Kept build caches grow until the job's disk is full issue #72 closed: completed andymccall opened this 15 days ago · 0 comments

Found 2026-10-20 01:45: oxygenhub's test-postgres (cache "target", 24 GiB disk) failed twice in a release run: first "No space left on device" writing an rlib, then the linker killed by SIGBUS writing its mmapped output. The kept target accumulates every build's artifacts (old hashes of every crate after each Cargo.lock change, both feature sets) and is never pruned; the cache is taken whole into the next run. Stopgap: that job's disk raised to 40 GiB (oxygenhub 51e67b9a). Fix: before keeping a disk, prune it (for cargo targets, drop artifacts not touched by this run, as cargo-sweep does by atime/mtime; generically, a per-cache size cap with oldest-first removal), and say in the job log how much was kept and pruned. Both tiers (microVM ext4 images and syscall directories, issue #70).