When the appliance in your own data centre can still see the flood, but the internet link in front of it is already full, nothing it does reaches the problem. That is the gap hybrid DDoS mitigation exists to close: an on-premises mitigation layer handles what fits inside your links, and cloud scrubbing takes over for what does not. It is an architecture decision, and the part a DDoS test spends its time on is the handover between the two layers, because that is where a hybrid design breaks.
How the two layers split the work
The on-premises layer is always in the path. It sees every packet, keeps state, and handles application-layer floods and anything below the link's capacity, with no diversion delay. Its ceiling is the uplink: an appliance cannot drop what never arrives.
The cloud layer is usually on demand. A scrubbing center takes the traffic by BGP announcement or DNS change, filters the volume, and returns the clean remainder over a tunnel or interconnect. That is the always-on versus on-demand choice made per layer, not once for the whole service.
What joins them is a signal. The appliance, watching the link fill, tells the provider to start scrubbing. That can be a person on the phone, a provider API call, or a standard protocol: the IETF's DDoS Open Threat Signaling (DOTS) was written for exactly this, with a signal channel designed to keep working over a congested path.
Where the handover fails
The signal leaves too late. If escalation is set close to the link's capacity, the link saturates before the provider acts, and the signal itself competes with the flood for the same pipe. A signal sent over a dedicated or out-of-band path avoids the second problem; only a lower threshold fixes the first.
The appliance blocks the provider. Where the cloud layer proxies connections rather than routing packets, traffic arrives from the provider's addresses instead of the users', so the appliance sees far fewer sources carrying far more traffic each. Per-source rate limits tuned for the open internet can then fire on the provider's own ranges and drop the clean traffic the cloud just handed back.
The appliance loses sight of the attack. After diversion it sees only scrubbed traffic, so by its own readings the attack is over. If it is allowed to signal "clear", the diversion is withdrawn into a flood that never stopped, and the design flaps. The end of an attack has to be judged from the cloud side's telemetry.
Return traffic and state disagree. The clean-pipe return path brings its own tunnel overhead and routing asymmetry, which the mitigation cutover page covers. In a hybrid design those faults land on a stateful appliance that was built expecting the old path.
What a DDoS test checks
A hybrid test runs in two bands. Below the escalation threshold, it confirms the appliance handles the flood alone and does not call the cloud, since a design that escalates every small attack pays the diversion cost for nothing. Above it, the test times the chain on one clock: appliance detection, signal sent, provider acknowledgement, traffic arriving at the scrubber, clean traffic reaching the servers.
Then it checks the failures above. Did the signal get out while the link was loaded? Did the appliance pass the returned traffic, or rate-limit the provider? Who decided the attack had ended, and on what data? Did the diversion withdraw once, or flap?
Driving traffic past your own uplink's capacity also loads your carrier's links, so that band needs the carrier's authorisation and is often run with the provider's involvement, rather than as a surprise to either. Measuring the appliance against the load it can actually see, and finding the link ceiling that sets the escalation point, is the on-premises half of the job that on-premise DDoS testing works through.