feat(kernel): bake skips a kernel BuildRun the registry already holds #2080
Loading…
Reference in a new issue
No description provided.
Delete branch "feat/bake-skip-published"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Kaizen run 4, improvement B (from profiling sf42: 755s of re-bake for already-published images). Kernel BuildRun skipped behind a registry manifest proof; node-bootc always rebuilds. Guard P10 + contract M15; 14/14.
Profiling the sf42 evidence-reuse release: 755s of its 26 minutes was the bake re-running BuildRuns for images already published — the git-clone layer is not byte-stable, so the BuildKit layer cache misses across releases and the kernel recompiles (~9m17s, ccache-warm) for byte-equivalent output. build/linux-kernel:${SHORT} is a pure function of pinned inputs (source SHA, immutable kernel-builder :vN, the build args), so an existing manifest IS proof this exact bake already succeeded. The bake now HEADs the registry and skips the kernel BuildRun on 200; any registry error falls through to building — the safe direction. The kernel digest always flowed from crane against the registry, so no downstream identity changes. node-bootc is NEVER skipped: it tracks builder-images main (the bootc-timer-mask class) and costs only ~3 minutes. Guard P10 pins both halves (skip only behind registry proof; node-bootc never gains one); contract M15 proves the second half bites — 14/14.