DDoS test abort criteria are the control that stops a DDoS test the moment it starts to harm something it was not authorised to harm. Whether they still fire when the network is loaded, and how fast every generator goes quiet once they do, is the one safety property a test has to prove about itself before it proves anything about the target.
What an abort criterion is made of
The plan-level basics are settled: the criteria are agreed before the window opens, and they are numbers a monitor can evaluate rather than a feeling. How they sit next to the kill switch and the go/no-go gate is laid out in the abort section of a test plan, so this page covers what makes a single criterion work.
Each one has four parts:
- A signal. The thing being watched, named precisely: the error rate of real users on one path, the CPU of one shared firewall.
- A threshold. The value at which the signal counts as harm.
- A dwell time. How long the breach must hold before it fires. Too short and one noisy sample ends a good test; too long and the harm runs on while the timer counts.
- A vantage point. Where the signal is read from, and that place must sit outside the path the test is loading.
Abort criteria are not failure criteria
The easiest mistake is to write the target's own failure as a reason to stop. In a DDoS stress test the break is the result being looked for, and the synthetic user traffic that falls when the target gives way is the measurement, not the alarm.
Abort criteria watch what sits outside the experiment:
- Real users beyond the agreed budget, on paths that were never meant to be touched or on a production path past the degradation that was signed off.
- Systems outside scope that share something with the target: a firewall, a NAT gateway, a transit link, a DNS provider. These are where a test's blast radius shows itself first.
- A mitigation acting wider than the scope, such as a provider placing a null route on the whole prefix when the test aimed at one address.
- The test's own faults: a generator above its rate cap, or traffic to an address that is not on the list.
Write the target's failure as an abort and every stress test ends at its first knee. Leave the collateral off and an outage can run while the test reports success.
Losing sight is a criterion of its own
Monitoring that shares a link or a collector with the target can go dark exactly when the link fills. A dashboard that stopped updating looks calm.
So a working set always includes one more rule: if a watched signal returns no data for longer than its dwell time, the test stops. Blind counts as breached.
Why the stop has to be fast and unarguable
Anyone named may stop it; restarting needs agreement
The plan names a short list of people, on both the asset owner's side and the test side, and any one of them can end the test without asking anyone. Stopping needs no consent. Starting again does: the cause is understood first, and the test resumes at a lower step than the one that tripped.
Stop latency has three parts
The time from harm to silence is the sum of how long it takes to see the breach (the sampling interval plus the dwell), how long it takes to decide, and how long the traffic takes to drain from every generator and queue. Automated criteria cut the middle term to nearly nothing. Only rehearsal shows the third.
Harm can also outlast the traffic. A mitigation that engaged during the test, an automatic block list or an upstream null route, may stay in place after the generators stop, so the criteria name who can ask for it to be lifted and how.
What a DDoS test records about its own stop
The stop is rehearsed before the window, at low rate, and timed from command to the last packet. A test in which no criterion fired has not shown that the criteria work; only the rehearsal has.
When one does fire, the DDoS test report records which criterion it was, the reading and the time, the drain time to zero, and what was left unrun because of it. Vectors the stop prevented are "not reached", never "held". The criterion that fired is a finding in its own right, because it tells you which shared component the next test, or the next real attack, will lean on first.
Choosing the numbers themselves, and tying each one to the success criteria it protects, belongs in the DDoS test plan, written while nobody is under pressure.