Can a Five-Story Villa Still Function Normally When the Central Controller Is Shut Down? A Smart Home Failure Acceptance Checklist
An acceptance plan for shared failure points, misleading latency percentiles and stale actions after recovery, with runnable calculations and test-record templates.
In a smart home demonstration, turning on lights from a phone, linking curtains to other devices, and viewing camera feeds can all look seamless. But for a five-story home of about 1,000 square meters, the more important question is: when one device fails, which household functions disappear with it?
This article proposes an acceptance method for a villa project that is still at the unfinished-shell stage. No on-site power-off tests have been performed. The targets in this article are suggested values for discussion among the owner, integrator, and contractor, not measured results or mandatory industry standards.
The first article in this series addresses wiring and capacity budgets; the second covers gateway selection. This article breaks “reliability” into requirements that can be recorded: what a failure affected, when devices recovered, and whether anything went wrong after recovery. The accompanying software calculations illustrate two pitfalls: two paths may share a single point of failure, and an impressive P95 may conceal many failures. All calculation data is artificially constructed; it is not test data from this villa.
Map Each Everyday Function as a Dependency Chain
Suppose a wall button’s action must pass through the cloud, then Home Assistant, before a Mesh gateway controls the light. Failure at any point will affect turning on the light. Even with a physical button on the wall, this cannot be called independent control.
An alternative is a panel that supports local binding and directly controls lights on the same Mesh network, while the central controller handles only coordination across systems. The distributed control capabilities described by the Bluetooth SIG provide a foundation for this architecture, but whether it works depends on the actual products and configuration.[1]
Acceptance requirements should therefore list functions first, then their dependencies:
| Everyday function | Local capability to retain | What may be temporarily lost |
|---|---|---|
| Turn on lights from a wall switch | Still works when the central controller is stopped | Phone-based scenes and remote control may be temporarily unavailable |
| Adjust air conditioning temperature | Manufacturer’s panel remains usable | Coordinated scheduling across rooms may stop |
| Control underfloor heating | Dedicated thermostats and necessary interlocks operate independently | Smart preheating strategies may stop |
| Record camera footage | NVR records independently while it and the camera network are operating normally | Video cards in the central controller may be temporarily unavailable |
| Smoke, gas, and similar alarms | Independent local alarms remain effective | Phone notifications may be delayed by network failures |
The table must also state its power assumptions. During a whole-home mains outage, lights without backup power will not stay on. “The lights still work when the central controller fails” and “the lights still work during a whole-home power outage” are two entirely different requirements.
Why Two Gateways Can Still Fail Together
A simplified model of a floor’s remote-control paths looks like this:
Path A: Central controller → Core switch → Floor switch → Gateway A → Target device
Path B: Central controller → Core switch → Floor switch → Gateway B → Target device
Assuming both paths have been configured and are actually usable, Path B may continue working when Gateway A alone is stopped. But if the floor switch fails, both paths are interrupted. This is a single point of failure caused by a shared dependency; adding more gateways does not solve it.
Mathematically, a function is available if at least one complete path survives; failure of any required component on a path breaks that path. The accompanying script calculates only this Boolean logic: a path remains after a single gateway fails, but none remains after the shared switch fails. It does not simulate wireless routing, reconnection time, or real-device takeover, and it cannot establish that the hardware has failover capability.
Acceptance therefore requires two diagrams: a communication dependency diagram and a power dependency diagram. Two switches plugged into the same power strip without backup power also share a power failure point. Two central controllers running on the same host machine share a host failure point. What matters is which failures they are independent under, not how many separate devices are visible.
Do Not Test Only an Internet Outage
Disconnecting the WAN does not prove the home can operate when the central controller fails. Stopping the central controller does not prove the network survives when a floor switch loses power. Acceptance should test at least the following separately:
- Disconnect the internet while keeping the local network and all devices powered; test local operation.
- Restore the internet, stop the central controller service, and test wall switches, manufacturer thermostats, and the NVR.
- Restore the central controller, stop one floor’s Mesh gateway, and record what is lost in lighting control and cross-floor scenes.
- Restore the Mesh gateway, shut down one Thread border router, and observe representative sensors and control devices.
- During an agreed test window, cut power to a floor switch and record the shared failure scope of PoE devices.
- As arranged by qualified personnel, test controlled power loss and restoration, checking for unintended actions, lost configuration, and recovery order.
Introduce only one failure at a time, and return the system to normal before the next test. HVAC protection, life-safety alarms, and work on energized equipment must not be tested by casually unplugging cables; the relevant professionals should follow the product procedures.
Before each test, fix a baseline covering models and firmware, network configuration, scene versions, device inventory, installation locations, power sources, and status-reporting intervals. After a failure, preserve logs first. Do not immediately restart the router, gateway, and central controller together: even if the system recovers, you will not know which action worked.
Write recovery steps into the plan as well. For example, restore switches and network services first, then start gateways and the central controller, and finally check device states and scheduled tasks. Adjust the exact order according to device documentation. The point is to follow the same procedure in each test and record deviations.
“Call Succeeded” Does Not Mean the Light Came On
For lighting control, record three moments: when the command was sent, when the platform received a status update, and when a change in the light was observed on-site. Recording only the platform’s execution trace may miss cases where a gateway receives a command but the downstream device does not execute it.
A preliminary target for common wall-switch operations could be that, under an agreed measurement method, 95% of responses take no more than 500 milliseconds; ordinary room scenes complete within two seconds. These are proposed experience targets to be agreed after testing a sample installation, not performance promises that can be applied directly to every device. For dimming, distinguish an intentionally configured transition time from network latency.
For each representative location, consider performing at least 100 controlled operations and recording failures, timeouts, and duplicate executions, rather than reporting only average response time. One failure in 100 operations can leave the average latency looking good. Even 100 successes do not imply a zero long-term failure rate; testing must also cover day and night, high concurrency, long-running operation, and recovery from abnormal conditions.
Do Not Let the “P95 of Successful Samples” Hide Failures
Suppose 100 operations are attempted: 90 complete in 200 milliseconds, while the target state is not observed within the two-second deadline for the other 10. If P95 is calculated only from successful samples, it is still 200 milliseconds, which could easily be described as “very fast response.” But the success rate within the deadline is only 90%.
This article suggests reporting four values together: total attempts, successes within the deadline, unconfirmed attempts, and P95 of successful samples. P95 uses the nearest-rank definition: sort successful samples by latency and take the sample at rank ceil(0.95 × number of successes). If there are no successful samples, P95 should display “no data,” not zero.
If the 10 unconfirmed samples are treated as infinite in the ranking of all attempts, the overall 95th percentile in this example is not finite. Because their actual completion times might fall after two seconds, a report should say “the overall P95 cannot be determined within two seconds,” rather than inventing a latency. The accompanying script retains both measures to prevent misinterpretation.
“Unconfirmed” does not necessarily mean the device did nothing: the light may have come on but its status report was lost, or a status report may have arrived while the light did not come on. Preserve on-site observations, platform reports, and timeouts separately; platform state cannot replace all other evidence.
| Recorded field | Purpose |
|---|---|
| run_id, device ID, configuration version | Tie results to the same test run |
| Failure condition and operation number | Distinguish normal tests, gateway outages, and recovery stages |
| Command time, on-site change time, status-report time | Measure action latency separately from feedback latency |
| Deadline, whether confirmed before deadline | Retain timed-out samples in statistics |
| Duplicate actions, unintended actions, recovery steps | Prevent errors during the process from being missed by checking only the final state |
When clocks in logs from different devices are not synchronized, directly subtracting their timestamps can produce false latency measurements. For the experience of lighting control, a single video can capture both the button press and the light, with timing measured frame by frame and the frame rate and error stated. For software timing on one machine, prefer a monotonic clock; use wall-clock timestamps in logs to align events. A phone stopwatch alone is unsuitable for proving targets measured in hundreds of milliseconds.
For a whole-floor scene, record when each light actually finishes and how much those completion times vary. Do not create extra load by polling every light at high frequency and then mistake the resulting congestion for inadequate coverage.
“Recovery time” also needs a defined endpoint. A gateway reappearing in the online list does not mean the entire room is usable. One possible agreement is that recovery for the test run is complete only after all tested device states have been reconfirmed, three consecutive basic operations have succeeded, and no old commands remain pending. Three operations here are an example acceptance procedure, not proof of a reliability rate.
Doing the Wrong Thing After Recovery Is Harder to Spot Than Staying Offline
Consider a specific scenario: a scheduled lights-off task for 22:00 does not complete during a network failure; an occupant manually turns the lights on at 22:03; the system recovers at 22:05. If the recovery program checks only that “the 22:00 lights-off task did not run,” it may turn off the lights the occupant just turned on.
Business rules must be written first: does an overdue task still make sense? Does a manual action cancel the old schedule? Can the control path identify the source of an action? What conservative policy applies if it cannot? There is no single answer suitable for every room. Bedrooms, corridors, and courtyards may use different rules, but they must be agreed before testing.
Add at least three recovery tests: recovery without human intervention; recovery after someone manually operates a device during the failure; and another disconnection during recovery. Record every physical action, not just the final state, so “turned off incorrectly, then turned back on automatically” is not counted as one success.
For motors, valves, and HVAC, do not treat resending a command as the default remedy. Repeating a state-setting command, a relative-movement command, or a toggle command has different consequences; whether a business operation is safe to retry must be checked against the interface semantics. This article specifies acceptance requirements only. Protection and interlocks for particular actuators remain governed by the manufacturer’s solution.
Check the UPS Along the Power Path
Putting a UPS in the main equipment cabinet does not mean floor-level PoE cameras have backup power. If they receive power from a floor switch, that switch must also be included in the backup-power design, or the inability to record in that area during a power outage must be accepted.
For example, at a critical load of 200 W and a desired runtime of 30 minutes, the load needs 100 Wh of energy. Assuming a conversion efficiency of 0.85, the battery must provide about 117.6 Wh of usable energy. This does not yet account for aging, temperature, or discharge characteristics; final sizing should use the UPS manufacturer’s runtime curve at the corresponding load.
Acceptance should record actual runtime, alarms, and shutdown behavior. Rated VA is one sizing parameter, not a measure of runtime. The NVR’s graceful shutdown and resumption of recording should also be verified separately.
The Acceptance Endpoint for Backups Is Restoration
“Backup task succeeded” proves only that some form of backup was produced. Central-controller data, Matter commissioning information, Thread network credentials, Mesh configuration databases, and manufacturer accounts may have different storage and restoration methods.[2][3]
Handover materials should therefore include at least device models and firmware, network topology, a cable and port schedule, configuration files, credential storage locations, and procedures for restoring service after replacing a device. Do not publish commissioning codes, access keys, or detailed diagrams of camera coverage in an article.
Conduct restoration rehearsals in an isolated environment or during a maintenance window to avoid starting two control systems that execute the same automations simultaneously. Record which data can be restored, which devices require reauthorization, and which functions still need the manufacturer’s app. A backup copy does not automatically provide high availability, nor does it mean devices can be restored without being reconnected.
Use Defect Severity to Decide Whether the Project Can Be Handed Over
I would divide results into three categories: failure of basic operations or protective functions; failure to meet the agreed primary experience; and incomplete supplementary functions. The first two cannot be closed out with only a promise to “upgrade the firmware later.” They need a defined fix, regression test, and responsible party. Whether supplementary functions may be deferred should be agreed in writing by both parties.
For example, if wall switches stop working when the central controller is shut down, that is a basic-function failure if the contract requires independent local operation. If a particular brand’s special lighting effect is not integrated, and it was explicitly excluded from the deliverables before procurement, it cannot later be treated as a defect. Acceptance must be measured against requirements, not revised after the fact to match the capabilities of the products already installed.
Retain at least three corresponding sets of handover materials: baseline configuration, raw records for each test run, and defect and retest records. Mark untested combinations as “not covered.” Passing 100 normal operations does not substitute for results from the farthest point in the basement or a whole-floor concurrent scene.
The final acceptance deliverable should be a table with recorded results: how the failure was injected, what was observed on-site, how long recovery took, and who is responsible for addressing it. Write “pending test” in untested cells; do not turn design targets into fictional results with check marks.
All actual-result fields in the accompanying acceptance template are blank and can be filled in during sample-installation testing. There is also a new per-operation record template, dependency and statistics calculation script, and software calculation results. This article follows the earlier articles on Matter state confirmation and Home Assistant execution after restart. Its focus is whole-home dependencies and handover acceptance, rather than repeating protocol implementation details.
Sources checked on 2026-11-15: 1 Bluetooth SIG Mesh, 2 Home Assistant Matter, 3 Home Assistant Thread. The failure-test procedures and experience thresholds are acceptance suggestions proposed in this article.
Series: Part 1: Wiring and Capacity Budget · Part 2: Mixed Protocols and Gateway Selection · Part 3: Failure and Recovery Acceptance
The templates and design documents in the attachments are in Chinese; the Python script depends only on the standard library. The software calculation results do not represent on-site measurements.
AI-generated scene illustration, not a photograph of this project or a wiring instruction.
Discussion
Share a question or your experience. Comments appear after moderation. Your email is never displayed publicly. Comment privacy
Loading comments…