mirror of
https://github.com/calibrain/shelfmark.git
synced 2026-09-28 18:15:08 +01:00
splitsec2
ca25448529
With the amount of changes and testing I've been doing lately I noticed how long the builds were taking and how much each one pulled, so I went and looked at the Dockerfile. I think this balances cache and efficiency. Three changes needed to be stacked to make it happen. **The version stamp sits above everything expensive.** `ARG BUILD_VERSION` and `ENV BUILD_VERSION` are at the top of the `base` stage and the value carries the commit sha, so it changes on every commit. This invalidates the layer and everything below it, which means the apt install, the dependency sync and the Chromium install. Nothing in the build reads either variable. They're only used at runtime by entrypoint.sh, tor.sh, wireguard.sh and genDebug.sh, so they can move to the end of the final stages. **`COPY . .` sits above the Chromium install.** It's in `base`, and the `shelfmark` stage installs Chromium and the seleniumbase drivers after it, so any source change rebuilds those too. Moving the source copy and the runtime-paths block to the end of each final stage solves that. **There's no cross-run cache.** Runners are ephemeral, so without `cache-from` and `cache-to` every layer is rebuilt on every run whatever the ordering, and a rebuilt layer gets a new digest even when the content is identical. That's why reordering on its own doesn't change the load. Measured with a source-only change between two builds: | | Build | Pull | |---|---|---| | before | 13m 20s | 526.5 MB | | after | 2m 25s | 3.7 MB | 15 of 18 layers get reused where it was 5. The cache sits at 0.86 GB, which leaves room under the 10 GB repo budget for the uv caches in ci.yml and e2e-platform.yml. I tried `mode=max` first and it built a bit quicker at 1m 43s, but it used 5.15 GB of cache and the potential to impact other workflows so it didn't seem worth 40 seconds, but that is a single line fix if you want to. This means the runtime-paths block is now duplicated, once per final stage, and that's most of the diff. It has to sit below each stage's heavy layers to do its job and I couldn't find a way around that short of another shared stage, which looked like more complexity for complexity's sake. Happy to take another run at it if you'd rather have it 'DRY'. I checked the built image against the current one. Same size, it boots, /api/health returns 200, and BUILD_VERSION and RELEASE_VERSION are still stamped correctly. Also, I'll slow down on the PRs. Promise.