The garage door rebooted because DNS failed

LiftMaster’s MyQ Home Bridge reboots in a loop when it cannot reach the vendor’s MQTT brokers. That is the incident. Everything after it is a decision about where you fix it. Nobody sat down with me for this interview. I am running it on myself, because I have taken both positions below in the last few years and I would take each of them again on the right device.

A conversation with both of my hats

The write-up that started this is a teardown. Bare SWD pins on a Cortex-M4, an Arduino doing the talking, an MQTT key lifted off the chip, then a local broker and a DNS redirect so the device believes the cloud answered. Hackaday covered it, and the comment section immediately split into “let her go” and “nice work”. Both.

Q: Start with the extraction. Was reading the key off the microcontroller the right first move?

A: No, and I say that as someone who owns a Bus Pirate. The first move is a packet capture, not a soldering iron. Put the thing on a VLAN by itself, point its DNS at a resolver you control, and read the query log for twenty minutes. Devices like this are loud. They name every host they want and show you how often they want it. In this build the answer was already visible before anyone opened the case: one broker domain, a handful of hosts under it, keepalives on a timer. That is a small enough surface to fake. Extraction becomes correct when the cheap version failed, or when the firmware refuses any answer from a broker that does not hold its key. Here it was the second case, which is why the SWD pins mattered at all. The pins were exposed and an evening with an Arduino was enough. Fine. Measure first anyway, or you will spend a weekend on a problem you never sized.

Q: And the other option. Bin the cloud path, put an ESP32 on the door.

A: The swap wins on almost every axis I care about. Ten dollars, ESPHome, a reed switch, no vendor broker, no key to lose, no update that wipes your work. It also fails in the one place the teardown does not: the opener is a mechanical install with a wall panel and safety sensors that stop the door on contact. If the actuator stays, you either talk to it or you bolt a second radio alongside it. Two costs get skipped. Nothing you learn about one vendor’s handshake transfers to the next device. And you inherit the glue: every firmware update is a possible outage, and you will be the only person alive who knows why the door stopped answering.

So the swap is right when the opener is dumb and separate, which covers most of the installs I have walked past. This unit was a Home Bridge, an accessory wired next to the motor. Extraction was defensible.

What a network engineer should do first, regardless

Q: Neither of those is a network fix. What belongs on the checklist before either one?

A: Instrument, isolate, redirect. Isolation is a VLAN whose egress policy denies by default, so a device that goes feral costs you nothing but log lines. Instrumentation is DNS logging, because the query log is the cheapest dependency map you will ever get. Then the redirect, which is usually one line:

# dnsmasq: answer the vendor namespace locally
address=/mqdevice.com/10.20.30.5

If the log shows one broker name and a keepalive, a stub that answers pings satisfies it. Done in an hour. If it shows a dozen names pulled from a config file, stop: that device is a client of a service, not an appliance, and you have just volunteered to run the service.

A watchdog that restarts hardware because DNS did not answer is not protecting the device. It converts a remote dependency into a local outage loop, and it will keep doing that until something upstream of the firmware lies to it convincingly.

Q: Which brings us to the firmware. What does a reboot loop like that tell you?

A: That its failure model is inverted. A watchdog should catch a hung task. This one catches an unreachable server, and the only remedy it owns is restart. Cheaper to write than a state machine, and it lets the vendor call the result self-healing. The testable version of my claim: drop the WAN for an hour while your resolver keeps answering, and see whether the loop stops. If it does, the cloud was never the requirement. The heartbeat was. I have not run that test on this unit, so the timeout interval is a guess on my side, and I would measure it before describing recovery behaviour to anyone.

Next on my list: a week of WAN drops with the stub answering. Then one firmware update, to see whether the key and the handshake survive it, since an update that re-keys the session puts an expiry date on the whole trick. That date decides whether I recommend it.

Press ⌘K to search & switch themes