PennockTech / ci

Continuous Integration image building

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ci

Continuous Integration images.

These are not for production deployments of code: they are large and bloated. These images contain Stuff We Often Need In CI Testing. The two use-cases are fundamentally different.

Sub-directories docker-FOO each contain a Dockerfile controlling the build of a Docker image.

We have a philosophical objection to encouraging the use of sudo, instead of capabilities and drop-only privileges, albeit not as great as our objections to mode 0777 directories, curl | sh or retrieval without checksum verification.

Beware that these add my own certificate authorities to the system trust stores.

docker-gobuild

Alpine images derived from upstream golang:${VERSION}-alpine images, with a strictly minimal set of packages added.

These images should be sufficient to build. They will not be extended to include enough packages to make it easier to test or debug. They do alas include the docker CLI to be able to talk to Docker, to build images, not just binaries. Tools for building other packaging artifacts will probably not be added, instead a new variant will be used.

$ docker-gobuild/build.sh 1.14.6

docker-purple

First-generation Pennock Tech CI image for ourselves, rather than for clients. Based on Alpine, this is rather heavy, with gcc, clang, python, go and much more.

Creates:

  • pennocktech/ci:purple-root: everything installed, uid is root
  • pennocktech/ci:purple: run-time user is ci

Known issues:

  • build style is incompatible with Docker build-step caching, needs a rethink to find a decent middle-ground while still avoiding the 20-line chained-&& pattern
  • missing kubectl still
  • too large

I do development work on Zsh, so the usual suspects such as "remove fancy shells", are not really tenable. I mostly wanted to get an image with up-to-date pip out, before the PyPI cut-over, so haven't yet prioritized sane shrinkage.

docker-pink

Stripped down from docker-purple, this has things required for a custom Go build but is missing most of the rest of the tooling added. The purple tooling is used to create this image.

docker-pastel

This is created with the purple tooling, but only installs the most critical of packages:

  • curl
  • git
  • jq
  • openssl

A few config files go in too, including the aforementioned certificate trust stores being setup.

Sizes

I periodically update the Docker Hub description with stats on the (horrendously large) sizes of these images. They're not small. Size is not a goal.

For production deployments, Small Is Good. The smaller it is, the faster it transfers, the less there is to go wrong. The less likely it is to be evicted from cache. You can be small relative to a large base image, as long as the base image truly is widely used across all the target hosts, so will already be present.

My typical pattern is to use a multi-stage Dockerfile to build Go code in golang:x.y.z or alpine:x.y, then copy it into a FROM scratch final image in the last stage, or sometimes FROM alpine if I want to be able to SSH in. I expect total image sizes to be less than 20MiB, or less than 50MiB if Alpine is used. Perhaps more if there's a lot of ancillary data which for some reason belongs in the image instead of a distinct Volume.

By contrast, these CI images take a "kitchen sink" approach. The goal is to not have to think too much about what I'm depending on when writing tests: being able to write and run decent tests is more important than keeping the CI image small, and an extra minute in CI runner start-up time is not a problem. As long as the build is small, so that it's comprehensible and easy to audit, I am happy.

So as a rough ballpark:

  • Pastel: 30 — 40 MiB
  • Pink: 1.5 — 1.7 GiB
  • Purple: 2.5 — 3 GiB

Those are insane. I should probably abandon Pink and Purple and use Ubuntu images for that sort of testing. If I'm pausing to think about it when writing tests, I might. But otherwise: these are available, they work, and they're a consistent environment.

But they're not a template for constructing an image for deployment to Production. And they're not intended to be.

About

Continuous Integration image building

License:MIT License


Languages

Language:Shell 75.1%Language:Dockerfile 14.7%Language:Vim Script 10.1%