Site Health Rollups: One Number per Site Without Hiding Problems
With one site, your monitoring screen is a list of devices, and a list is fine. With twenty sites, or twenty clients, the list is six hundred rows. Nobody can read six hundred rows at eight in the morning and answer the only question anyone is asking: "is anything wrong, and where?" I have watched a good technician try. He scrolled for a minute, said "looks fine", and missed the dead branch on row four hundred.
The answer is the rollup: one status per site that summarizes everything inside it. Twenty tiles instead of six hundred rows. Green means leave it alone; anything else means click. That is the entire idea, and in most tools it takes about an hour to set up.
It takes about the same hour to set up badly. A bad rollup fails in one of two directions. It shouts about everything, turning a whole site red because one printer is out of toner. Eventually people learn that red is a colour the wall just is. Or it hides problems, averaging a dead core switch into a comfortable "94% healthy" because the thirty desktops behind it are fine. This guide shows you how to design a rollup that does neither. It covers what counts, how much it counts, and what happens when the link to the site is down. It explains which summary method to use and how to make sure every rollup opens into its reason.
In this guide
- Why rollups exist, and how they go wrong
- What counts toward site health, and what does not
- Dependency awareness: the WAN goes down once
- Worst-component-wins versus weighted score
- The rollup states
- The rollup rules table
- A worked example
- Drill-down: the rollup must open into its reason
- Client-facing versus internal site dashboards
- Where to go next
Why rollups exist, and how they go wrong
A rollup is a summary status computed from the statuses of the things inside it. Most tools give you some version of it. That might be a group whose color is derived from its members, or a "site" object. It might be a map node that turns red when anything under it does. The purpose is attention management. A human watching a wall of sites needs to know which one to look at, in priority order, without reading the details of any of them.
The two failure modes come from the two lazy ways to compute a summary. "Any red makes it red" is honest but indiscriminate. A printer, a test VM, and the core switch all have the same power to turn the site red. So the site is red most of the time and red stops meaning anything. "Average everything" is calm but dishonest. A site with fifty devices and one dead firewall reports "98% healthy," and the one thing that matters is invisible until a user calls. Neither is wrong arithmetic; both are wrong design. The fix is not a cleverer formula but a clearer decision about what a site's health actually means — and that starts with membership.
What counts toward site health, and what does not
Ask a simple question of every device at the site: if this failed, would the site's users be unable to work? The devices that pass that test are the site's health components. Everything else is monitored, graphed, and alerted on in its own right, but it does not move the site's rollup.
The WAN link usually counts (the circuit connecting the site to the world, and to you). So do the router or firewall it terminates on and the core switch that everything plugs into. The servers the site depends on to work usually count (a domain controller, a file server, a local application server). The phone system or voice gateway usually counts. So does Wi-Fi as a group when the site is wireless-first. Individual printers and individual access points usually do not count (one dead AP is a dead zone, not a dead site). Neither do workstations, test machines, and anything a user would describe as an annoyance rather than an outage. The printer still gets its own alert if you want one. It simply does not get a vote on whether the site is up.
Membership alone fixes most of the "everything is red" problem. The second lever is weighting by criticality: not every component that counts, counts equally. A tiered scheme works well and maps naturally onto the way you already think about severity. The reasoning behind tiers and levels is laid out in Severity Design, and a site rollup simply reuses it.
- Tier 1 — site-defining. If it is down, the site is down: WAN link, firewall, core switch. One failure here decides the rollup outright.
- Tier 2 — service-defining. If it is down, the site is degraded: a key server, the voice gateway, the Wi-Fi group as a whole. One failure makes the site amber; several can make it red.
- Tier 3 — noticed but not site-affecting. Individual APs, secondary switches on a single floor, the UPS's self-test. These are visible in drill-down and may nudge a health score, but they cannot on their own change the site's color.
- Untiered. Printers, desktops, lab gear. Monitored, never rolled up.
Dependency awareness: the WAN goes down once
Here is the trap that catches almost every first rollup. The WAN link to a branch fails. From your central platform, every device behind that link stops answering. If your rollup counts each of those non-answers as a failure, the branch reports "1 WAN link down, 1 firewall down, 1 switch down, 3 servers down, 12 APs down". That means eighteen failures, eighteen alerts, and a red tile that looks like a fire when it is actually one cut cable. Worse, if you use a weighted score, eighteen failures at once will drag the score to the floor whether or not the devices are genuinely down. And you have no idea which.
The rollup needs to understand dependency: that the reachability of everything at the site depends on the link and the router in front of it. When the parent is down, the children are not "down"; they are unknown. You cannot see them, which is a different fact from them having failed. A dependency-aware rollup counts the WAN failure exactly once, marks the children as unreachable rather than failed, and suppresses their individual alerts. I built the eighteen-alert version once; the alerts arrived on my phone in the order the poller found them, which was not the useful order. The general habit of mapping what depends on what is covered in Dependency Thinking. The branch-specific version is worked through in Remote Site WAN Monitoring. It covers telling a dead circuit from a dead site from a power failure, and where a probe inside the site changes the picture.
A probe at the site, as described in MSP-Style Monitoring Architectures, changes the rollup in an important way. When the WAN is down, the probe keeps polling locally and buffers the results. The rollup at that moment can only say "unknown because unreachable." But when the link returns and the probe forwards its backlog, history is retroactively filled in. The site's rollup for that hour becomes "WAN down, everything else up." That distinction matters enormously when you are writing the month's availability report, which is exactly what the reporting guide in this pillar is about.
Worst-component-wins versus weighted score
With membership, tiers, and dependencies sorted, you can choose how to combine components into one status. There are two honest methods, and each misleads in a predictable way.
Worst component wins
The site's state is the state of its worst member. Down anywhere means down; degraded anywhere means degraded. It is simple, explainable, and impossible to fool with averaging: if the firewall is down, the site is down, no matter how happy the servers look. Where it misleads is the middle. With sloppy membership, a single tier-2 or tier-3 component keeps the site permanently "degraded". A site that is amber for a month is a site people stop looking at. Worst-wins depends entirely on disciplined membership. Applied to tier 1 and tier 2 only, with tier 3 excluded, it is the most trustworthy method there is.
Weighted score
Each component gets a weight, and the site's health is the weighted fraction that is healthy. Thresholds turn the number into a color: above one line is healthy, between lines is degraded, and below is down. It expresses partial failure well — two of six access points down really is a different situation from five of six. It also lets you show a client a number over time. Where it misleads is at the top: unless tier-1 weights are enormous, the score can stay comfortable while the one thing that matters is dead. A score also invites people to argue about the weights rather than the outage. (They will bring a spreadsheet.)
The hybrid most teams settle on
Use worst-wins for tier 1, and a weighted score for tiers 2 and 3. A tier-1 failure forces the site to "down" regardless of the score (call this the floor rule). With tier 1 healthy, the score decides between healthy and degraded. In that case, a low enough score — say most tier-2 components gone — can still push the site to down. You get the honesty of worst-wins where it matters and the nuance of a score where partial failure is real.
The rollup states
Resist the temptation to have two colors. A site needs at least four states, and the fourth is the one most tools forget. Each state should have a one-line definition that anyone on the team could recite.
| State | Definition | What the viewer should do |
|---|---|---|
| Healthy | All tier-1 components up; tier-2 score above the healthy line | Nothing |
| Degraded | All tier-1 up; one or more tier-2 components down, or the score below the healthy line | Open the site; work it in business hours unless a client-facing service is affected |
| Down | Any tier-1 component down, or the score below the down line | Open the site now; this is an incident |
| Unknown (unreachable) | The path to the site (WAN, router, or probe) is down; child states cannot be observed | Treat as an incident on the path, not on the children; expect history to fill in later |
| Maintenance | A scheduled window is active; failures are recorded but do not change the color or page anyone | Nothing, unless the window overruns |
"Unknown" deserves its own color and its own wording. When a site is unreachable, saying "down" is a guess. A guess in the wrong direction produces a very awkward client conversation: "your dashboard said our servers were down for three hours, but they were fine." Say what you know. "Unreachable since 09:14; last known state healthy" is true, useful, and calm.
The guess in the other direction is quieter and worse. Ashgrove Housing ran eleven neighbourhood offices from one platform with a probe in each. The platform had no "unknown" state; a silent probe simply left its tile at the last colour it had reported. The probe at the Millbrook office lost power when a contractor needed its socket. The Millbrook tile stayed green for eight days, through a switch failure that took the office's phones down for a morning. The failure was reported by the office ringing the helpdesk. The review noted that the dashboard had been "consistently green throughout the period". Millbrook's tile now turns grey after three missed probe heartbeats, and the socket has a label.
The rollup rules table
Write the rules down before you configure them, in a table that a new technician can read. A rule that exists only inside the tool's configuration is a rule nobody can review, and a rule nobody can review drifts. Here is a template for a typical branch; adjust the members and weights to what the site actually depends on.
| Component | Tier | Depends on | Effect on rollup when down | Weight (if scored) |
|---|---|---|---|---|
| WAN link | 1 | — | Site becomes Unknown (unreachable); children suppressed | n/a |
| Firewall / router | 1 | WAN link | Site Down (floor rule) | n/a |
| Core switch | 1 | Firewall | Site Down (floor rule) | n/a |
| Domain controller | 2 | Core switch | Degraded; contributes to score | 3 |
| File / application server | 2 | Core switch | Degraded; contributes to score | 3 |
| Voice gateway | 2 | Core switch | Degraded; contributes to score | 2 |
| Wi-Fi (group of APs) | 2 | Core switch | Degraded when the group is below half; score by fraction up | 2 |
| Single access point | 3 | Wi-Fi group | No color change; visible in drill-down | 0.25 each, via the group |
| Floor switch | 3 | Core switch | No color change; visible in drill-down | 0.5 |
| UPS | 3 | — | No color change; "on battery" raises its own alert | 0.5 |
| Printers | — | Floor switch | None; monitored separately | 0 |
Two details in the table repay attention. The WAN link's effect is "Unknown," not "Down": you have lost sight, not confirmed failure. And the Wi-Fi group is a component made of components. That is how you stop one access point from moving the site while still noticing when half the floor has gone quiet. Most tools let you nest groups like this; if yours does not, treat the wireless controller as the tier-2 component and the individual APs as tier 3.
Tip: a rollup rule is a claim about what the site depends on, and sites change. Every time a server is added, a service moves to the cloud, or a second circuit is installed, revisit the site's rules. A rollup that still treats a decommissioned server as tier 2 will hold the site at "degraded" forever. The team will learn to ignore the tile.
A worked example
Take an imaginary branch — call it Riverside — with the components in the table above. It has one WAN link, one firewall, one core switch, a domain controller, a file server, and a voice gateway. It has six access points, two floor switches, one UPS, and three printers. The UPS is an uninterruptible power supply — the battery that carries the rack through a power cut. Using the hybrid method, tier 1 is worst-wins with the floor rule. The tier-2 score is the healthy weight divided by the total weight, with degraded below 90% and down below 50%. Here is how five ordinary mornings play out.
Scenario 1: one printer offline Tier 1 ......... all up Tier 2 score ... 10 / 10 = 100% Rollup ......... HEALTHY (printer alert exists, but has no vote) Scenario 2: two of six access points down Tier 1 ......... all up Wi-Fi group .... 4 of 6 up = 67% of its weight 2 = 1.33 Tier 2 score ... (3 + 3 + 2 + 1.33) / 10 = 93% Rollup ......... HEALTHY, drill-down shows "Wi-Fi: 4 of 6 APs up" (four APs down would score 87%: DEGRADED) Scenario 3: file server down Tier 1 ......... all up Tier 2 score ... (3 + 0 + 2 + 2) / 10 = 70% Rollup ......... DEGRADED "File server down since 07:52" Scenario 4: WAN link down Tier 1 ......... WAN link unreachable Children ....... suppressed, marked unreachable Rollup ......... UNKNOWN "Unreachable since 09:14; last known HEALTHY" One alert ...... WAN link, not eighteen Scenario 5: core switch down, WAN up Tier 1 ......... core switch down -> floor rule Rollup ......... DOWN "Core switch down since 11:03" Children ....... behind the switch marked unreachable, not down
The printer never touches the tile. Two dead access points are visible but do not alarm anyone; four would. A dead file server is honestly amber, with the reason in the tile. A cut circuit produces one alert and an honest "unknown". When Riverside's probe forwards its buffered results an hour later, the history for that hour shows the servers were fine. And a dead core switch is red no matter what the score says, because the floor rule exists precisely so that a score cannot hide it. Five scenarios, five different readings, none of them a lie.
Drill-down: the rollup must open into its reason
A rollup that cannot explain itself is a rumor. The moment a tile turns amber, the viewer's next question is "why?" And the answer must be one click away, not a hunt through a device list. Design the drill-down path deliberately.
- The tile carries a reason line. Not just a color: "Degraded — file server down since 07:52" or "Unknown — unreachable since 09:14." The reason is the worst contributing component and when it changed.
- One click opens the component list, sorted by tier and state. Tier 1 first, failures at the top, unreachable items visibly distinct from failed ones. A viewer should see the cause in the first row.
- One more click opens the failing check. The graph, the last successful poll, the error text, the dependency it sits behind. This is where a technician starts working.
- The path back up is obvious. Breadcrumbs or a persistent site header, so nobody gets lost three levels deep during an incident.
Test the drill-down the way you would test an alert: break something at a site. Time how long it takes a colleague who did not break it to find the cause starting from the wall of tiles. If it takes longer than a minute, the rollup is hiding its reason, and hiding the reason is the same failure as hiding the problem.
Client-facing versus internal site dashboards
The same rollup serves two audiences, and they should not see the same screen. The internal dashboard is the wall of tiles: every site, every client, tiers and reasons visible. It is sorted by state so that anything not green floats to the top. It is a working tool, and it may be busy. The client-facing site dashboard is a storefront: one site (or one client's sites), in their words. It shows the status of the services they care about, the reason line if anything is degraded, and the recent availability figure. There are no tiers, no internal notes, no other clients, and no scoring arithmetic. A client does not want to know their site scored 83%. They want to know that the Wi-Fi on the second floor is weak and you are on it.
Two rules keep the two views honest. First, they must be derived from the same rollup. If the client's view says healthy while your internal tile says degraded, someone will eventually notice. Then the conversation will be about trust rather than access points. Second, the client view should use the "unknown" state generously and clearly. "We cannot currently reach your site; last confirmed healthy at 09:14" is a message a client can act on. They can look out of the window and check whether the power is on. "Down" is a message they will repeat to their boss, and you will spend the afternoon un-saying it. The mechanics of building scoped, read-only client views are part of tenant separation, which has its own guide in this pillar.
Put the site's rollup state on the network map as well as on the tile wall, if your tool draws maps. A map with twenty sites colored by rollup state is the fastest way anyone will ever get the "where is the problem" answer. It is the view a manager or a client understands without a legend.
Where to go next
A good rollup is a small piece of design that makes the rest of multi-site monitoring possible. Honest membership. Tiers instead of equal votes. Dependencies, so the WAN fails once. A hybrid of worst-wins and scoring, five states including an honest "unknown," and a drill-down that opens into the reason. With that in place, twenty sites fit on one screen, the screen can be trusted, and the technician on row four hundred gets to stop scrolling. The other guides in Monitoring Across Multiple Sites and Clients cover the architecture that feeds the rollup and the tenant separation that keeps client dashboards clean. They also cover the probe fleet that keeps each site's data flowing and the reporting that turns rollup history into evidence.
Put this into practice with SysPeep Monitor
The rollup design in this guide needs a tool that understands dependencies and can draw the result. SysPeep Monitor evaluates thresholds and dependencies centrally so a failed WAN link does not become eighteen alerts. Its live dashboards and network maps show site status from one web console, with remote probes keeping each site's data flowing. The free trial runs on any Windows machine.
Download Free TrialLearn About SysPeep