Setting Internal Uptime Targets You Can Actually Hit
At some point somebody, a manager, an auditor, a form from an insurance company, will ask you: "what's our uptime target?" There is a powerful temptation to answer with whatever number sounds professional. 99.9% has a nice ring to it. The trouble is that a target is not a mood. It is a promise, and 99.9% is a very specific promise. Measured around the clock, it allows fewer than nine hours of total downtime per year. That covers every failure your systems, your vendors, and your one pair of hands can produce.
I have heard "we need five nines" said, warmly and sincerely, about an environment with one server, one internet line, and one admin. The admin was me. The person saying it had heard the phrase somewhere and liked the sound of it, and that was not their fault. Nobody had yet shown them what the phrase costs, and the person who should have was busy nodding.
Here is the principle this whole guide hangs on: an uptime target must follow from your architecture, your staffing, and your budget, never from aspiration. The network with one internet line, one firewall, and one admin who sleeps at night has a ceiling on what it can honestly promise. No amount of wanting changes that ceiling. What does change it is money and structure: a second line, a second box, a second person. So target-setting is really two separate conversations: "what can we hit today?" and "what would it cost to hit more?" Confusing them is how IT teams end up quietly promising things they cannot deliver.
This guide walks you through the honest version. It gives you a capability audit you can run in an afternoon and a worked example of turning the audit into a defensible number. It gives you a tiering scheme so different services get different promises. It also shows how to price the next tier, so "we need more nines" becomes a budget line item instead of a guilt trip. At the end there is a fill-in worksheet to make it concrete.
In this guide
- A target is a promise, not a wish
- The capability audit: four questions
- From audit to a defensible number
- Tiering: different services, different promises
- Pricing the next nine
- When to revisit targets
- The burnout math
- The worksheet
A target is a promise, not a wish
First, vocabulary. An uptime target is a stated percentage of time a service is supposed to be working, over a defined measurement window. In larger organizations you will hear SLO, service level objective, or SLA, service level agreement, when it is written into a contract. Every part of the target definition carries weight. "Working" needs a definition. A server that answers ping while its application is dead is not working. That is why this pillar has a whole guide on defining "down". And the measurement window changes the promise enormously: 99.5% measured around the clock is a different world from 99.5% measured during business hours.
Why does an inflated target matter, if it is only internal? Because targets drive expectations, and missed expectations compound. Promise 99.9%, deliver 99.5%, and you have taught management that IT misses its numbers — even though 99.5% might be genuinely excellent for your architecture. The same performance under an honest target of 99.5% is a promise kept, quarter after quarter. Kept promises are the currency you will spend later when you need budget. There is a strange freedom in a modest number you always hit. Nobody remembers the ambition of the target; everybody remembers whether you made it.
Tip: never write a target as a bare percentage. "99.5%" means nothing; "99.5% during business hours, 7:00–19:00 weekdays, per the outage definition" is a promise someone can check. A target without a measurement window attached is an argument waiting to happen.
The capability audit: four questions
Before any number gets written down, audit what your environment can actually do. Four questions, in order: what can break alone, how fast do you find out, how fast can someone act, and how fast can it be made whole. Be ruthless — this document is for you, not for show.
1. The single-point-of-failure inventory
A single point of failure (SPOF) is any component whose lone failure takes a service down — no backup, no twin, no workaround. Walk the stack and count. The honest answers are usually humbling:
| Layer | The question | If the answer is "one" |
|---|---|---|
| Internet | How many lines into the building? | Any ISP incident is a full outage for everything cloud |
| Edge | One firewall? | A single power supply decides your year |
| Core | One core switch? | Same story, one layer down |
| Servers | One host running all the VMs? | One hardware fault takes every virtual machine on it |
| Power | One feed, one UPS? | Grid blips reach the servers directly |
| People | One person with the admin passwords? | Their vacation extends every outage |
| Support | One lapsed vendor contract? | Parts arrive on the vendor's schedule, not yours |
Notice the last two rows. People and paperwork are infrastructure. The admin who is the only one who knows the firewall password is exactly as much a SPOF as the firewall — and considerably harder to keep in stock.
2. Detection speed: how fast do you know?
Downtime starts when the service breaks, not when you notice — and the gap between those two moments is pure, unrecoverable loss. Trace your own path honestly. Suppose checks run every five minutes and your alert rule waits for two consecutive failures before firing (a sensible anti-noise setting). Up to ten minutes pass before the system even decides something is wrong. Then the alert has to reach a human: an email lands — are you at your desk, or in a meeting? During business hours, ten to twenty-five minutes from failure to aware is a realistic honest answer for an email-alerted shop. Off-hours is where the truth gets stark. If nobody reads alerts overnight, an outage at 9 p.m. is discovered at 8 a.m. That is eleven hours of downtime before anyone lifts a finger. That single fact decides more about your achievable target than any hardware. And detection quality matters as much as speed. A check ladder that verifies the actual service, not just ping, is what makes "we know" mean something — see Ping-Up vs. Service-Up. If you have no systematic checks at all yet, start with Network Monitoring Fundamentals before setting any target — you cannot promise what you cannot see.
3. Response speed: who can act, from where, how fast?
Knowing is not fixing. Who is allowed and able to act on an alert — and from where? Can the firewall be administered remotely, or does someone drive in? Suppose remote access rides the very VPN (the encrypted tunnel you use to reach the office from outside) that just died. Your response path then has a circular dependency (a surprisingly common discovery). How long is the drive? Who covers when the primary person is sick, on leave, or at a conference? Is that coverage real — does the backup person have current passwords and have they actually done the procedure before? A one-person shop's honest answer is often: thirty to sixty minutes to hands-on during business hours, and "tomorrow" outside them. Write that down without shame; it is the truth the target must be built on.
4. Recovery capability: how fast can it be made whole?
Once hands are on the problem, what bounds the fix? Spares: is there a switch, a power supply, an access point on a shelf, or does a dead device mean a purchase order? Backups: they exist, but have they been restored from lately — and how long does a full restore of the file server actually take? (If you have never timed it, the honest audit answer is "unknown," which for planning purposes means "long.") Vendor contracts: read the real terms. "Next business day" means a part arrives tomorrow at best — a Friday afternoon hardware failure is bridged Monday. A four-hour contract means something entirely different and costs accordingly. Your recovery capability is the slowest of these links, not the average.
From audit to a defensible number
Now convert the audit into arithmetic. Use business-hours measurement for the worked example: 7:00–19:00 weekdays is 3,120 hours — 187,200 minutes — per year. The downtime allowance at each target level (full tables and what each tier demands live in The Nines, Translated):
| Target (business hours) | Allowed downtime per year | Per quarter |
|---|---|---|
| 99.9% | 187 min (about 3.1 h) | about 47 min |
| 99.5% | 936 min (about 15.6 h) | about 3.9 h |
| 99% | 1,872 min (about 31 h) | about 7.8 h |
| 95% | 9,360 min (156 h) | about 39 h |
Next, sketch a realistic year for a typical small-shop stack. It has a single ISP (internet service provider), a single firewall, one virtualization host plus an aging physical file server. It has monitoring with email alerts, and one admin with a helpdesk colleague. Not a disaster year; a normal one. A couple of ISP incidents that overlap business hours, say two and a half hours in total. One hardware failure bridged on next-business-day parts: roughly six business hours down. Two or three software or configuration incidents at half an hour to an hour each: call it two hours. One mystery that takes an afternoon to diagnose: ninety minutes of user-visible impact. Total: around twelve hours of business-hours downtime, from ordinary events, none of them negligent.
Hold that twelve hours against the table. It fits inside the 99.5% allowance of 15.6 hours — with margin, but not lavish margin. It blows through the 99.9% allowance of 3.1 hours before the first quarter is out; the ISP incidents alone spend most of it. The conclusion writes itself: this stack supports a 99.5% business-hours target, and no amount of effort makes it a 99.9% stack. The failure classes that dominate (ISP outage, dead hardware waiting for parts) are architectural, not effort-shaped. So write down 99.5%. It will feel underwhelming compared to the brochure numbers. It has the advantage of being true. At year's end you can compute your achieved availability honestly (method in Honest Availability Math) and stand next to the result.
Tiering: different services, different promises
One target for everything is a mistake in both directions — it overspends on the trivial and shortchanges the vital. Sort services into classes and promise each class only what it needs. Three tiers cover a small shop; here is a sensible starting scheme:
| Class | What belongs here | Target (business hours) | Allowed downtime / year | Response commitment |
|---|---|---|---|---|
| Critical | Internet access, sign-in/authentication, the main line-of-business app, email path | 99.5% | ~15.6 h | Drop everything; work begins within 15 min |
| Important | File shares, Wi-Fi, intranet tools, printing servers | 99% | ~31 h | Work begins within the hour |
| Best effort | Individual printers, meeting-room gear, test systems, signage | 95% or no formal target | ~156 h | Queued; next business day is fine |
Two notes on using the table. First, the response commitment column is a promise about people, and it must survive contact with the staffing audit. A fifteen-minute response commitment on a day the only admin is at a dentist appointment is fiction. So state it as "during staffed hours" if that is the truth. Second, put the tier list in front of management and let them argue with the sorting, not the physics. If someone insists the signage screens are critical, that is a genuinely useful conversation to have now, at classification time, rather than mid-outage. Classification disagreements are cheap on Tuesday and expensive during an incident.
Pricing the next nine
Sooner or later someone will look at your honest 99.5% and say the magic words: "we need more nines." This is not an attack; it is an opening. The capability audit already told you exactly which failure classes dominate your downtime, so raising the target is not a motivational exercise — it is a shopping list. Price it:
| What you buy | Failure class it removes | Kind of cost |
|---|---|---|
| Second internet line, automatic failover | Multi-hour ISP outages | Recurring circuit + failover-capable edge config |
| High-availability firewall pair | Edge hardware death | Hardware + licensing |
| Second virtualization host with replication | "One host died, everything died" | Hardware + licensing + real complexity |
| Cold spares on the shelf (switch, PSU, AP) | Waiting for parts | Modest, one-time |
| Four-hour vendor hardware contract | Next-business-day waits | Recurring contract |
| On-call rotation (needs two-plus people) | Overnight and weekend blindness | Staffing — the expensive one |
| Tighter detection: one-minute checks, alerts that reach a phone | Tens of minutes of not-knowing | Nearly free — configuration and habit |
Present it exactly like that, attached to the specific outages in your history. "Our two longest outages were the ISP and the file server host; here is what removing each failure class costs." Now "more nines" is a line item with a price, and the decision belongs where it should — with the people who control the budget. If they fund it, the target rises honestly, after the new capability is installed and proven. If they decline, the current target stands, documented as a choice someone made with the price in view. Either outcome is healthy. The unhealthy outcome is the target rising while the architecture stays flat — that gap gets paid for in weekends, and eventually in credibility. Notice the last row, though: detection speed is the one upgrade that is nearly free. Before anyone buys hardware, make sure you are not donating half an hour of every outage to "nobody noticed yet."
When to revisit targets
Targets are not tattoos. Review them on a calendar — once a year is plenty. Re-run the capability audit after anything that changes the answers. That includes a second line installed, a host added, a person hired or departed, or a vendor contract lapsed. The departure one is easy to miss: when the second admin leaves, your response capability just changed. The honest move is to say so out loud rather than silently absorbing the old promise into one person's evenings. A big incident is also a natural checkpoint — not to slash the target in a panic, but to ask whether the audit assumptions held. If the outage revealed a SPOF you had not inventoried, the audit gets a new row before the target gets a new number.
The burnout math
The most common failure in small-shop target setting is not a wrong number on paper. It is the quiet, undocumented promise of enterprise nines with a one-person team. Run that math with a straight face. A 24/7 availability promise carries an implied 24/7 response capability. A year contains 8,760 hours. One person "informally" covering all of them is on call every evening, every weekend, every holiday, every night of every vacation they answer the phone during. Which is to say, it is not coverage; it is a person slowly being sanded down. And here is the perverse part: heroics hide the problem. When the lone admin drives in twice at 3 a.m. and the quarterly number comes out at 99.8%, management learns that 99.8% is what the current budget buys. It is not. It is what the current budget plus an unsustainable donation of one human's life buys. And the invoice for that donation arrives later, as burnout or a resignation letter — either of which, per the SPOF table above, is itself an availability event.
Bellweather Logistics wrote "99.9%, 24/7" for their dispatch system into an internal standards document, and for three quarters they hit it. The system ran on one host behind one firewall. And the number was met because the one admin drove in at three in the morning four times and never put any of it in a report. Management learned, reasonably, that 99.9% was what the current budget bought. In the fourth quarter the admin took a job with a company that staffed the role with two people. The next quarter came in at 98.7% on the same hardware, the same firewall, and the same internet line. The hardware had not changed. The target was reset to 99.5% during business hours and a second person was hired, which is roughly what the first number had been costing all along.
Tip: an honest target protects people, not just numbers. Writing "99.5% business hours, best effort outside them" is not lowering the bar — it is stating what is actually covered. That makes raising real coverage a staffing decision someone can fund instead of a sacrifice someone silently makes.
The worksheet
Pull it all together in one table — one row per service, five columns. This is the artifact worth an afternoon: it captures the audit, states the honest target, and prices every gap, which means it doubles as your next budget conversation. A filled example for the small shop above:
| Service | Current capability (audit notes) | Honest target | Gap | Cost to close |
|---|---|---|---|---|
| Internet access | Single ISP; detection ~10 min; no failover possible | 99.5% business hours | Multi-hour ISP outage possible any quarter | Second line + failover-capable edge |
| Sign-in / authentication | Two domain controllers — but both on one host | 99.5% business hours | Host failure takes both at once | Move one DC to separate hardware |
| File server | Single host; restore tested twice a year; NBD parts | 99% business hours | Hardware failure means a day-plus down | Four-hour contract + shelf-spare PSU |
| Printing | Assorted devices, no spares; a failure is tolerated | Best effort | None accepted | Nothing — and say so in writing |
Copy the columns, fill a row per service, and be exactly as unflattering as the audit demands. The rows where the gap column stings are not embarrassments. They are your pre-written answers for the next time someone asks why the network went down. They are also your pre-priced requests for the next time there is money to spend. The printing row is worth copying as written: no spares, no tears if one dies, and a line saying so, which turns neglect into a decision.
Where to go next
The method in one breath: audit what can break alone, how fast you know, how fast you act, and how fast you recover. Derive the number the architecture supports and tier it per service class. Price every step up, and re-audit when reality changes. A target built this way is modest, defensible, and, the part that matters, hittable, which over a few quarters buys you the credibility that inflated numbers never do. The next time someone says "five nines" warmly in a meeting, you will have a worksheet to hand them instead of a nod. The neighboring guides carry it on: what the tiers cost, what counts as down, and how to compute achieved availability to compare with the target. They are all collected at Uptime & Availability.
Put this into practice with SysPeep Monitor
Half of every honest target is detection speed - and that half is cheap to fix. SysPeep Monitor auto-discovers your devices, checks them over SNMP, WMI, and SSH, and alerts you the moment something fails. So the clock between broken and known stays short. Dashboards and reports show exactly where you stand against the target you set. The free trial installs on a Windows machine in minutes.
Download Free TrialLearn About SysPeep