From Proactive to Predictive: Knowing Before It Breaks
"Predictive" is the most abused word in monitoring. It appears on boxes, in brochures, and on login pages, usually next to a graph with a dotted line extending into the future. It is sold as something you acquire. It is not. For a small team, predictive is a set of habits with a calendar attached. Every one of those habits can be done with the graphs you already have and a spreadsheet.
The habits are worth having. A team that is proactive finds most faults before a user does. A team that is predictive finds most of them before they are faults. The disk that will fill has a date beside it. The failover that will be needed has been tested. The plan for next quarter contains three items that came from graphs rather than from complaints. It is calmer, and it is not magic; it is looking at the same data on a schedule and writing down what it says.
By the end of this guide you will be able to define predictive in a sentence that a vendor cannot sell you. You will be able to recognise it on a Tuesday and run the six practices that make it real. You will be able to tell which parts of the marketing describe something you can do and which describe something you can buy. And you will be able to state its limit honestly, because predictive does not mean the end of outages.
In this guide
- Predictive, honestly defined
- What a Tuesday looks like
- The six practices
- What is marketing
- The honest limit
- The predictive practices template
- The proactive-to-predictive checklist
- Where to go next
Predictive, honestly defined
Predictive, in the model this guide belongs to, means three plain things. You review trends and capacity on a calendar, not when something looks odd. You rehearse changes and failovers before they are needed, so the first time the standby firewall takes over is not during an outage. And you act on forecasts before they become incidents, which means the quarterly plan contains items whose only justification is a graph. The four levels, plainly places this as level 3 of four. It says more than once that a level is a map, not a grade. That applies with extra force to the top one.
Notice what the definition does not contain. It does not say machine learning. It does not say anomaly detection, artificial intelligence, or a feature you enable. Those things exist, some of them are useful, and none of them are the level. A team with a plain graph, a ruler, and a recurring calendar entry is predictive. A team with a forecasting feature nobody opens is proactive at best, and proactive only if the other practices are in place. The word describes what the team does on a schedule, not what the software could do if asked.
It is also not compulsory. A small team solidly at level 2 is doing well. The monitoring finds faults first, runbooks exist, reviews happen, and capacity is planned. Most of the value lives in the moves that get there. Level 3 is for teams whose level 2 has been stable for a couple of quarters. Those teams have also noticed that the remaining surprises all had a graph that saw them coming.
What a Tuesday looks like
A predictive team's Tuesday is quiet in a specific way: the things that happen were on a list. The capacity register is a table of every finite resource, its current use, its growth rate, and the date it is projected to run out. It has a row for the backup volume that says nine weeks. The purchase order for the replacement went in a fortnight ago. Nobody is surprised by the row. Somebody is mildly surprised that the purchase order was approved on time.
Monday's trend review found something, as it does about one week in three. The VPN concentrator's peak session count has climbed for six consecutive weeks and will meet its licence limit in about four months. A ticket exists with the graph attached. The last maintenance window included a failover test to the standby firewall, which worked. It also included a restore test of the file server from backup. That took forty minutes longer than the runbook said, so the runbook was corrected. The quarterly plan has three items on it that nobody asked for and every one of which came from a graph.
There was an outage last month. A core switch's power supply failed, and the switch had two, so nothing stopped. The alert said which supply, the runbook said which spare, and the replacement went in during a window. The point is not that nothing breaks. The point is that when it does, the team had mostly already thought about it. (The one thing they had not thought about was that the spare was in the other building. It is now in this one.)
The six practices
Six practices make a team predictive, and each has a cadence, a piece of evidence, and an owner. They do not have to be done in a strict order the way the reactive-to-proactive moves do. These practices stand on the level 2 practices rather than on each other. Start with the first two; they are the cheapest and they produce the material the others use. Between them the six touch every dimension of the model. Trend alerts move alerting, the register moves data and reporting, and rehearsals move response. Synthetic checks move coverage, and the SLA and the lab move process. The owner column beside each one is how the people row moves, provided the same name is not in every row.
1. Trend alerts and the weekly review
A trend alert fires not when a metric crosses a line but when its direction says it will cross the line within a stated period. For example: "disk will reach ninety per cent within thirty days at the current rate". It turns the weekly review's best question into something the monitoring asks for you, every night, for every metric you point it at. Alerting on trends covers how to build them without drowning in false forecasts. Cadence: the alert runs continuously; the review is weekly. Evidence: a trend alert rule on each capacity metric, and a weekly note.
2. Capacity planning from history
The capacity register is the heart of predictive, and it is a spreadsheet. One row per finite thing: disk volumes, licence counts, address pools, link bandwidth, the power budget of a switch. For each, the current value, the value ninety days ago, and a projected date. Capacity planning basics explains how to read the growth honestly, including why a line drawn through a quiet season lies. Cadence: updated monthly, reviewed quarterly. Evidence: the register, with dates, and a purchase or change raised because of one.
The projection is arithmetic. With the current used figure in column B, the figure from ninety days ago in column C, and the capacity in column D, the projected date of full is:
=IF(B2<=C2, "not growing",
TODAY() + (D2 - B2) / ((B2 - C2) / 90))
It assumes the next ninety days look like the last ninety. That is the assumption to state out loud when you show the date to anyone. A date with the assumption beside it is a forecast. A date without one is a promise you did not mean to make.
3. Rehearsed failover and restore tests
A failover test deliberately fails the primary of a redundant pair during a maintenance window to prove the secondary takes over. A restore test brings a system back from backup onto spare hardware and checks that it works. Both are rehearsals for the day the real thing happens. Both find the stale route, the expired certificate, and the backup job that has been reporting success while writing nothing. Monitoring backups is about the last of those. Cadence: quarterly for each redundant pair and each critical restore. Evidence: notes from the last rehearsal, including what it found.
4. Synthetic checks for the services users feel
A synthetic check performs a whole user action on a schedule and times each step. For example, it logs into the portal, searches, opens a record, and logs out. It is the service check from level 2 taken one step further. And it is the only monitoring that notices "the system is up but everything takes eleven seconds" before a user does. The synthetic transactions pillar covers building them. The same row of the model asks for a coverage review. This is a scheduled comparison of the service list against the check list so that new services do not arrive unmonitored. The guide to coverage reviews describes the meeting. Cadence: the checks run continuously, with timings reviewed weekly; the coverage review is quarterly. Evidence: a synthetic check per service on the "users feel first" list, a graph of its step times, and notes from the last coverage review.
5. An SLA you measure yourself
An internal SLA is a written promise to the business about a service — "the order system is available during trading hours at least this much". It is measured by your own monitoring and reported by you before anyone asks. It makes the capacity register and the synthetic checks answer a question the business recognises. The internal SLAs pillar covers choosing what to promise and reporting it without spin. Cadence: measured continuously, reported monthly. Evidence: the SLA document and last month's report against it.
6. Change rehearsal in a lab
A lab is a small copy of the production network where a change is tried before it is made for real. Picture an old desktop, a managed switch with a handful of ports, and a used small-business firewall. Firmware upgrades, a new alert rule, a threshold change on a busy metric: run them in the lab first. Then the production change becomes a repeat rather than a first attempt. The homelab monitoring practice pillar covers building one. Cadence: before every change that touches a critical path. Evidence: lab notes attached to the change record.
What is marketing
The word "predictive" on a product means one of two things, and it is worth knowing which. Sometimes it means the product can draw a trend line and alert on where it points, which is practice one above, and is useful. Sometimes it means the product applies a statistical model to your metrics and raises an alert when they behave unusually. That is anomaly detection. It can also be useful, if someone is going to read the alerts. Neither makes a team predictive. A team is predictive when it reviews what the line says on a calendar and does something. The feature draws the line; the level is what happens next.
"AI-driven" is the same word with a larger budget. The observability marketing translator renders the vocabulary into plain terms, gently, and it is worth a read before a demo. The translation for this guide's purposes is short. If a feature tells you something will happen, ask what you would do about it. Ask whether you would have done it anyway with a graph and a calendar entry. Often the answer is yes. Sometimes it is no, and then the feature is worth its price. The level does not change either way.
The honest limit
Predictive does not mean no outages. It means fewer surprises. Hardware still fails on its own schedule, which it does not publish. Upstream providers still have bad days. A change that was rehearsed in the lab still meets a production quirk the lab did not have. The difference is that a predictive team is surprised by the cause and not by the consequence. The failover works because it was tested. The restore takes the time the runbook says because it was timed. And the capacity that was going to run out did not, because it was replaced first.
The other limit is attention. Every practice above produces a document, and a document nobody reads is level 2 with extra files. The register has to be opened; the trend note has to be read by someone who can raise a ticket. The rehearsal notes have to change the runbook. The calendar entry is not the practice. The practice is what the calendar entry makes someone do.
Bellweather Logistics kept an excellent capacity register, reviewed quarterly. The row for the archive volume had said "fills in the second week of the quarter" since the previous review. That review was cancelled to make room for a warehouse system go-live, and the next one was not rebooked. The volume filled in the second week of the quarter, on a Saturday, precisely as the register said it would. The alert fired at ninety-five per cent, and its runbook, which had been written by someone thorough, said "consult the capacity register". The register's nearest date is now the first line of the monthly page, where cancelling a meeting cannot hide it.
Tip: put the nearest date from the capacity register on the monthly management page every month, even when it is comfortably far away. A date that moves closer by a month each month is the most persuasive graph you will ever not have to draw. And the month it stops moving is the month the purchase went through.
The predictive practices template
This is the six practices as a table you can keep. One row per practice, with its cadence, when it was last done, the evidence that it was, and a name. It is written plainly so that it can be pasted into a quarterly plan or a budget document. The "last done" column is the one that tells you whether you are at level 3 or merely own the template.
| Practice | Cadence | Last done | Evidence | Owner |
|---|---|---|---|---|
| Trend alerts on capacity metrics | Continuous; rules reviewed quarterly | Trend alert rule per capacity metric; last rule review notes | ||
| Weekly trend review | Weekly, fifteen minutes | One-line note per week; tickets raised from findings | ||
| Capacity register update | Monthly | Register with current figures and projected dates | ||
| Capacity register review | Quarterly | Review notes; purchases or changes raised from the register | ||
| Failover test per redundant pair | Quarterly, in a maintenance window | Rehearsal notes with findings and runbook corrections | ||
| Restore test per critical system | Quarterly | Restore notes with the time taken and the runbook updated | ||
| Synthetic check per user-facing service | Continuous; timings reviewed weekly | Check list matches the "users feel first" list; step-time graphs | ||
| Internal SLA measurement and report | Continuous; reported monthly | SLA document; last month's report against it | ||
| Change rehearsal in the lab | Before each critical-path change | Lab notes attached to the change record | ||
| Quarterly plan items from graphs | Quarterly | Plan with at least one item whose justification is a graph |
A row with a blank "last done" is a practice you intend to have, and intentions belong in the plan rather than the score. A row whose "last done" is more than two cadences ago is a practice that has stopped. And it is better to write that down than to leave the old date in place and hope. If the same name fills every owner cell, the people dimension is at level 1 whatever the rest of the table says. In that case, that is the row to fix before adding a practice. The template is honest only if the dates and the names are.
The proactive-to-predictive checklist
Use this at the quarterly review. Each line is a yes or a no, and a yes should have a date and a file behind it.
- Every capacity metric has a trend alert, and the last false forecast was investigated and the rule adjusted.
- The weekly review happened every week this quarter, and at least one ticket came from it.
- The capacity register was updated each month and has a projected date on every row.
- At least one purchase or change this quarter was raised because of a register date, before the resource ran out.
- Every redundant pair had a failover test this quarter, with notes.
- Every critical system had a restore test this quarter, and its runbook time was corrected if needed.
- Every service on the "users feel first" list has a synthetic check with step timings.
- The internal SLA was reported last month, by the team, before anyone asked.
- Every critical-path change this quarter has lab notes attached.
- Next quarter's plan contains at least one item that came from a graph.
- The nearest register date appears on the monthly management page.
Where to go next
Predictive is three habits: review trends and capacity on a calendar, rehearse before you need to, and act on the forecast before it is an incident. Six practices make the habits real, each with a cadence and a date, and a table keeps the dates honest. None of it needs a feature with the word on the box, and none of it prevents every outage. It prevents the ones that had a graph, which turns out to be most of the ones that hurt.
A team at level 3 has a story worth telling in the budget room. The guide to turning maturity into budget and credibility shows how the register, the rehearsal notes, and the SLA report become the case for whatever the next step needs. If the practices in this guide feel a level too far, the self-assessment will show which level 2 statements still need evidence. All of them are part of A Monitoring Maturity Model for Small Teams.
Put this into practice with SysPeep Monitor
The capacity register and the trend review both live on long-term graphs. The failover test needs a monitoring system that is itself still standing when the primary is not. SysPeep Monitor keeps long-term graphs and reports in one web console, with thresholds, dependencies, and alert rules beside them. It collects over SNMP, WMI, and SSH from a central Windows service with optional remote probes. It offers high-availability clustering for the teams whose monitoring cannot be the single point of failure. The free trial is enough to draw the first ninety days of history the register needs.
Download Free TrialLearn About SysPeep