Navigation Menu

Skip to content

ordinerf/docker-slim

 
 

Repository files navigation

docker-slim: Make Your Fat Containers Skinny and Tough (optimize and secure*)

Docker Global Hack Day #dockerhackday project (status: ACTIVE!)

Just because the hack day is over doesn't mean the project is done :-) The project needs your help even if you don't know Docker or Go!

IRC (freenode): #dockerslim

DEMO VIDEO

DockerSlim demo

Demo video on YouTube

DESCRIPTION

Creating small containers requires a lot of voodoo magic and it can be pretty painful. You shouldn't have to throw away your tools and your workflow to have skinny containers. Using Docker should be easy.

docker-slim is a magic diet pill for your containers :) It will use static and dynamic analysis to create a skinny container for your app.

CURRENT STATE

It WORKS with the sample node.js, python, and ruby images (built from sample_apps). More testing needs to be done to see how it works with other images.

Sample images (built with the standard Ubuntu 14.04 base image):

  • nodejs app container: 431.7 MB => 14.22 MB
  • python app container: 433.1 MB => 15.97 MB
  • ruby app container: 406.2 MB => 13.66 MB
  • java app container: 743.6 MB => 100.3 MB (yes, it's a bit bigger than others :-))

You can also run docker-slim in the info mode and it'll generate useful image information including a "reverse engineered" Dockerfile.

DockerSlim now also generates an AppArmor profile for your container.

Dependencies:

To run docker-slim you need to export docker environment variables. If you use docker-machine you get it when you run eval "$(docker-machine env default)".

USAGE

./docker-slim <IMAGE_ID_OR_NAME> [rm-artifacts | image-info-only]

Example: ./docker-slim 6f74095b68c9

By default, docker-slim doesn't remove the artifacts it generates. To remove them set the rm-artifacts flag.

Example: ./docker-slim 6f74095b68c9 rm-artifacts

To generate a Dockerfile for your "fat" image without creating a new "slim" image set the image-info-only flag.

Example: ./docker-slim 6f74095b68c9 image-info-only

DEMO STEPS

The demo run on Mac OS X, but you can build a linux version.

  1. Clone this repo

    git clone https://github.com/cloudimmunity/docker-slim.git

  2. Create a Docker image for the sample node.js app in sample_apps/node

    cd docker-slim/sample_apps/node

    eval "$(docker-machine env default)" <- optional (depends on how Docker is installed on your machine)

    docker build -t my/sample-node-app .

  3. Run docker-slim:

    cd ../../dist_mac

    ./docker-slim my/sample-node-app

    DockerSlim creates a special container based on the target image you provided.

  4. Use curl (or other tools) to call the sample app (optional)

    curl http://<YOUR_DOCKER_HOST_IP>:<PORT>

    This is an optional step to make sure the target app container is doing something. Depending on the application it's an optional step. For some applications it's required if it loads new application resources dynamically based on the requests it's processing.

    You can get the port number either from the docker ps or docker port <CONTAINER_ID> commands. The current version of DockerSlim doesn't allow you to map exposed network ports (it works like docker run … -P).

  5. Press any key and wait until docker-slim says it's done

  6. Once DockerSlim is done check that the new minified image is there

    docker images

    You should see my/sample-node-app.slim in the list of images. Right now all generated images have .slim at the end of its name.

  7. Use the minified image

    docker run --name="slim_node_app" -p 8000:8000 my/sample-node-app.slim

Notes:

You can explore the artifacts DockerSlim generates when it's creating a slim image. You'll find those in dist_mac/container/artifacts. One of the artifacts is a "reverse engineered" Dockerfile for the original image. It'll be called Dockerfile.fat.

If you'd like to see the artifacts without running docker-slim you can take a look at the sample_artifacts directory in the repo. It doesn't include any image files, but you'll find:

  • a reverse engineered Dockerfile (Dockerfile.fat),
  • a container report file (creport.json),
  • and a sample AppArmor profile (which will be named based on your original image name).

If you don't want to create a minified image and only want to "reverse engineer" the Dockerfile you can use the image-info-only option.

You can get the current binaries for Macs here

BUILD PROCESS

Before you build the tool you need to install GOX and Godep (optional; you'll need it only if you have problems pulling the dependencies the old fashioned way :-))

1: go get github.com/tools/godep

1: go get github.com/mitchellh/gox

2: gox -build-toolchain -os="linux" -os="darwin" (note: might have to run it with sudo)

Local Build Steps

Once you install the dependencies (GOX - required; Godep - optional) run these scripts:

  1. Pull the dependencies: ./src.deps.get.sh
  2. Build it: ./src.build.sh

You can use the clickable .command scripts on Mac OS X:

  1. mac.src.deps.get.command
  2. mac.src.build.command

Builder Image Steps

You can also build docker-slim using a "builder" Docker image.

  1. Create the "builder" image: ./docker-slim-builder.build.sh (or click on docker-slim-builder.build.command if you are using Mac OS X)
  2. Build the tool: docker-slim-builder.run.sh (or click on docker-slim-builder.run.command if you are using Mac OS X)

DESIGN

CORE CONCEPTS

  1. Inspect container metadata (static analysis)
  2. Inspect container data (static analysis)
  3. Inspect running application (dynamic analysis)
  4. Build an application artifact graph
  5. Use the collected application data to build small images
  6. Use the collected application data to auto-generate various security framework configurations.

DYNAMIC ANALYSIS OPTIONS

  1. Instrument the container image (and replace the entrypoint/cmd) to collect application activity data
  2. Use kernel-level tools that provide visibility into running containers (without instrumenting the containers)
  3. Disable relevant namespaces in the target container to gain container visibility (can be done with runC)

SECURITY

The goal is to auto-generate Seccomp, AppArmor, (and potentially SELinux) profiles based on the collected information.

  • AppArmor profiles (the auto-generated profiles are almost usable :-)).

CHALLENGES

Some of the advanced analysis options require a number of Linux kernel features that are not always included. The kernel you get with Docker Machine / Boot2docker is a great example of that.

DEVELOPMENT PROGRESS

PHASE 1 (DONE)

Goal: build basic infrastructure

Create the "slim" app that:

  • collects basic container image metadata [DONE]
  • "reverse engineers" the Dockerfile used to create the target image [DONE]
  • creates a container replacing/hooking the original entrypoint/cmd [DONE]
  • creates a new "slim" image from the collected information and artifacts [DONE]

Create the "slim" launcher that:

  • starts the original application (based on the original entrypoint/cmd data) [DONE]
  • monitors process activity (saving events in a log file) [DONE] (note: doesn't work with all kernels)
  • monitors file activity (saving events in a log file) [DONE]

PHASE 2 (DONE)

  • Fix new image permission errors [DONE]
  • Use env data from the original image [DONE]

MILESTONE 1 - MINIFIED TEST DOCKER IMAGE (DONE)

The minified sample_app docker image now works! We turned a 430MB node.js app container into a 40MB image.

PHASE 3 (ACTIVE)

  • Do a better job with links [DONE] The test image is now even smaller (was: 40MB, now: 14.22MB)
  • Make sure it works with other images [WIP, now: node,python,ruby,java].
  • Refactor the time-based container monitoring phase.
  • Automated interaction with the target container (requires app code analysis).
  • Auto-generate AppArmor profiles [WIP].
  • Auto-generate Seccomp filters.
  • Split "monitor" from "launcher" (as it's supposed to work :-))
  • Add scripting language dependency discovery to the "scanner" app.
  • Support additional command line parameters to specify CMD, VOLUME, ENV info.
  • Build/use a custom Boot2docker kernel with every required feature turned on.
  • Explore additional dependency discovery methods.
  • "Live" image create mode - to create new images from containers where users install their applications interactively.

NOTES

  1. The code is really really ugly at this point in time :)
  2. Each app directory contains a dummy .git directory because godep fails to work without it.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Go 84.7%
  • Java 7.3%
  • Shell 6.4%
  • Other 1.6%