testing.md

Testing

Automated tests

gofmt -l .
go vet ./...
go test ./...

The unit suite covers strict configuration, atomic reconfiguration, credential permissions, archive round trips, corruption detection, partial cleanup, native-client argument construction, parent/descendant cancellation, reaping, per-profile locks, abandoned-execution recovery, wizard cancellation, secret redaction, systemd unit generation/lifecycle, installer integrity, backup failure cleanup, restore verification, and retention safety. Windows scheduler tests also cover generated task XML, preset/custom cadence, daily/interval behavior, DST boundary construction, profile task naming, interactive-token least-privilege principals, missing tasks, unavailable service/query failures, snapshot validation, cancellation, and verified disable/remove/restore transitions. Installation-isolation tests use two canonical configuration homes with the same profile name and verify that task names, task actions, disable, and removal cannot cross installation boundaries. Rollback tests insert foreign tasks after both present-task and absent-task snapshots and verify that restoration fails closed without mutation. Legacy profile-only task names are migrated only after executable, argument, principal, and configuration identity checks pass. Additional tests cover password-mode registration, absence of Windows passwords in TOML/XML/history, password requirements during rollback, Task Scheduler local wall-clock conversion across winter/summer and DST boundaries, and manual-run requests that Task Scheduler accepts but never starts. Background (S4U) tests cover task XML, owner security descriptors, password-free registration and snapshots, rejection of network backup folders, and the profile identity used by the elevated registration helper. Wizard tests cover the database picker, credential failures during listing, preset preselection, and byte-order-mark handling in piped answers. Windows tests also exercise the one-shot password pipe: protected current-user DACL, round-trip delivery, timeout cancellation, invalid names, broker stderr redaction, and non-terminal prompt failure. CI parses the PowerShell Direct helper with the Windows PowerShell parser.

Disposable MariaDB acceptance test

Docker is optional and used only by the development fixture:

bash test/integration/run.sh

See test/integration/README.md in the repository. The test must never target a production database.

Manual MariaDB acceptance evidence

The initial implementation was exercised against a disposable MariaDB 11.8.6 instance on loopback using mariadb-dump and mariadb clients. That run covered:

• 50,000 generated rows and mixed InnoDB/MyISAM tables
• a view, procedure, function, trigger, and scheduled event
• successful compressed backup, checksum verification, and local history
• restore into a separate target database with matching 50,000 rows and object counts
• empty database backup
• corrupted archive rejection
• invalid credential rejection without publishing an archive
• insufficient routine privileges rejection without publishing an archive
• keep_last = 2 retention after three real backups
• password absence from execution logs in the invalid-credential case

This evidence does not replace running the reproducible Docker script or a Linux acceptance test.

Linux checks without installing services

• systemd-analyze validates generated service/timer files in Linux tests.
• Every generated interval calendar is passed to systemd-analyze calendar, including arbitrary start times and daylight-saving transitions.
• systemctl/loginctl enable, disable, remove, logging, and next-time operations are tested with mocked command execution.
• Failure injection covers unit creation, reload, activation, disable, removal, unconfirmed activity, stale manual-mode timers, and configuration rollback.
• Readiness tests cover Unix socket and TCP recovery, permanent credential and database errors, bounded timeout, cancellation, and archive suppression.
• State-query tests distinguish enabled/active, disabled/inactive, legitimate not-found, D-Bus failure, permission denial, timeout, cancellation, malformed output, and unexpected states.
• Cancellation tests cover before the transaction, after timer pause, during persistence/activation, during removal, and failure of rollback itself.
• The local installer test builds a Linux binary, installs it into a temporary prefix, rejects a bad checksum, uninstalls it, and verifies that user data is preserved.
• The Windows installer test performs the equivalent per-user install, upgrade, checksum rejection, uninstall, and user-data preservation flow, without changing the user's PATH.
• Actual persistent timer registration is intentionally not performed on the development host. It requires a disposable Linux VM or explicit authorization.
• A transient user service with the production hardening directives successfully executed the backup binary, read configuration and credentials, wrote state and archive files, and connected through a Unix socket.
• Native Windows tests inspect actual DACLs for configuration directories, credentials, reconfiguration/rollback temporaries, and native-client option files, including Unicode/space paths and repair of broad ACLs. The socket probes are in test/systemd/.

Windows Task Scheduler checks

Normal Windows CI validates generated XML with the native Task Scheduler COM parser and queries a legitimately missing task without registering anything. An explicitly guarded lifecycle test can register a disposable task, disable it, snapshot/restore its state, and delete it:

$env:FIVEMESH_TEST_TASK_SCHEDULER = "1"
go test -count=1 -run TestWindowsTaskSchedulerLifecycleIntegration ./internal/taskscheduler

Run that command only on a disposable Windows account.

Before a release, also check on a disposable Windows Server VM: setup with a background schedule, a scheduled run while signed out, a reboot with a missed run caught up, disable and re-enable, and a restore into a separate database.