- Decision
- ADR-101
- Date
- 2026-09-26
- Status
- SupersededSuperseded. The root-cause finding (below) and the RAM/swap/crypto exhaustion on the signing host are still correct and kept for the record; the fix chosen was wrong.
- Mentions
- ADR-102
- Mentioned by
- ADR-102
1Decision
Split sign-ota.sh's ceremony into sign-ota-pending-verify.sh plus finalize-ota.sh to move the OOM-prone OTA re-verification to the build host. Superseded before first use, same day — see ADR-102.
2Why and evidence
check_ota_package_signature.py's VerifyPackage() reads the whole OTA zip into memory (package_bytes = package_file.read()) to check its whole-file signature — stock AOSP behavior, confirmed by reading the actual pinned source, not assumed. This OOM-killed the re-verification step of candidate 2026092501 on the 1.9GB-RAM signing host (anon-rss ~1.84GB, kernel log confirmed). The signing host has no hypervisor access to add RAM, its LVM volume group has 0 free space to grow swap, and it is intentionally network-isolated so installing cryptsetup was rejected outright.
3Consequences
The owner pointed out this project already has a working, simpler tool for exactly this situation (tools/signing/sign-baseline-images.sh, proven 2026-08-25, ceremony-log entry kind=baseline-images result=SUCCESS): it produces a directly-flashable, release-signed image bundle instead of an OTA package, and its own independent verification (avbtool verify_image plus a small otacerts.zip extraction) never reads a multi-gigabyte file into memory, so it does not hit this OOM at all. Writing two new ceremony scripts was unnecessary — I had not checked for existing tooling before designing a fix, contrary to this project's own prior-art discipline.
4Review when
New ceremony scripts were never deployed or run and are removed from the repo (never reached the signing host).
Read from docs/governance/decision-log.md, entry ADR-101.