Honest Availability Math
Two admins can take the same month, the same outages, and the same calculator. They can walk into the same meeting with different availability numbers, both produced in good faith. One counted the maintenance window and one excluded it. One charged the full ninety minutes of a branch-office outage and one weighted it by the third of the company it touched. Neither is lying yet. The lying starts when the choices are invisible: when a number arrives with no footnotes, and every judgment call inside it silently happened to favor the person presenting.
I have produced that number. The first availability figure I ever reported went up by half a percentage point between the draft and the meeting, and nothing on the network had changed. I had changed how I counted, and I had not written down that I had. Nobody caught it, which was the problem. A number that goes up when you change the counting is not a measurement of the service. It is a measurement of the counting.
Availability arithmetic is division a ten-year-old could do. What makes it hard is that the two inputs, "downtime" and "total time," each hide half a dozen policy decisions. Every decision can be gamed, fumbled, or defended, depending on whether it was made in advance and written down. This guide walks every place the number bends. It covers maintenance windows, partial outages, degraded performance, gaps in your own measurement, and the resolution limits of polling. It also covers multi-service rollups and the difference between a component failing and users suffering. Then it computes one messy month end to end, two defensible ways, so you can see exactly where the choices live.
In this guide
- The formula, and what it hides
- Maintenance windows: show both numbers
- Partial outages: weighting without cheating
- Degraded performance: is slow down?
- Measurement gaps: unknown is not up
- What your polling interval lets you claim
- Rollups and redundancy
- A month, computed two honest ways
- Footnotes beat exclusions
The formula, and what it hides
The base formula is the one everyone knows:
availability = (total time - downtime) / total time x 100 Example month (30 days): 43,200 minutes total One 47-minute outage: (43,200 - 47) / 43,200 = 99.89%
Clean, until you poke either input. Total time: total according to whom — every minute on the calendar, or only the hours the service is promised? Does an announced maintenance window belong in the denominator? Downtime: down according to what — a failed ping, a failed transaction, a flood of complaints? Does slow count? Does an outage that hit one department out of four count in full, in part, or not at all? And what about the twenty-five minutes when the monitoring server itself was rebooting and nobody was measuring anything?
None of those questions have a single correct answer. They have policy answers — choices your organization makes once, in writing, before the incidents happen. The what-counts-as-down half of the policy is the subject of defining "down". This guide takes those classifications as given and handles the arithmetic half. It explains how each policy choice changes the number, and how to present the result so it survives hostile questioning. The one rule that spans everything: a choice disclosed is a methodology; a choice hidden is a thumb on the scale.
Even the measurement window is a choice worth stating. A monthly number is volatile — one bad afternoon dents it visibly. That is exactly why it is the honest reporting unit: problems surface in the month they happened. An annual number smooths that same afternoon into a rounding error, which makes it flattering and slow to sound alarms. The defensible pattern is to compute monthly, then let the annual figure be the aggregation of twelve months everyone has already seen. An annual number that arrives as a surprise is an annual number someone was saving up.
Maintenance windows: show both numbers
Planned maintenance is the biggest single fork in the math. Suppose one four-hour patch window (240 minutes) in an otherwise perfect 30-day month. Two computations:
Counting the window: (43,200 - 240) / 43,200 = 99.44%
Excluding the window: (43,200 - 240 - 0) / (43,200 - 240) = 100%
^ removed from downtime ^ and from total
Note how exclusion works: the window leaves both the numerator and the denominator. You are not claiming the service was up during the window. You are declaring those minutes outside the promise, the way a shop is not "closed" at 3 AM. It was never open Tuesdays at 3 AM to begin with.
Excluding is legitimate under exactly two conditions, both boring and both mandatory. A written policy that predates the window said planned maintenance is excluded. And users were actually told in advance through the normal channel. Miss either one and the exclusion is retroactive storytelling. The number's most popular inflation technique is quietly reclassifying Tuesday's surprise outage as "emergency maintenance" on Friday. The criteria for what counts as properly planned — notice period, approved window, communication — belong in your outage definition. And the discipline of running windows well is its own guide: planned downtime. The presentation rule is simpler than the policy: report both numbers. "100% against our maintenance policy; 99.44% wall-clock" is one line. It answers the skeptic before they ask. And it costs you nothing because the second number was one subtraction away anyway.
Partial outages: weighting without cheating
A switch failure knocks one department — 30 of 120 users — off everything for 60 minutes. Everyone else works normally. Three ways to score it:
Ignore it ("most users were fine"): 43,200 / 43,200 = 100%
Full charge ("an outage is an outage"): (43,200 - 60) / 43,200 = 99.86%
User-weighted (30/120 = 25% affected): 60 x 0.25 = 15 min charged
(43,200 - 15) / 43,200 = 99.965%
Ignoring it is indefensible — thirty people lost an hour, and "the number didn't move" is exactly the kind of claim that destroys trust in every future number. Full charge is defensible but blunt: it scores a one-department outage identically to a whole-company blackout. Over time, that makes the metric useless for distinguishing bad days from disasters. Weighting is the honest middle: multiply the duration by the fraction of users affected. Where user counts mislead, use the fraction of business function affected. An order-entry outage matters more per head than an archive-server outage.
Weighting carries its own honesty rules, because it is also the number's favorite place to hide. The weights must be defined before the incident — in your outage definition's severity and scope rules — never negotiated afterward, when everyone knows which weight produces which result. Use units you can count at 2 AM: a site, a VLAN (one segment of the network), or a department roster. State the weight in the report ("90 min weighted at 33%: 40 of 120 users"). Never weight the same incident twice. A degraded state that was already discounted to 50% does not also get a "but only some users cared" haircut. One weighting per event, stated in the footnote.
Degraded performance: is slow down?
The slow-is-down question is a policy fork covered properly in defining "down", so here is just the arithmetic consequence. Take one 35-minute period where the system worked at a crawl. Under a strict policy, it is 35 minutes of downtime. Under a 50%-credit policy, it is 17.5 minutes. Under a separate-lanes policy it is 0 minutes of downtime plus a second reported figure: "35 degraded minutes." Same event, three numbers — 35, 17.5, 0 + disclosure — and all three are defensible if the policy was written down first and the report says which one is in force. What is not defensible is choosing per incident. If this month's slowdown got 50% credit and last month's identical slowdown got ignored, you do not have a methodology; you have a mood.
Measurement gaps: unknown is not up
Sooner or later, the monitoring system itself goes down — a reboot for patches, a full disk, a crashed probe. When it comes back, the availability graph shows a smooth green line across minutes nobody actually observed. The graph looks like "up" because up is what the absence of recorded failures looks like. This is the subtlest inflation in the whole subject: nobody chose to lie; the tooling's default just assumed health during its own blackout. It is a first cousin of the "dashboard is green, so everything is fine" fallacy from the monitoring myths guide — green pixels are only as good as the measuring behind them.
There are three honest ways to book unmeasured time, in descending order of conservatism:
- Count it as down. The strictest floor: "we cannot prove it was up, so it does not count as up." Use this when the number carries real weight — commitments, disputes, anything contractual.
- Report it as unknown, separately. "99.91% measured, over the 99.4% of the month we were measuring; 25 minutes unmeasured." Most precise, slightly more work to present, and completely audit-proof.
- Corroborate and reclassify. Use secondary evidence — application logs, server event logs, the absence of a single ticket — to argue the service was healthy during the gap. And write down what evidence you used. Acceptable for internal reporting; weakest of the three, because "no one complained" is testimony, not measurement.
The one option that is never on the list is the silent default: counting unknown as up because the graph happened to render it that way. The strict option has a pleasant side effect, too: it makes monitoring outages expensive, which gets them fixed. And since gaps are inevitable, monitor the monitor. Even the simple tricks in the fundamentals guides shrink the unknown minutes you will one day have to account for. One example is a heartbeat alert whose absence means trouble.
What your polling interval lets you claim
Your checks are not a continuous gaze; they are snapshots. A check every 5 minutes samples the service 8,640 times in a 30-day month, and between snapshots anything can happen unrecorded. A 3-minute outage that falls entirely between two polls simply never existed as far as your data is concerned. One that straddles a single failed poll gets booked at your tool's resolution — roughly one interval — whether it really lasted 90 seconds or four and a half minutes. The error runs both directions, and it is bounded by the interval. Your availability number has error bars about the size of your polling cycle, and claims finer than that are decoration.
Now put that against the tiers from the nines. Four nines monthly allows 4 minutes 19 seconds of downtime — less than one 5-minute polling interval. With 5-minute polls, you cannot demonstrate 99.99% for a month. A single missed-then-caught blip is already the size of the whole budget, and the blips you missed entirely are unknowable. The practical rules: poll your truth-tier checks on important services every minute or faster. These are the content or transaction checks that prove the service works, not merely that the box answers. Require consecutive failures before declaring down. But book the outage from the first failed poll, not from when the confirmation finished. Round your public claims to what the instrument can support. A 5-minute poll supports "99.9% ± a few minutes." Saying so out loud is not weakness — it is the sound of someone who knows how their own measurement works.
Rollups and redundancy
Two last distortions live at the reporting layer, after the per-service math is done.
The averaged rollup. Ten services; nine ran a flawless month; the order system ran 92%. The "overall availability" average is 99.2% — a number that looks like a decent quarter and describes a disaster. Averaging across services launders your worst number in your best ones. The more trivia you monitor, the better the laundering. Add enough lobby TVs and test boxes at 100% and no outage can dent the average again. Report per-service numbers for the services people actually depend on — the per-service targets you set in internal uptime targets give you the list. If management insists on one company number, give the honest pair: the average and the worst important service, side by side.
Component vs. user-experienced availability. Your firewall pair had a terrible month: the primary died and sat dead for six hours before replacement. The service — internet access — never blinked, because the standby took over in seconds. Which number is "availability"? Both, for different audiences. The component number (node down 6 of 720 hours: 99.17%) belongs to engineering — it drives repair, spares, and the risk conversation about running on one node. The user-experienced number (100%) is the availability of record, and you should say it plainly. A redundant component failing without user impact is not downtime. It is redundancy doing its job. Reporting user-experienced availability as 100% with a note is what makes the redundancy budget defensible next year. Log the near miss where such things live — the uptime evidence file — and resist anyone who wants to average the two numbers into something meaning nothing.
A month, computed two honest ways
Now the whole method on one messy, realistic month for one service. The events, as classified under the outage definition:
| # | Event | Raw duration | Policy facts |
|---|---|---|---|
| 1 | Patch window, Sunday 02:00–04:00 | 120 min | Announced 7 days ahead; policy allows exclusion |
| 2 | Core switch failure, Tuesday morning | 47 min | All users; unplanned hard down |
| 3 | Branch WAN outage | 90 min | 40 of 120 users (weight 33%) |
| 4 | Degraded period (3× slower) | 35 min | Definition scores degraded at 50% |
| 5 | Monitoring server offline for patching | 25 min | No corroborating evidence either way |
Method A — wall-clock, everything counts. The conservative floor: calendar denominator, every event charged in full, unknown counted as down.
Downtime = 120 + 47 + 90 + 35 + 25 = 317 min Availability = (43,200 - 317) / 43,200 = 99.27%
Method B — policy basis. The window is excluded because policy and notice both check out; the branch outage is user-weighted; degraded gets its defined 50%. The measurement gap still counts as down, because unknown is not up:
Total = 43,200 - 120 (planned, excluded) = 43,080 min
Downtime = 47 + 90 x (40/120) + 35 x 0.5 + 25
= 47 + 30 + 17.5 + 25 = 119.5 min
Availability = (43,080 - 119.5) / 43,080 = 99.72%
99.27% and 99.72% — same month, same facts, both defensible. Every step from one to the other is a stated policy applied consistently. There is the window exclusion (with its notice), the 33% weight (with its user counts), and the 50% degraded credit (from the signed definition). The gap is charged as down under both methods. Consider a third number with no stated method: say, a suspiciously round 99.9% built by quietly dropping events 3, 4, and 5. That is the only wrong answer on the table. The gap between A and B is not spin. It is the measured size of your policy choices, and showing both is how you prove it.
Footnotes beat exclusions
Which brings us to presentation, where honesty is mostly formatting. The strongest possible way to report availability is a headline number with its working attached:
Order system, this month: 99.72% (unplanned, user-weighted) 1. Planned window (120 min) excluded per maintenance policy; announced in advance. 2. Branch outage (90 min) weighted at 33% of users (40 of 120). 3. Degraded period (35 min) counted at 50% per outage definition v1.0. 4. Monitoring gap (25 min) counted as downtime: unknown is not up. Wall-clock, everything-counts basis: 99.27%.
Six lines. Every judgment call visible, every skeptic's question pre-answered, and the conservative number sitting right there proving you have nothing buried. Compare the alternative: a bare "99.9%" that later turns out to have eaten a reclassified outage. Remember that these numbers compound like credit. The first inflated figure someone catches retroactively discounts every number you have ever reported and every one you ever will. Footnotes are cheap. Rebuilding trust is not. Keep the reports, the policies they cite, and the raw incident records together in one place. That is the uptime evidence file, and it is where a defensible number goes to stay defensible.
Tollgate Foods reported their warehouse system at 99.9% for eleven months in a row. In the twelfth month a new finance director asked a polite question. How had a system she had personally been unable to use for most of a Thursday scored 99.9% again? The answer was assembled over an uncomfortable week. The Thursday had been reclassified as "emergency maintenance" the following Monday, as had two earlier Thursdays. Nobody had written down a rule under which that was allowed. The number for month twelve was restated at 99.4%, and at her request so were the previous eleven. The warehouse system had not got worse. The counting had got honest, and every report before it was now read with a pencil.
Tip: do the arithmetic within a day or two of each incident, while the timestamps and scope are fresh, and file the per-incident line then. Month-end math becomes addition instead of archaeology — and numbers computed close to the event are far harder to argue with later than numbers reconstructed under deadline.
Where to go next
You now have the whole honest pipeline: a formula and the six places it bends. You also have a presentation format that turns every bend into a footnote instead of a secret. Use it and your number will sometimes go down, which is the first sign that it has started measuring the service instead of the counting. Upstream of the math sit the classifications that feed it — defining "down" — and the checks that generate the timestamps: ping-up vs. service-up. The rest of the Uptime & Availability guides cover the targets, the maintenance discipline, and the evidence file where all of this lives.
Put this into practice with SysPeep Monitor
Honest math needs raw material: timestamps, check results, and history you did not have to reconstruct. SysPeep Monitor collects it continuously: device and service checks over SNMP, WMI, and SSH, with dashboards and reports in one web console. So when someone asks for the number, you can show the working behind it. The free trial installs on any Windows machine in minutes.
Download Free TrialLearn About SysPeep