Defining "Down": Writing an Outage Definition Everyone Accepts
The worst part of an outage is often the meeting afterward. The system is back, the adrenaline is gone, and five people are arguing about what just happened. Sales says the system was down for two hours, because it was unusable. You say forty minutes, because that is when it stopped responding entirely. Accounting says it does not count at all, because only their VLAN (the slice of the network their department sits on) was affected and they "made do." Someone asks whether it matters that half of it happened after 5 PM. Everyone is using the word "down," and nobody means the same thing by it. Whatever number ends up in the report will be remembered as either an excuse or an accusation.
I have sat in that meeting from every chair. I have also sat through the short version: a two-minute blip in the order system. Three of us stood in a corridor arguing about whether a two-minute blip was an outage. The blip ended before the argument did. None of those people were being unreasonable. They were answering different questions, because nobody had ever written down the question.
An outage definition is a short document, agreed in advance, that settles what counts as down, how badly, for whom, and exactly when the clock starts and stops. This guide walks through the arguments the definition must settle and gives you a complete template you can adapt in an afternoon. It classifies the classic borderline cases. It also shows how to wire the definition into your monitoring so the clock runs on evidence instead of memory.
In this guide
- The arguments a definition prevents
- Five dimensions every definition must answer
- The template, ready to adapt
- Borderline cases, classified
- Degraded states and partial credit
- Getting it signed before it is needed
- Wiring the definition into monitoring
The arguments a definition prevents
Every organization without a written definition has the same three fights, on repeat. Knowing them in advance is useful, because your definition is finished exactly when it can settle all three without a meeting.
"It was slow — that counts!" The system technically responded to every request, at a pace that made real work impossible. To the person clicking, that is downtime with extra frustration. To the person reading a monitoring dashboard that only measures dead-or-alive, it never happened. Both are looking at real evidence; the definition decides how slowness is scored, and by how much.
"Only accounting was affected." A switch failure takes one department offline while everyone else works normally. Is that an outage of the network? A partial outage? A footnote? Suppose seven people out of a hundred and twenty lost everything for an hour. Someone has to decide whether the availability number moves — and by a lot or a little. Scope is the dimension people argue about most bitterly, because it decides whose bad morning officially existed.
"It was after hours." The file server was down from 9 PM to midnight. If nobody was supposed to be using it, most teams agree that should not count the same as three business hours. But "most teams agree" is not a rule. The one time it happens during a payroll deadline weekend, the informal understanding evaporates. Service hours have to be written down per service, along with what happens in the gray zones.
There are two quieter fights behind those three. First, when did it start? At the first user report, the first monitoring alert, or the moment reconstructed later from logs? Second, when did it end? When the fix was applied, or when someone verified users could actually work again? Ten minutes of clock ambiguity on each end of every incident adds up to an availability number nobody can reproduce. And a number nobody can reproduce is a number nobody trusts.
Five dimensions every definition must answer
Strip away the formatting and every workable outage definition answers the same five questions. If your draft answers all five with numbers instead of adjectives, it will survive contact with a real incident. "Significantly impacted" is not a threshold. It is a fight waiting for a calendar invite.
- Scope: which services are covered, and how much impact makes an incident count — how many users, which sites, which functions. A definition that covers "the network" as one blob will be re-argued every time; scope lives per service.
- Severity levels: the difference between hard down, degraded, and a single-user problem, with a concrete, measurable threshold for each. That lets two people reading the same incident reach the same level.
- Service hours: when each service is promised. A 24×7 service counts every minute; an 8-to-6 service counts business minutes, with named exceptions for the weeks when after-hours suddenly matters.
- Planned vs. unplanned: what earns the "planned" label — typically advance notice of a stated length, an approved window, and user communication — and how planned time is reported. Work that misses any requirement is unplanned by definition, which is what keeps the label honest.
- Clock rules: exactly what event starts the outage clock, what event stops it, and how the times are established when detection lagged reality.
A practical note on scope, because it is the dimension that looks easy and is not. You need a way to count affected users that works at 2 AM with partial information. Head-counting during an incident is fantasy, so scope thresholds should map to things you can name instantly — a site, a VLAN, a department, "everyone behind switch B." If your org chart says accounting is 30 of 120 people, then "accounting is out" is 25% by construction. Nobody has to survey the building to classify the incident. Define scope in units your network diagram already has.
Notice what is not on the list: blame, root cause, and whose fault the internet provider is. An outage definition measures impact, not guilt. Keeping fault out of the document is half the reason people will agree to it; the other half is that it is short.
The template, ready to adapt
Here is a complete definition you can adapt. It fits on two pages on purpose — a definition nobody reads is a definition nobody agreed to. Numbers in it are sensible starting points for a small or mid-size organization, not laws of physics: change them to fit, but change them to other numbers, never to adjectives.
SERVICE AVAILABILITY AND OUTAGE DEFINITION (v1.0) Covers: order system, email, file service, phones, internet access Owner: IT. Signed by: IT lead, operations manager, finance lead. Review: annually, and after any incident whose classification was disputed.
Section 1 — Severity levels. Every incident gets exactly one severity per service affected, decided by the thresholds below, not by how loud the reporting was:
| Level | Name | Threshold (any one qualifies) | Counts toward availability as |
|---|---|---|---|
| S1 | Hard down | Primary function fails for 25% or more of the service's users, or for an entire site or department, or for a function with no workaround | 100% of elapsed service-hours time |
| S2 | Degraded | Primary function works but key operations take more than 3× their normal time for 10+ minutes, or more than 1 attempt in 10 fails, affecting 25% or more of users | 50% of elapsed time (see Section 4) |
| S3 | Limited | Fewer than 25% of users affected and a documented workaround exists, or any single-user issue | 0% — tracked as tickets, not downtime |
| P | Planned | Work inside an approved window, announced at least 5 business days ahead through the standard channel | Excluded from the unplanned number; reported separately every time |
Section 2 — Clock rules. The outage clock is a machine, not a memory. It starts and stops on events, and "I rebooted it" is a hopeful act, not an ending:
- Rule 1 — Start: the clock starts at the earlier of two events. These are (a) the monitoring system's first failed check for the incident, or (b) the first user report later verified as this incident. If logs later prove impact began earlier, the log time wins — evidence beats detection.
- Rule 2 — End: the clock stops when the service is verified working. The same check that failed now passes, or a user-level test succeeds. Applying the fix does not stop the clock.
- Rule 3 — Relapse: if the same incident recurs within 30 minutes of verification, it is one incident and the clock includes the gap. This kills the temptation to declare five short outages instead of one long one.
- Rule 4 — No minimum: if it met a severity threshold, it counts, however short. Blips below your monitoring's ability to measure are handled by the measurement rules in honest availability math, not by pretending.
- Rule 5 — Service hours: each covered service lists its hours next to its name (for example: order system 24×7; file service Mon–Fri 07:00–19:00). Outside listed hours the incident is logged and severity-classified but accrues no downtime — unless a published exception (payroll week, month-end close) was in effect.
Section 3 — What is out of scope. Single-workstation problems, individual account lockouts, and third-party cloud services you only consume (tracked separately, since you can measure them but not fix them). Out of scope does not mean unimportant; it means measured elsewhere.
To see the clock rules run, walk one incident through them. The order system starts failing quietly at 08:41 — nobody knows this yet. Monitoring's transaction check fails at 08:47. The first user ticket lands at 09:12. The fix is applied at 09:55, and the failed check passes again at 10:03. Under Rule 1 the clock starts at 08:47 (first failed check). When the application log later shows errors beginning at 08:41, the start moves there — evidence beats detection. Under Rule 2 the outage ends at 10:03, not 09:55, because "fix applied" is not "verified working." Total: 82 minutes of S1, and every number in that sentence came from a timestamp rather than an argument. That is the entire trick: the meeting afterward reads the ledger instead of writing it.
Tip: resist the urge to add severities. Three levels plus "planned" settles almost every argument; seven levels creates a new argument about levels. If two people can honestly disagree about whether an incident is S1 or S2 under your thresholds, fix the threshold — do not add an S1.5.
Borderline cases, classified
A definition earns its keep on the ugly cases, so classify the classics in the document itself — people remember worked examples far better than rules. Here is how the template above scores the four incidents most likely to cause a meeting:
| Scenario | Classification | Why |
|---|---|---|
| Everything works but takes 5× normal, all users, for 20 minutes | S2 degraded, 20 min at 50% weight | Over the 3×-for-10-minutes threshold for more than 25% of users; it worked, but not usefully |
| One department's VLAN (30 of 120 users) loses everything for 40 minutes | S1 hard down, weighted by the 25% of users affected | An entire department qualifies as S1; the availability math then weights the 40 minutes by scope |
| Primary firewall dies; standby takes over in seconds; users notice nothing | No outage — logged as a component incident | The service never failed; the definition measures user impact, and the near miss goes in the risk log, not the downtime ledger |
| Saturday 3-hour outage of the Mon–Fri file service | Logged, zero downtime accrued | Outside service hours with no exception in effect — but if it is still broken Monday 07:00, the clock starts then |
| Announced 2-hour maintenance window runs 40 minutes long | First 2 hours planned; the overrun is unplanned S1 | The window was the promise; exceeding it is exactly the failure planned-work rules exist to catch |
The third row is the one that surprises people, in a good way. Redundancy that works means the outage did not happen as far as availability is concerned. Saying so explicitly is what makes engineers willing to invest in redundancy. The component failure still gets recorded, because a spent spare is risk you are now carrying, but it belongs in a different ledger than user-facing downtime.
Degraded states and partial credit
Slow-but-alive is the hardest state to score, and there are only three honest ways to do it. Strict binary: degraded counts as fully down. It is the simplest possible bookkeeping, with no arguments about weights. But it punishes a 20-minute slowdown exactly like a 20-minute blackout. That quietly teaches everyone to argue that slowness "wasn't really S2." Partial credit: degraded time counts at a fixed fraction, like the 50% in the template. Pragmatic and proportionate. The fraction is admittedly arbitrary, which is fine as long as it was agreed in advance. An arbitrary rule applied consistently beats a fair-sounding rule negotiated per incident. Separate lanes: report availability from hard outages only, with degraded minutes reported alongside as their own number. Most honest and most work, and it only functions if the degraded number actually gets reported every time rather than quietly dropped when it is embarrassing.
Any of the three can be defended. What cannot be defended is deciding after the incident, when everyone already knows whose number benefits from which rule. Pick one, write it into Section 1, and let honest availability math handle the arithmetic consequences. That guide shows the same month computed under different policies, which is a useful thing to show your stakeholders when you pick.
Getting it signed before it is needed
The definition's power comes entirely from when it was agreed. Before an incident, severity thresholds are abstract and nobody's reputation is attached. People negotiate in good faith because they cannot know which side of a future argument they will be on. After an incident, every clause has a dollar value and a name attached, and you are not writing a definition anymore — you are litigating one. Agreement before the next incident is not a nice-to-have; it is the entire point of the document.
Harbourview Clinic wrote their first definition the week after their patient-records system spent a Monday morning answering every request in forty seconds. Reception called it down. IT called it slow, because every request did, eventually, come back. The practice manager, who had to report the number, asked which it was. The three of them spent the following week writing a definition. Every threshold had been chosen with that particular Monday in mind. The next incident was a Saturday, affected one consulting room, and matched nothing in the document. The second draft was written in a quiet month by people who could no longer remember whose side they had been on. And it has held ever since.
Who signs? Three roles, kept small on purpose. IT — whoever runs the monitoring and reports the numbers — because they have to be able to measure every clause. A business owner per major service — the person who feels the outage — because their signature is what stops "it was slow, that counts" from being relitigated. Whoever the availability numbers are reported to — a manager, director, or owner — also signs. The numbers only carry weight if their reader accepted the rules that produced them. Walk each signer through the borderline-case table, ask "would you accept this scoring on a bad day?", and adjust until yes. Then set a review date: annually, plus any time a real incident's classification was disputed. A dispute means the definition has a hole, and holes get patched in calm weather.
Keep the signed copy somewhere it can be produced in thirty seconds, alongside the numbers it governs. The uptime evidence file is built for exactly this. A definition nobody can find is a definition nobody agreed to, retroactively.
Wiring the definition into monitoring
A definition on paper and a monitoring system that measures something else will drift apart by the second incident. The fix is a deliberate mapping. Every severity threshold in the document should correspond to a check and an alert rule in the monitoring system. So the clock runs on timestamps instead of recollection.
The mapping leans on the layered checks from ping-up vs. service-up, because the check ladder and the severity ladder line up naturally. A check that fails is a start time, a check that passes again is an end time, and neither has any optimism in it:
- S1 evidence: the service's truth-tier check — content, transaction, or protocol — failing repeatedly, or failing for the monitoring probes that represent whole sites. The first failed check in the sequence is your Rule 1 start time.
- S2 evidence: the same checks passing but breaching their response-time thresholds — which is why every check needs a slow threshold, not just a dead one. Set the alert to require the definition's duration (10 sustained minutes, in the template) so the monitoring system itself applies the S2 rule.
- S3 evidence: single-host or single-rung failures with everything user-facing still green — a ticket, not a clock.
- Recovery evidence: Rule 2's "verified working" is simply the failed check passing again — automatic verification, with a timestamp.
Two practical notes. First, your monitoring's check interval bounds how precisely the clock can run. A five-minute polling cycle cannot produce a to-the-second start time, and your reports should be honest about that granularity. Second, planned windows should be entered into the monitoring system as maintenance periods before the work starts. Then planned minutes tag themselves and nobody sorts alerts by hand afterward. The discipline around those windows is covered in planned downtime.
Where to go next
Write the two pages, score the borderline cases, collect three signatures, and map the thresholds to checks. The next incident then produces a classification instead of a meeting. And the number in the report is one that everybody in the room agreed to before they knew whether they would like it. The corridor argument ends too, because the answer to "was that an outage?" is on page one, and has been since a quiet afternoon when nobody minded. A natural companion to this guide is honest availability math, which turns classified incidents into a defensible percentage. See the rest of the Uptime & Availability library for targets, maintenance windows, and keeping the evidence organized.
Put this into practice with SysPeep Monitor
An outage definition needs a clock it can trust. SysPeep Monitor gives you one. Its service checks have thresholds you can map straight onto your severity levels. Timestamped state changes provide start and end times. Dependency-aware alerting makes the record show causes instead of echoes. The free trial installs on any Windows machine in minutes.
Download Free TrialLearn About SysPeep