Skip to content
Obsidian Integration Testingv10.3.2

Leftover cleanup

A run that dies mid-flight cannot clean up after itself. On Android that is the normal case, not the exception: teardown removes the vault through the WebView, and a dead WebView is exactly what most failures are (Vault cleanup error (non-fatal): no such window). So every failure leaves a temp-vault-* directory behind — and, worse, leaves it registered, which is work Obsidian has to redo at every startup, inside the same WebView-readiness budget the run is already straining. Failures therefore make the next failure likelier. One real emulator had accumulated 103 leftover vaults.

Every run sweeps at both ends, and the start sweep is the one that matters, because it runs before anything that can die:

  • On the device (Android) — before the Appium session launches Obsidian, every temp-vault-* directory under vaultBasePath is removed over adb, and their stale entries are pruned from Obsidian Mobile’s localStorage vault registry when the run registers its own vault. Unregistering a vault removes its device directory over adb whether or not the WebView answered.
  • On the host — leftover temp-vault-* staging directories and owned userdata-* instance profiles in the system temp directory are removed.

The two halves gate differently, on purpose:

SweepGateWhy
DeviceUnconditionalAndroid runs hold an exclusive lock, so no concurrent run can own a device vault — and an age gate would let a vault leaked minutes ago poison a rerun.
HostOnly older than leftoverMaxAgeInMillisecondsDesktop runs are deliberately not serialized, and every project on the machine shares one temp directory, so a young directory may belong to a live run.

Both knobs are available on either transport:

  • shouldSweepLeftovers (default true) — set false to disable both sweeps entirely.
  • leftoverMaxAgeInMilliseconds (default 7200000, i.e. two hours) — the host age gate. Raise it if a run of yours can outlive the default; 0 removes every host match regardless of age.

Sweeping is best-effort throughout: a directory another process still holds is skipped, never thrown.

One directory at a time, and the result is measured

Section titled “One directory at a time, and the result is measured”

The device sweep removes each vault with its own rm -rf and then re-lists what is left, so the count it reports is what actually went away rather than what it asked for. Both halves of that matter, and both come from the same measured failure:

  • An Android emulator can end up holding a directory whose name the FUSE layer cannot express — rm -rf, find -delete and force-stopping Obsidian first all answer Operation not permitted, and it is permanent. Removing the whole set in one command let that single entry decide the fate of every other: one device was found carrying 26 leftover vaults that the sweep had been “removing” every run. Per-directory, it costs one warning per run instead of the whole sweep.
  • A removal that leaves the directory behind is otherwise invisible, because rm -rf runs with its exit code ignored. A run that passed end to end was still leaking a vault apiece, silently. Anything that survives is now named in the log, at both ends — the start-of-run sweep and the teardown — and is retried by the next run’s sweep.
Section titled “Related”