A global retail company's Ceph cluster nodes were experiencing out-of-memory kills, with PID 1 (systemd) consuming ~70 GB of RAM instead of the expected 8 GB. Canonical Support engineers traced the root cause across three system layers: a cephadm orchestration misconfiguration was repeatedly attempting to 'zap' a non-existent OSD, causing a ceph-volume Docker container to spin up and tear down in a continuous loop. This generated millions of mount table entries, triggering a kernel race condition that caused systemd to continuously reparse the mount table while glibc's allocator retained the memory without returning it to the kernel. The fix involved correcting the cephadm loop as a workaround, then isolating a minimal reproducer (mounting/unmounting tmpfs in a tight loop) to characterize the kernel race condition precisely. A patch was merged into the mainline Linux kernel and subsequently landed in several Ubuntu kernel releases via the Stable Release Update process.
