A Working Glossary of Monitoring Terms (With Examples)
This is the short glossary: 24 terms, each with an example. For the full A-to-Z reference of 152 terms, from bandwidth and jitter to NetFlow, syslog and IPAM, see Network Monitoring Terms Defined.
Open any monitoring tool's documentation and you meet a wall of words it assumes you already know: pollers, thresholds, flapping, rollups, acknowledgments. None of these ideas is hard. Most are five-minute concepts wearing intimidating names. But nobody ever sits you down and explains them, so they get absorbed slowly, by osmosis and embarrassment, over your first year on the job. I learned "hysteresis" from a settings page at eleven at night, with the alerts still coming in.
This page is the sit-down. Here are twenty-four terms, each with a plain definition and a short "in real life" scenario. The scenario is what makes a definition stick, and the definition on its own rarely does. The terms are grouped into four families that mirror what a monitoring system actually does. It collects numbers, judges them, alerts people, and keeps data. So the glossary doubles as a map of the machine. Read it straight through in twenty minutes, or keep it open beside your tool's settings pages and raid it as needed. Nobody is checking which.
In this guide
- The four families at a glance
- Collecting: how numbers arrive
- Judging: how numbers become verdicts
- Alerting: how verdicts reach people
- Data: what the tool remembers
- The vocabulary in action: one small outage
- Terms you can safely ignore for now
- Where to go next
The four families at a glance
| Family | The question it answers | The terms |
|---|---|---|
| Collecting | How do numbers get from devices into the tool? | poller, check, agent, SNMP, WMI, OID, counter vs. gauge, polling interval |
| Judging | How does a number become "fine" or "not fine"? | threshold, baseline, state, flapping, hysteresis & confirmation |
| Alerting | How does a verdict reach a human? | alert vs. notification, severity, escalation, acknowledgment, maintenance window, false positive/negative, alert fatigue |
| Data | What does the tool remember, and for how long? | time series, rollup & retention, availability, dependency |
One warning before the definitions: vendors disagree on names far more than on ideas. What this page calls a check, four different tools will call a sensor, a monitor, a service, and a probe — while meaning the same thing. (A fifth will call it a "monitor object", and mean it too.) When a term in your tool's interface looks unfamiliar, do not assume the concept is new. Look at what the thing does, find the matching idea here, and translate. The machinery underneath is remarkably standard across the whole industry.
Collecting: how numbers arrive
Everything begins with getting a number out of a device and into the monitoring system. This family is the plumbing.
Poller
The component of a monitoring system that asks devices questions on a schedule. It runs a permanent loop over three lists — devices, credentials, and what to ask each one — and it never gets bored. In real life: your poller asks the firewall "what is your CPU right now?" every 60 seconds. That is 1,440 identical questions a day, which is precisely the point: the value of monitoring is relentless repetition no human would tolerate.
Check (also called sensor, monitor, or probe)
One configured question about one thing, with expectations attached. Every vendor picks a different name for this object, which causes more confusion than any concept on this page — they are all the same idea. In real life: "ping 10.0.0.1," "free space on drive C:," and "does the web server answer within two seconds?" are three checks. Your whole monitoring setup is, at heart, a list of checks.
Agent
A small program installed on a machine so it can report from the inside. The opposite approach — asking through interfaces the machine already has, like SNMP or WMI — is called agentless. Agents see more; agentless deploys with less effort. In real life: a server answers pings cheerfully while its backup service sits crashed. An agent (or a WMI check) reports the dead service; a ping never would.
SNMP
Simple Network Management Protocol: the standard question-and-answer language spoken by network equipment — switches, firewalls, printers, UPS units. The monitoring tool sends a small request; the device returns a value. Access in older SNMP versions is controlled by a community string, essentially a shared password. In real life: port status, traffic counts, and device uptime on your switch all reach your dashboard as SNMP answers.
WMI
Windows Management Instrumentation: the question-answering interface built into Windows itself. It is how a monitoring system asks a Windows server about disks, memory, services, and event logs without installing anything. It just needs an account with permission to ask. In real life: "how full is drive D: on the file server?" is a WMI query, answered in milliseconds from across the network.
OID
Object Identifier: the numeric address of one specific fact in SNMP's giant filing cabinet. Every fact a device can report lives at a dotted path like 1.3.6.1.2.1.1.3.0 (that one is uptime). In real life: tools hide OIDs behind friendly names, until the day you add a vendor-specific sensor and the docs say "poll this OID." Now you know that is just a filing-cabinet address.
Counter vs. gauge
The two personalities a metric can have. A counter is a running total since boot — bytes received, errors seen — that only climbs. One reading is meaningless, but the difference between two readings gives you a rate. A gauge is a right-now value that moves both directions — CPU percent, temperature — meaningful the moment you read it. In real life: your bandwidth graph is a counter read every minute, subtracted, and divided by time. Your CPU graph is a gauge, plotted raw.
Polling interval
How often a check runs. Shorter means fresher data and faster detection, at the cost of more load on the network and devices. The interval also sets your reaction clock, since nothing can be noticed faster than it is checked. In real life: the uplink is polled every 60 seconds, printer toner every 15 minutes. An alert rule of "down for 3 polls" on a one-minute interval means you find out roughly three minutes after the crash.
Judging: how numbers become verdicts
A pile of numbers is not monitoring. This family is how "47" becomes "fine" — or "get up."
Threshold
The configured line between acceptable and not, usually set as a pair: one value for "worth a look," a higher one for "act now." A threshold is an opinion — ideally yours, based on your network, rather than a default someone shipped. In real life: disk usage warns at 85% and goes critical at 95%. The numbers feel arbitrary until you tune them against your own machines, and then they feel like policy.
Baseline
Your network's learned normal — the reference that gives any reading its meaning. Nobody can hand you one; it grows out of watching your own graphs for a couple of weeks. In real life: 70% CPU on the application server at 9 AM is fine, because it is 70% every 9 AM. The same value at 3 AM is a story, because the baseline says nights idle at 10%.
State
The current verdict attached to each check: OK, warning, or critical, plus unknown when the question could not be asked at all. Alerting runs on state changes, not on values — the moment of crossing is the event. In real life: the green, yellow, and red tiles on your dashboard are states. Treat unknown with respect: "I could not reach it to ask" is not the same as "it is fine," and gray tiles have hidden real outages.
Flapping
A check bouncing rapidly between states — up, down, up, down — usually because something is genuinely unstable or a value hovers exactly on a threshold. Each bounce is a state change, and naive systems send a message per bounce. In real life: a dying patch cable takes a port up and down every couple of minutes. Without protection, that is forty emails by lunch, each individually true and collectively useless.
Hysteresis and confirmation
The two standard flap-killers. Confirmation means a condition must persist — several failed polls, several minutes — before the state changes. Hysteresis means the trigger level and the all-clear level differ, so a value hovering at the line cannot oscillate across it. In real life: "alert after 3 failed polls" is confirmation; "warn at 85% but do not clear until 80%" is hysteresis. Together they are most of the difference between a quiet tool and a hated one.
Alerting: how verdicts reach people
The judging family decides something is wrong. This family decides who hears about it, how loudly, and how often — which is where monitoring is won or lost.
Northgate Retail learned the first term in this family the expensive way. One Sunday night a till server's disk crossed its critical line. The notification rule, which nobody had read since installation, re-sent the message every five minutes until somebody acknowledged it. Nobody did, because it was Sunday. By Monday morning the store manager had one hundred and eight emails and a firmly worded request to "turn off the alerts". The admin very nearly did. What had actually happened was one alert, one hundred and eight notifications, and one missing acknowledgment. The fix was a repeat interval of an hour plus an escalation to a second phone. Every word in those sentences is defined below.
Alert vs. notification
An alert is the system's recorded judgment that a problem exists — a logged state change. A notification is a message sent about it. One alert can produce many notifications, or none at all. In real life: the file server's disk crosses 95% once: one alert. You receive an email, then a push message, then a repeat an hour later: three notifications. When someone says "I got six alerts," it was usually one alert and six notifications — a distinction that matters when you tune the noise.
Severity
The label saying how bad an alert is — commonly info, warning, and critical. Severities only work when each level maps to a distinct human response, decided in advance. In real life: critical means "wake someone": firewall down, disk nearly full. Warning means "look during business hours": errors climbing. Info means "dashboard only." If every alert is critical, nothing is.
Escalation
The rule that an unanswered alert climbs a ladder: after a set time without response, it goes to the next person or the next louder channel. Escalation is the safety net under a missed notification. In real life: critical alert at 02:10, email and push to you; at 02:25, still unacknowledged, it texts your colleague. In a one-person shop, escalate across channels instead — email first, then a sound your phone cannot sleep through.
Acknowledgment
Telling the system "seen it — it's mine." Acknowledging an alert stops the repeat notifications and shows everyone else the problem has an owner. It fixes nothing by itself. In real life: the disk alert fires at 9:02; you acknowledge at 9:03 so it stops re-sending while you clear space. The classic failure is acknowledging and then getting pulled into a meeting — which is exactly what escalation and re-alert timers exist to catch.
Maintenance window
A scheduled period when alerts for chosen devices are suppressed because you are working on them on purpose. In real life: patch night reboots six servers. With a 22:00–24:00 maintenance window, silence. Without one, thirty alerts. After a few such nights the team learns that late-evening alerts are "probably just patching." That is precisely how a real outage sneaks through.
False positive and false negative
A false positive is an alert without a real problem; a false negative is a real problem without an alert. Every threshold choice trades one against the other. Tighten and you catch more real issues plus more noise. Loosen and you quiet the noise plus miss real events. In real life: the latency alert that fires during every backup is a false positive — raise the line or schedule around it. The disk that filled at 97% while the alert waited for 99% was a false negative. Positives cost trust; negatives cost outages.
Alert fatigue
The numbness that sets in when alerts are too frequent: people stop reading them, then stop noticing them, then rule them into a folder. It is the disease every other term in this family exists to prevent, and the most common way monitoring systems die. In real life: fifty alerts a day, mostly noise, until the one that mattered scrolled past unread. The cure is never "try harder to read them" — it is fewer, better alerts.
Data: what the tool remembers
The last family covers the memory of the system — what gets kept, what gets summarized away, and the relationships that make sense of it.
Time series
The stored history of one metric: a list of timestamped values in time order — a two-column diary of when and what. Monitoring databases hold one series per measured thing. In real life: every graph you have ever looked at is a time series, drawn. "CPU on dc-01" is one series; a 20-port switch with three metrics per port contributes sixty.
Rollup and retention
How the diary stays affordable. Rollup condenses aging data — per-minute readings are averaged into five-minute or hourly summaries as they age. Retention is the schedule for how long each level of detail is kept before deletion. In real life: last week's graph shows every minute; last year's shows hourly averages, which is why old spikes look mysteriously smaller. They were averaged, not healed — so investigate oddities while the raw data still exists.
Availability (and uptime)
The percentage of a period a check spent in the OK state — the number that ends up in reports. Watch for the sibling word uptime, which also means "time since this device last booted"; the two get conflated constantly. In real life: "99.9% availability" sounds immaculate and still allows nearly nine hours of downtime a year. Meanwhile a switch reporting uptime of 20 minutes has just rebooted, and knowing that solves half the mystery.
Dependency
A declared parent-child relationship between checks that mirrors your topology: these forty devices are reachable only through that switch. When the parent fails, the children's alerts are suppressed as consequences, not causes. In real life: the core switch dies at 6 AM. With dependencies: one alert, correctly aimed. Without: forty-one alerts, and the actual culprit buried somewhere in the pile while you triage alphabetically.
The vocabulary in action: one small outage
Here is a fifteen-minute incident, narrated in glossary. At 02:11 the poller asks the branch router its usual question and gets silence. The reachability check does not panic — its confirmation rule wants three failed polls. At 02:14, third miss: the check's state flips from OK to critical, and an alert is recorded. Dependencies immediately mark the nine devices behind that router as unreachable-by-consequence and suppress their alerts — one story, not ten. Because the severity is critical, the routing rules fire a notification to your phone. You acknowledge it at 02:16, which stops the repeats and cancels the escalation that would have texted your colleague at 02:30. The graphs — each one a time series — show the router's latency gauge (latency being the delay before an answer comes back). It climbs for twenty minutes before the drop, well past its usual baseline: not a crash, a sick line. The ISP (internet service provider) confirms and fixes. At 02:49 a poll succeeds. Confirmation holds the all-clear until 02:52, then state returns to OK. In the morning you log the 38 minutes against the month's availability and conclude the alert was no false positive. Because the line may wobble again tonight, you set a maintenance window for the ISP's announced repair slot. That way, nobody's sleep is spent on planned work. Every term, one ordinary night.
Terms you can safely ignore for now
Monitoring has a deep end, and vendors love to describe it. None of the following is nonsense — each is genuinely useful in the right context. But none belongs in your first months, and no salesperson should convince you otherwise:
- NetFlow, sFlow, IPFIX: formats for traffic-conversation records. Relevant the day your question becomes "who is eating the bandwidth?"
- Syslog collection: gathering devices' text logs centrally. A worthy project — a different project.
- Streaming telemetry: a push-based alternative to polling used in large networks. Polling serves a small network perfectly well.
- Observability and "the three pillars": umbrella jargon for metrics, logs, and traces together. You are already building the pillar that matters first.
- Percentiles (p95, p99): statistics for describing latency at scale. Averages and eyeballs are fine to start.
- SLA / SLO: formal availability promises and targets. They matter when contracts appear; the availability number above is the raw material.
- Anomaly detection: software-judged weirdness instead of fixed thresholds. Learn what your own normal looks like before delegating the judgment.
- Event correlation engines: the big-network cousin of the humble dependency, which does most of the same job for you already.
Tip: when a term you meet is not on this page, apply the family test: is it about collecting, judging, alerting, or data? Nearly everything in monitoring slots into one of the four, and knowing which family a mystery word belongs to is halfway to understanding it.
Where to go next
Vocabulary is leverage: settings pages stop being intimidating the moment their words mean something. You can now tell a poller from a probe and an alert from a notification without slowing down. That was the whole point of the sit-down. The next time a settings page says "hysteresis", you will nod, at a reasonable hour, with nothing coming in. The natural next step is to put the words to work. What Should You Actually Monitor? The First 10 Checks for Any Small Network turns this vocabulary into an afternoon of concrete setup. The rest of Network Monitoring Fundamentals fills in the deeper how and why.
Put this into practice with SysPeep Monitor
The fastest way to make these terms stick is to watch them work. SysPeep Monitor is built from exactly the pieces in this glossary. It has checks collected over SNMP and WMI, thresholds, dependencies, and alert rules designed to keep the noise down. All are in one web console. Install the free trial on a Windows machine and you can point at each concept on a live dashboard within the hour.
Download Free TrialLearn About SysPeep