... whatever other argument

This works just fine when running the docker image locally, but completely fails when running in the CI pipeline:

download:
  stage: download
  image: <url>/my-image:<tag>
  variables:
    URL: <URL>
    API_KEY: <API_KEY>
    CI_DEBUG_TRACE: "true"
  script:
    - download -f zip -u true test-download.zip

This is the error that is returned:

Executing "step_script" stage of the job script 00:01
Using docker image sha256:<sha> for <url>/my-image:<tag> with digest <url>/my-image:<tag>@sha256:<sha> ...
SLF4J: Failed to load class "org.slf4j.impl.StaticLoggerBinder".
SLF4J: Defaulting to no-operation (NOP) logger implementation
SLF4J: See http://www.slf4j.org/codes.html#StaticLoggerBinder for further details.
usage: tool [-h] ACTION ...
tool: error:  invalid  choice:  'sh'  (choose  from  'upload',  'download',
'apply')

I have tried following the suggestion in gitlab-runner doesn't run ENTRYPOINT scripts in Dockerfile but I can't seem to get the CI part to work correctly.

I would like to avoid using the entrypoint directive as it needs to be used on multiple files, so I rather fix the issue at the root.

Does anyone have an idea of what is happening or how to fix it?

Answer 1

I would like to avoid using the entrypoint directive as it needs to be used on multiple files, so I rather fix the issue at the root.

You can change your Dockerfile instead to keep default ENTRYPOINT (as openjdk:18 doesn't define any entrypoint, it will be empty):

FROM openjdk:18

# ...

# ENTRYPOINT ["./run.sh"] # remove this

# Add run.sh to path to be able to use `run.sh` from any directory
ENV PATH="${PATH}:/opt/app"

And update your run.sh to specify full path to jar:

#!/bin/bash

java $JAVA_OPTS -Djava.security.egd=file:/dev/./urandom -jar $PROJECT_HOME/cli.jar "$@"

Now your container will start in Gitlab without having to specify entrypoint keyword for job. You can then setup something like this:

download:
  stage: download
  image: <url>/my-image:<tag>
  variables:
    URL: <URL>
    API_KEY: <API_KEY>
    CI_DEBUG_TRACE: "true"
  script:
  # Specify command using run.sh
  # This command is run from within your container
  # Note that script is not an argument passed on your image startup
  # But independent commands run within your container using shell
  - run.sh download -f zip -u true test-download.zip

Notes:

• Gitlab won't run your script in Dockerfile's WORKDIR but in a dedicated directory where your project will be cloned.. Using ./ will look for script and jar in current directory at the moment your command is run, but they wouldn't be found if not run from /opt/app. Specyfing full path to jar and adding your run.sh script to PATH make sure they'll be found wherever your run.sh from. Alternatively you could run cd /opt/app in your job's script but it may cause unwanted side effects.
• Without ENTRYPOINT you won't be able to run Docker commands like this

  docker run " <url>/my-image:<tag>" download ...
  

  You'll need to specify either COMMAND or --entrypoint such as

  docker run "<url>/my-image:<tag>" run.sh download ...
  docker run --entrypoint run.sh "<url>/my-image:<tag>" download ...
  

• You specified not wanting to do this, but overriding image entrypoint on your job seems a much simpler and straightforward solution. Using multiple files you may leverage Gitlab's extends and include.

---

And now for the fun part

what is happening

When Gitlab run your container for a job it will use the entrypoint defined in your Dockerfile by default. From doc:

1. The runner starts a Docker container using the defined entrypoint. The default from Dockerfile that may be overridden in the .gitlab-ci.yml file.
2. The runner attaches itself to a running container.
3. The runner prepares a script (the combination of before_script, script, and after_script).
4. The runner sends the script to the container’s shell stdin and receives the output.

And what the doc doesn't say is that Gitlab will try to use various form of sh as Docker command. In short for step 1. it's like running this Docker command:

# Gitlab tries to run container for your job
docker run -it "<url>/my-image:<tag>" sh

It doesn't work as Gitlab will use default entrypoint and the final command run in Docker is:

./run.sh sh

Where ./run.sh is the entrypoint from Dockerfile and sh is the command provided by Gitlab. It causes the error you see:

tool: error: invalid choice: 'sh' (choose from 'upload', 'download', 'apply')

You never reach your job's script (step 4). See ENTRYPOINT vs. CMD for details.

Furthermore, the script you define is a command itself. Even if your container started, it wouldn't work as the following command would be run inside your container:

download -f zip -u true test-download.zip

# 'download' command doesn't exists
# You probably want to run instead something like:

/opt/app/run.sh download -f zip -u true test-download.zip

Answer 2

So, after a bit of research, I have been able to find a solution that works for me.

From my research (and as Pierre B. pointed out in his answer), Gitlab essentially tries to inject a shell script that performs a check for which shell is available.

Now, my solution is in no way elegant, but does achieve what I wanted. I modified the Dockerfile like so:

FROM openjdk:18-bullseye

ENV JAR_NAME "my-jar.jar"

ENV PROJECT_HOME /opt/app
RUN mkdir -p $PROJECT_HOME

WORKDIR $PROJECT_HOME
COPY run.sh $PROJECT_HOME/run.sh
RUN chmod +x $PROJECT_HOME/run.sh

COPY target/$JAR_NAME $PROJECT_HOME/cli.jar

RUN  echo '#!/bin/bash \njava $JAVA_OPTS -Djava.security.egd=file:/dev/./urandom -jar $PROJECT_HOME/cli.jar "$@"' > /usr/bin/i18n &&  \
     chmod +x /usr/bin/i18n

ENTRYPOINT ["./run.sh"]

And also modified the run.sh script this way:

#!/bin/bash

if [[ -n "$CI" ]]; then
    echo "this block will only execute in a CI environment"
    exec /bin/bash
else
    echo "Not in CI. Running the image normally"
    java $JAVA_OPTS -Djava.security.egd=file:/dev/./urandom -jar ./cli.jar "$@"
fi

This works because Gitlab, in its list of predefined variables, provides a CI env var that is set when the script is running on the CI. By doing so, I skip the java invocation but leave it in the case I need to use it when not on a CI.

Now when I need to use my image, all I need to specify in my .gitlab-ci.yml file is this:

download:
  stage: download
  image: <url>/my-image:<tag>
  variables:
    URL: <URL>
    API_KEY: <API_KEY>
    CI_DEBUG_TRACE: "true"
  script:
    - i18n download -f zip -u true test-download.zip # This calls the script that was injected in the Dockerfile

This way I essentially mimic an actual CLI, and can use it in all my projects that require this tool.

I am not sure though why I need to "echo" the script for the CLI, and I can't simply copy it. For some reason the env variables are not passed down and I couldn't spend any more time debugging it. So for now, it will stay like this.

If you have any recommendations on how to clean this up, please leave some comments and I will edit my answer!

Answer 3

Try to wrap your script in single quotes:

script:
    - 'download -f zip -u true test-download.zip'

EDIT:

Oh, this open bug in gitlab could be relevant to you