Cloud software lives in a clean world of JSON and retries. Biometric gate hardware lives in a world of fixed-length packets, flaky network switches, and devices that power-cycle at the worst moment. The Centralized Gate Pass System had to make those two worlds agree on a single question: is this person allowed through this gate, right now? ## Two sources of truth, one verification table The ZKTeco devices hold their own user and fingerprint records locally so a gate keeps working if the network drops. The cloud holds the authoritative roster. The danger is drift: a revoked pass that still opens a gate because the device never got the memo. The whole design centers on keeping the `dop_verification` table and the on-device records converging, not diverging. ## Push, don't hope Early on I leaned on the device's own polling. It was unpredictable. The reliable pattern was to treat the backend as the source of intent and push changes to devices, then confirm: ``` def sync_user(device, user): device.set_user(user.pin, user.name, privilege=user.role) device.set_fingerprint(user.pin, user.template) ack = device.read_user(user.pin) # read back to confirm Verification.objects.update_or_create( pin=user.pin, device=device, defaults={"synced": ack == user, "last_sync": now()}, ) ``` The read-back is the part that turns "I sent it" into "it is actually there." ## Idempotency and retries, because devices disappear Network calls to hardware fail constantly — a switch reboots, a cable is bumped. Every sync operation is idempotent so it can be retried safely, and failures go into a queue rather than crashing the request: ``` @shared_task(bind=True, max_retries=5) def sync_user_task(self, device_id, user_id): try: sync_user(get_device(device_id), get_user(user_id)) except DeviceUnreachable as exc: raise self.retry(exc=exc, countdown=2 ** self.request.retries) ``` Exponential backoff means a device that's down for thirty seconds doesn't generate a thousand failed calls. ## Keeping the verification database honest Optimizing `dop_verification` was half indexing and half discipline: index the columns every gate-check filters on (pin, device, active), and run a reconciliation job that compares device records against the table and flags mismatches for review. The gate's job is to be fast and correct; the reconciliation job's job is to make sure "correct" stays true overnight. ## The takeaway Integrating physical hardware taught me to design for the failure, not the demo. The happy path — device online, network up — writes itself. The system earns its keep in how it behaves when the gate's switch reboots mid-sync.