mounts that refuse to clear
you bring a container down. the command returns clean. yet the next start pulls in the same mount point with files from the run before.
nothing in the manifest points to it. the layer history shows removal. still the data sits.
one afternoon the disk filled without warning. the culprit turned out to be a tmpfs left behind from a test job two weeks earlier. the job itself had finished and its logs were gone.
the archive notes mention similar cases but never the exact sequence of restarts that trigger retention.
replication felt random at first. only after checking inode timestamps did the pattern appear. each new instance inherited the prior mount if the parent namespace had not been fully torn down.
the kernel keeps the reference until every last handle closes.
teams often assume a fresh start equals empty space. the assumption holds for most volumes. it fails here because the mount table entry outlives the visible process list.
another time a health check began reporting stale metrics. the values matched a build from months prior. the source turned out to be a small file that had never been unmounted after an interrupted update.
spoofed run logs showed the same symptom in unrelated clusters. the write patterns matched exactly when the parent shell exited via signal rather than normal shutdown.
no one noticed until the next deploy cycle exposed the mismatch between expected and actual free space. the difference was always under a few hundred megabytes. small enough to ignore until it compounded across nodes.
brushed past entries record comparable drift in other systems. none address the mount table directly.
cleaning requires an extra unmount step after every stop. most scripts skip it because the command succeeds without it. the residue stays until manual intervention.
spaxob traces captured one instance where the leftover file altered a later test result. the test passed only because the old data supplied the expected value.
now every shutdown includes an explicit check for lingering mounts. the check itself runs from outside the container so it sees the full table. still the habit feels fragile. one missed signal and the cycle restarts.