charmcraftcache

charmcraftcache

Reinventing the wheel

Fast first-time builds for charmcraft—on a local machine or CI

Installation

Install pipx: https://pipx.pypa.io/stable/installation/

pipx install charmcraftcache

Usage

For best results, use charmcraft's Poetry plugin or add charm-strict-dependencies: true to your charmcraft.yaml.

Step 1: Update charmcraft.yaml to supported syntax

Only ST124 - Multi-base platforms in craft tools "shorthand notation" syntax is supported

Example

platforms:
  ubuntu@22.04:amd64:
  ubuntu@22.04:arm64:
  ubuntu@24.04:amd64:
  ubuntu@24.04:arm64:

Under the charmcraft.yaml platforms key, build-on and build-for syntax are not supported

The base and bases charmcraft.yaml keys are not supported

Step 2: Add your charm to the cache

ccc add

Step 3: Pack your charm

ccc pack

How it works

Why are charmcraft builds slow?

Instead of downloading wheels from PyPI (which pip does by default), charmcraft builds Python package wheels from source (i.e. with pip install --no-binary).

Caching mechanism

charmcraft builds each charm platform in a separate LXC container¹. Within each container, pip has an internal cache for wheels built from source & for HTTP responses.

charmcraft 2.5 moved the pip internal cache to the LXC host machine, so that one pip cache is used for all LXC containers. (This increases the chance of a cache hit—a faster build.)

However, charmcraft builds are still slow the first time the wheel is built. This happens on CI runners, when you use a new machine/VM, or when you contribute to a new charm.

charmcraftcache solves the slow first build.

charmcraftcache-hub maintains a list of charms. For each charm, charmcraft pack is used to build Python dependencies from source and the pip wheel cache is uploaded to a GitHub release.

ccc pack downloads these pre-built wheels to charmcraft's pip cache (and then runs charmcraft pack).

Note: Within the GitHub release, each charm has an isolated cache. If the same charm (same GitHub repository and relative path to charmcraft.yaml) is added to the list of charms more than once (with different git refs), the wheels are combined into a single cache. If there are duplicate wheels, the wheel is selected from the ref that is earlier in the list.

Isn't this just a worse version of PyPI?

Pretty much. The only difference is charmcraftcache-hub wheels are built from source on our runners, instead of built by the package maintainer.

Why does charmcraftcache only support "shorthand notation" syntax in charmcraft.yaml platforms?

"Shorthand notation" is used when build-on is identical to build-for. (For example, "ubuntu@22.04:amd64" means build on Ubuntu 22.04 amd64 and build for Ubuntu 22.04 amd64.)

For charms that depend (directly or indirectly) on Python packages with C extensions (e.g. pyyaml), charmcraft will build wheels where the C extensions only work on build-on.

For example, if a charm is built with

platforms:
  foo:
    build-on: ubuntu@22.04:amd64
    build-for:
      - ubuntu@22.04:amd64
      - ubuntu@22.04:arm64

the wheels in the *.charm file will contain C extensions that only work on amd64, not arm64.

The vast majority of charms have at least one Python dependency with C extensions, so the vast majority of charms should use "shorthand notation".

Footnotes

1. Unless --destructive-mode is enabled ↩