Skip to content

Latest commit

 

History

History
135 lines (110 loc) · 5.13 KB

README.md

File metadata and controls

135 lines (110 loc) · 5.13 KB

skinnywhale

Skinnywhale helps you make smaller (as in megabytes) Docker containers

--- Skinny whale is positively starving ---

                    ##        .
              ## ## ##       ==
           ## ## ## ##      ===
       /""""""""""""""""\___/ ===
      /	rX
  ~~~{ /\ ~ ~~~ ~~~~ ~~ ~ /  ===- ~~~
       \______////////////__/

If, for example, you want to run a python script in a container, normally you'd have to download a 600MB image that had the python interpretor along with most of an OS inside it. Skinnywhale helps you isolate runtime environments like python, ruby, and java, and throw the rest of the stuff away, so the images you make with it are normally hundreds of megs smaller than their bloated counterparts. Starting from a generic Ubuntu Docker image, my Skinnywhale Python images usually come out to be around 30MB. Here's how you use it:

step 1. Start with a normal fat image

docker run -ti ubuntu

step 2. Add the stuff you need to it (creates a new container)

apt-get update
apt-get install python3
exit

docker ps -a
<copy the container ID you just created eg.. 8efbc5497abb>

step 3. use skinnywhale to isolate the stuff you just added

skinnywhale 8efbc5497abb

Skinnywhale requires that you're using an aufs-backed docker that is version 1.6 or later (there are bugs earlier than that which cause "file not found" errors on images imported from tarballs)

You can see what kind of back-end you have with:

#  docker info
Containers: 2
Images: 7
Storage Driver: aufs
 Root Dir: /var/lib/docker/aufs
 Backing Filesystem: extfs
 Dirs: 11
 Dirperm1 Supported: false

If you're on ubuntu and your docker is not using aufs, it's probably because you're missing the linux-image-extra package. Try running:

apt-get -y install linux-image-extra-$(uname -r)
service docker restart
docker info

ANYWAY if everything went ok, skinnywhale will write a "skinny" version of the container you listed as an image. You should see it listed in docker images. You can specify this image in a Dockerfile like so:

FROM skinny_8efbc5497abb
ADD  myPythonScript.py /
CMD ["/usr/bin/python", "myPythonScript.py"]

Now build it:

docker build -t myTeensyAppContainer .

...and now you have a dockerized version of your app that contains only your script and the runtime needed to execute it (and not the entire rest of ubuntu or whatever).

A few Options

Skinnywhale uses environment variables to control a few options:

  • DEBUG : enable debug mode
  • NOCLEANUP : don't clean up the working directories in /tmp
  • BRUTELIB : brute-force copy the entire /lib directory from the fat image (try this if you're having name resolver issues etc..
  • BRUTEUSRLIB : brute-force copy the entire /usr/lib directory from the fat image.

How does this work?

Skinnywhale is pretty simple, it finds the aufs path for the container ID you give it, and walks through it looking for binary files that are dynamically linked. Skinnywhale resolves each of these dependencies by copying the linked library files in from the parent image. It then tars up the resulting filesystem and docker imports it.

There are two caveats to this process. First, some binary-distributed runtime environments (like the oracle JDK), contain files that link to libs that may not be installed on the client (if, for example you're using a base-image that doesn't have Xorg installed, the JDK WILL have binary files that are linked to X11 libs that don't exist on your base image). This is not usually a problem; as long as your app works with the base image you're trying to use, it should also work post-skinnywhale.

Second, Skinnywhale can't detect if your runtime uses dlOpen(). You're on your own there I'm afraid. FWIW I've been using skinnywhale for python and java stuff at work, and in practice I've found this not to be an issue. If you're having trouble getting stuff to run after it's been skinnywhaled, try setting the BRUTE variables described above. Chances are, if someone is using dlOpen(), those libs are going to be part of the distribution files for the runtime (ie you aren't going to assume a lib file is lying around somewhere unless you put it there, (unless you're silly or mean)), and since skinnywhale grabs the entire runtime, it ought to be pretty safe with respect to dlOpen'd stuff. YMMV. Caveat emptor. etc...

No I mean HOW does this work?

Oh. Pretty well I guess? Java images are still stupid big, because java is stupid big. Skinnywhale is irrelevant for Go because static binaries. It gets a python runtime down to about 33MB, but yeah, the jvm environments are still like 370MB (it is what it is. Still beats the gigabyte-sized java images floating around in the public registries).

Since skinnywhale is really just using the diff functionality built-in to aufs, it's completely agnostic to factors like the OS of your base-image. So if you start with a busy-box image, and use skinnywhale to isolate the java-runtime from that, you may end up with something smaller. The relevant factor is how much crap the package-manager in your base-image vomits into the file system when you use it to install something like Java. Good luck!