Home › Topics › Baselining & Knowing Normal › Rebaselining

When Normal Changes: Rebaselining After Growth and Change


Picture a baseline register written eighteen months ago, back when it was accurate and everyone was proud of it. Since then, the company hired thirty people, moved to a bigger office, and migrated half its applications to the cloud. Nobody updated the register. Now the WAN alert fires most mornings. A threshold reasoned from the old normal is being trampled by the new one. Everyone has learned to swat the alert away. Meanwhile the old batch server's generous margins, sized for a workload that no longer exists, would hide a real problem for days. Same register, both failure modes at once: stale normal crying wolf on one metric, stale normal asleep on watch on another.


Baselines expire. That is not a flaw in the method. It is the method meeting reality, because a baseline is a photograph of a moving subject. That is the whole framing from what a baseline really is. The skill this guide teaches is rebaselining: deliberately re-measuring and re-documenting normal when normal has legitimately changed. It avoids the one trap that makes rebaselining dangerous: blessing a change nobody can explain.


You will learn the three ways normal changes (and which of them justify a refresh), and the triggers that should start one. You will also learn the comparative method and the attribution discipline that keeps problems from being grandfathered into "normal." Then comes the follow-through that most teams skip. At the end: a decision tree you can pin to the wall, and the annual ritual that keeps a register alive for years.


In this guide

  • Three ways normal changes: drift, step, mystery
  • The triggers: when to rebaseline
  • The method: collect, compare, explain, record
  • Attribution before acceptance
  • Partial rebaselining: scope the refresh
  • Threshold follow-through: the other half of the job
  • The decision tree and the annual ritual
  • Where to go next

Three ways normal changes: drift, step, mystery


Everything in this guide hangs on one distinction: how did normal move? There are three shapes, and they call for three different responses.


Drift: the slow climb of organic growth

Drift is gradual movement over months. WAN utilization creeps up a percent or two per month as headcount grows. The file server's disk consumption accelerates as projects accumulate. Backup windows stretch as data does. Drift has no single cause and no single day it happened; it is the business growing, translated into metrics. Drift is a legitimate new normal — the register should follow it. One caution before you call something drift: make sure it is not rhythm. A school's September surge and a retailer's holiday climb are seasonal patterns that will reverse on schedule. Rebaselining on top of a seasonal peak bakes a temporary high into your permanent normal. When a rise might be seasonal, check it against your pattern library (see weekly rhythms and seasonal patterns) before treating it as growth.


Drift has one more property worth exploiting: it is invisible to thresholds (no single day crosses any line) but obvious to reviews. Comparing this quarter's numbers against the register's is the reliable way to catch it. When you do, record not just the new level but the drift rate: "climbing roughly two points a month for a year." A rate turns a baseline into a forecast. Two points a month against a link whose comfortable ceiling is twelve points away means about six months of runway. That is exactly the sentence a budget conversation wants to hear before the problem arrives rather than after.


Step change: the overnight jump with a name

A step change is a jump: normal is one thing on Tuesday and visibly another on Wednesday, and it stays there. Legitimate step changes come from known events. These include a migration to a cloud app (WAN up, LAN-to-server traffic down) or a new office joining the network. They also include a big deployment or a deliberate configuration change. A step change is a legitimate new normal if and only if it is explained. You can point at the change record, the project, the rollout email. The explanation is not decoration; it is the entire difference between a new normal and an unsolved incident with good posture.


Mystery: the change nobody ordered

And then there is the step nobody can attribute. Traffic doubled and no project explains it; CPU found a new plateau and no deployment matches the date. This is not a rebaselining trigger. It is an investigation trigger. Confusing the two is the single most dangerous mistake in all of baselining. Write this rule somewhere permanent: never rebaseline over a mystery. Rebaselining an unexplained change is how problems get grandfathered into "normal." The malware's beaconing becomes part of the traffic baseline. The memory leak becomes "that server just runs hot." The misconfigured replication storm becomes background. Each future comparison then measures against a sick normal. And the monitoring system — whose whole job is noticing change — has been taught to un-notice this one. Investigate first; the register only ever absorbs what you can explain.


The triggers: when to rebaseline


Three kinds of trigger should start a rebaselining pass — two you schedule, one that finds you.


  • Planned: after any change big enough to move normal — migrations, office moves, major deployments, ISP or circuit changes, big onboarding waves. Best practice bakes it into the change plan itself. The project is not "done" at go-live. It is done when the post-change baseline is captured and the register updated. (You are already capturing a pre-change snapshot as part of the change ritual; the post-change baseline is its other half.)
  • Periodic: the scheduled review. The register's next-review column — part of the template from collecting your first baselines — exists precisely so rows do not silently fossilize. Annually is the floor for every row; every six months is kinder for fast-moving metrics like WAN utilization and disk growth.
  • Reactive: the alert system tells you. A false-alarm review may find the metric genuinely outside its documented normal with nothing wrong. That is the "normal moved" branch of the baseline-first alert review. Or the same alert may keep firing for a cause everyone can name ("that's the new video platform"). In either case, the alerts are ringing a doorbell. Answer it with a scoped rebaseline rather than another round of threshold-nudging.

Just as useful is knowing what is not a trigger. One strange week is not a trigger. A fluke is a fluke, and two cycles of agreement were the standard for calling anything normal in the first place. An incident is not a trigger. An outage week's numbers describe the outage, not the network, so let the register sit until things are genuinely back. And a new admin's discomfort with inherited numbers is a reason to verify a baseline, not to discard it. Re-measure, compare, and keep whichever survives contact with the data.


The method: collect, compare, explain, record


Rebaselining is the first collection project run again, with one upgrade. This time you have something to compare against, and the comparison is where the value lives.


Collect a fresh observation window for the affected metrics, using the same standard as the first time: at least two full business cycles. That means two ordinary weeks minimum, a month where monthly rhythms matter. Use the same polling settings as before so the numbers are comparable. After a step change, let the dust settle first. The first days after a migration are full of one-time noise (re-syncs, catch-up jobs, users exploring) that you do not want enshrined as normal.


How long is "settled"? Watch the graphs for the shape to stop changing day over day. A small rollout usually settles within a few days. A big migration can churn for a week or two while initial synchronizations finish and users find their new habits. An office move settles only after the last team actually moves in. The tell is simple. When two consecutive weeks agree with each other — the same standard as ever — the new normal has arrived and is ready to be measured. Start the official capture window there. And note in the register that the interim weeks were excluded and why, so nobody later wonders about the gap.


Compare old and new side by side — the fortnight-reading skills from the first collection apply unchanged. Did the busy-hour range move, or just the peaks? Did the shape change (a new 20:00 bump that never used to exist), or only the level? Did the floor move — the quiet hours getting louder is its own finding. Explain every delta that matters, in writing, per the attribution discipline below. Then record: the register row gets its new values, and — this is the upgrade — a history entry. Four fields do it:


MetricNormal wasNormal isWhyWhen
WAN utilization (busy hours)35–50%, p95 58%50–65%, p95 74%Company-wide video conferencing rolloutCaptured the month after rollout
File server disk growth~15 GB/week~6 GB/weekDepartment file shares migrated to cloud storagePost-migration window
DNS queries/min (business hours)~900–1,400~1,300–1,900Thirty new hires onboarded over two monthsSemi-annual review

Never delete the old rows — strike through or move them to a history section, but keep them. A register that only shows today's normal answers one question; a register with history answers a better one. The trend of normals is its own insight: three years of "what busy-hour WAN utilization counted as normal" tells the growth story. It shows how demand actually tracks headcount and app choices, more honestly than any capacity report. Every entry was written down when it was true, not reconstructed for a budget meeting. When someone asks "do we need a bigger circuit next year?", the answer is already plotted.


Attribution before acceptance


The discipline that keeps rebaselining honest is the explained-delta rule. Every delta you accept into the register must carry a cause you could defend to a colleague. Two versions of the same event show the rule at work.


The explained double. Six months after the last capture, WAN busy-hour p95 has gone from 30% to 60%. The delta is enormous — but the explanation is sitting in plain sight. The company rolled out video conferencing to every meeting room. The rise dates from the rollout week, and the new traffic profile matches (steady daytime streams, worst during the morning meeting block). Accepted: the register gets the new numbers, and the why column gets the rollout. The calendar-of-causes gets the meeting-block peak, and capacity planning gets a heads-up that the circuit's headroom halved. Rebaselining here is just bookkeeping for a change everyone already knew about.


The unexplained double. Same graph, same doubling — but no rollout, no project, no memo. Now the rule bites: this is a mystery, so it goes to investigation, not to the register. The investigation is ordinary detective work, and the baseline itself is your best witness. Pin down when the change started, to the day or hour if the graphs allow. Ask what happened then — change logs, deployments, new devices on the network. Characterize the new traffic: which hosts, which direction, which hours. A doubling that lives at 3 a.m. has a very different suspect list than one inside business hours. Per-conversation visibility helps if you have it. There are a few usual endings. One is a legitimate-but-undocumented change (someone enabled cloud backup and told nobody — document it, then accept). Another is a misconfiguration (fix it, watch normal fall back, no rebaseline needed). Or it is something genuinely unwelcome (escalate; the baseline just earned its keep). Only when the delta has a name does it become eligible for the register.


Make the acceptance itself explicit, not ambient. A new normal should enter the register as a small, dated decision: who accepted it, on what evidence. That holds even if "who" is just you and "evidence" is one sentence. The failure mode this prevents is quiet ratification. A number gets glanced at, tolerated, and three months later everyone treats it as official because nobody remembers it is not. Deciding on purpose takes thirty seconds; undoing an accidental normal takes considerably longer.


Tip: attribution has a half-life. The week after a change, "what happened around the 12th?" gets answered in one hallway conversation. Six months later it is archaeology. This is the practical argument for reviewing baselines often. Young mysteries are cheap to solve, and old ones get expensive — or worse, get accepted out of fatigue.


Partial rebaselining: scope the refresh


Rebaselining is not all-or-nothing, and treating it that way is why it gets postponed. When one thing changed, refresh what it touched and leave the rest alone. A mail migration to the cloud moves WAN utilization, the mail server's whole row, and DNS query patterns. But the core switch CPU, the Wi-Fi client counts, and the backup window did not attend that project. Scope the refresh to the metrics the change plausibly reaches. Think one hop outward: the changed system, whatever carries its traffic, whatever depended on it. Stamp only those rows with new capture dates, and the job shrinks from "re-audit the network" to an afternoon. Scoping also keeps histories honest: a row should change when its normal changed, not because a spreadsheet got regenerated. The full-register pass still has its place — that is the annual ritual below — but between rituals, rebaselining travels in small, well-labeled batches.


Threshold follow-through: the other half of the job


A rebaseline that stops at the register is half a job — and the half you skipped is the half that pages people at night. Remember where thresholds came from: baseline plus reasoned margin. Move the baseline and every threshold derived from it is now sitting at the wrong height. It is too tight if normal rose (false alarms, and the reflexive "just raise it" fix that erases the reasoning). It is too loose if normal fell (a silent watchman with obsolete margins). So the last step of every accepted rebaseline is mechanical. Walk the affected rows through the threshold worksheet again — margin, duration, schedule, low side, written reason — exactly as in using baselines to set smarter alerts. The low side deserves special attention after downward shifts. When the file server's traffic legitimately fell because its data moved to the cloud, yesterday's "suspiciously quiet" is today's Tuesday. In that case, a too-quiet alert still tuned to the old floor will cry wolf every morning. New normal, new margins, new reasons, new dates. Then the loop is actually closed.


The decision tree and the annual ritual


Everything above compresses into one table — the wall-pin version:


What you seeRebaseline?First action
Slow climb tracking business growthYes, at the next reviewConfirm it is drift, not a seasonal rhythm; then refresh the row and re-derive thresholds
Step change matching a known project or change recordYes, once settledWait out the post-change noise, capture two clean cycles, record delta + why
Step change nobody can attributeNoInvestigate: date the change, hunt the cause, name it — only then decide
Seasonal peak (month-end, enrollment, holidays)NoRecord it in the pattern library as rhythm; normal has not moved, it is oscillating
Repeated false alarms; metric truly outside old normal; cause nameableYes, scopedPartial rebaseline of the affected rows, then threshold follow-through
Old normal restored after a fix (the change was a fault)NoKeep the original baseline; note the incident in history

And once a year, the full ritual — an afternoon that keeps the register alive:


  • Walk every register row: is the capture date within its review period? Refresh the overdue.
  • Compare each row against a fresh fortnight: any quiet drift the alerts never surfaced?
  • Reconcile against the year's change log: any project that should have triggered a rebaseline and did not?
  • Update the pattern library and calendar of causes alongside — rhythms drift too.
  • Re-derive thresholds for every row you touched, reasons included.
  • Read the trend of normals and write the one-paragraph growth story — then send it to whoever owns the budget. It is the cheapest capacity report you will ever produce.

Where to go next


This guide closes the loop the series opened. Define normal, measure it, and learn its rhythms. Wire it into alerting — and now, keep it true as the business changes underneath it. A register maintained this way stops being a document and becomes a habit. Normal is measured, explained, and current. So every deviation is worth your attention. That is the entire promise of the practice. The full series lives at Baselining & Knowing Normal.


Put this into practice with SysPeep Monitor

Rebaselining leans on long history: old normal and new normal on the same graph, with the change date visible between them. SysPeep Monitor keeps that history for the metrics it collects over SNMP, WMI, and SSH. It provides dashboards and reports for the side-by-side reading. It also provides thresholds and alert rules you can re-derive as soon as a new normal is accepted. The free trial installs on any Windows machine.

Download Free TrialLearn About SysPeep