Every solution review with an owner eventually hits the same question: "What happens when this gateway fails?" The previous article covered power and lightning protection — that is about making the gateway live long. This one covers the other half: if it still dies, what happens to the network? The answer is not "gateways are reliable so don't worry." It is an often-underestimated property of LoRaWAN architecture: adding a second gateway adds redundancy, and end devices don't need to change anything. But when is the second gateway worth it, how should it be deployed, and what changes in the network afterwards — these deserve calculation before purchase orders.
Before discussing redundancy, answer a prior question: how reliable can a single gateway be? Engineering uses availability:
Availability = MTBF ÷ (MTBF + MTTR)
MTBF is mean time between failures; MTTR is mean time to repair. The formula hides a lever people overlook: MTTR affects availability as much as MTBF does. A gateway with a five-year MTBF still looks bad if a failure means three days waiting for a spare plus two hours of travel; conversely, a deployment with spares on the shelf and a ten-minute swap keeps decent availability even with average hardware.
Converting to annual downtime makes it concrete: 99% availability means about 87.6 hours of downtime per year — over three and a half days. 99.9% is 8.76 hours. 99.99% is about 53 minutes. Do not casually write "four nines" into a quotation — that is a 53-minutes-per-year tier, which a single box plus fast response rarely sustains; it usually takes a redundant architecture.
So the first judgment is not "one gateway or two" but: how much downtime can your business tolerate. Temperature sensors reporting every 15 minutes where half a day of outage is acceptable — one gateway plus a spare on the shelf is enough. Fire-linked controls, production lines billed by the hour of downtime, unattended remote sites — those enter the redundancy conversation.
Engineers arriving from Zigbee or NB-IoT instinctively ask "how does handover work?" The answer: there is no handover. LoRaWAN is a star-of-stars broadcast architecture — when a device transmits an uplink, it names no gateway. Every gateway within radio range receives the same packet, each one timestamps it and attaches RSSI/SNR metadata, and each forwards it over its own IP backhaul to the network server.
The network server then does three things:
The direct conclusion: adding a gateway to the network improves coverage, redundancy, and capacity simultaneously, with zero device-side changes. Devices store no "my gateway is X" state. Kill one gateway and devices keep broadcasting; surviving gateways keep receiving. No reconnection, no rejoin, no site visit. For operations, this is LoRaWAN's most practical advantage over mesh protocols.
Adding a gateway is one kind of redundancy — not the only layer, and not always the right one. A complete path has three failure points, hence three redundancy layers:
The most common engineering mistake is spending at the wrong layer: three-tier backhaul on a building whose single gateway takes the whole floor down when it dies; or three gateways for coverage that all share one switch and one mains feed — the bottom-layer single point remains, and the upper-layer redundancy is decoration. Before buying the second gateway, walk through power and backhaul first.
Before installing the second unit, know what it brings and what it alters.
Uplink: pure gain. Devices in the overlap zone go from "reachable by one gateway" to "reachable by two," and packet delivery improves markedly — for edge points whose signal sat on the availability red line, this is a second life. There is no extra airtime cost: the device still transmits once; channel occupancy is unchanged.
Downlink: routing logic, with boundaries. The NS routes downlinks through the best gateway, but note two things. First, NS-to-gateway backhaul latency must stay within the receive-window timing budget — if the round trip exceeds roughly one second, downlink windows start being missed, a familiar problem from the antenna and backhaul articles wearing a new redundancy hat. Second, a downlink always goes out from exactly one gateway: two gateways do not add transmit power, and downlink coverage does not double.
ADR re-converges on its own, but takes time. Devices that sat under gateway A at high data rates will naturally fall back when A dies — lower rates, longer airtime, trading throughput for reliable contact with gateway B. The fallback is automatic; the cost is partial packet loss during convergence and a rise in network-average airtime. What you observe in a drill is exactly this convergence process, not the crude question of whether devices are still online.
Deduplication and capacity. Every uplink in the overlap zone gets forwarded by multiple gateways; NS-side dedup consumes a little compute and the backhaul carries a few extra copies — negligible in practice, but if your NS runs on an edge box in a TKE-class deployment, factor the dedup overhead into capacity planning at device scale (the headroom logic mirrors the timeseries sizing from the data pipeline article).
"Install two" comes in two topologies, an order of magnitude apart in cost-effectiveness.
1+1 hot standby at the same point: both gateways mounted at the same location (or very close), one takes over instantly when the other fails. The upside is zero-perception failover with unchanged coverage; the downside is that this is the worst value for money — the two devices' footprints almost completely overlap, the second contributes nearly nothing while both are healthy, and it is pure insurance. Only one scenario truly fits: a single point holds the entire business (one well house, one equipment room), and downtime costs far exceed the hardware.
Mutual backup across zones (the recommended default): two gateways at two locations, each owning its primary area, with a planned overlap band at the edges. In normal operation this is just a standard two-gateway deployment — coverage and capacity both double. Under failure, the survivor pulls the neighbor's edge devices back at lower data rates — degraded, not interrupted. In effect you buy redundancy as capacity: the same two devices create coverage value every day and only switch to the insurance role during failures. For most "do we need redundancy" questions, this is the correct answer.
One design point for mutual backup: plan the overlap band deliberately; do not let it form by accident. The site survey (article 36) must verify that "with either gateway powered off, the other can still cover the neighbor's edge points at a workable data rate" — write that into the survey report so acceptance has teeth. Accidental overlap is usually enough RSSI and not enough SNR; when a failure actually happens, the band cannot rescue anything.
Redundancy only counts once it has been drilled. At acceptance, do three things deliberately:
Pull one. Cut power to one gateway and observe: whether edge devices keep reporting (some loss expected), how device online/offline status moves in the NS (mind the three online-status definitions from the troubleshooting article — don't mix them), and how many uplink cycles ADR needs to converge. Restore power and watch the fall-back.
Check downlink routing. Push one RPC or configuration change to a device in the overlap zone and confirm the downlink leaves from the expected gateway and arrives — the trigger-rules article stressed that "receiving uplinks does not prove downlinks work"; redundancy adds a routing variable to that.
Document it. The site survey article covered installation records; redundancy adds three items: the coverage responsibility split between the two gateways, the overlap-band point list, and the drill log (when it was pulled, how many packets dropped, how long convergence took). Six months later, that log decides whether operations is tracing a map or rediscovering the network.
One platform-side note: gateways join ThinkLink via the gateway management whitelist — register the second gateway the same way, or you will discover during the drill that the NS does not recognize it at all. Device offline determination follows the platform's heartbeat rules; configure alert thresholds against your business tolerance.
Add a second gateway when any of these holds:
Skip it when:
One closing principle: redundancy does not eliminate failures; it demotes their cost from "business outage" to "performance degradation." In a mutual-backup design, the second gateway works every day and insures on the bad day — that is what spending in the right place looks like.