Reading Growth Curves: Linear, Step, Seasonal, and Runaway
Once you have set a graph to twelve months and found a line going up, the next question is what kind of up. A disk that fills by the same amount every week and a disk that fills twice as fast every month both draw rising lines. But one gives you a year to act and the other gives you a fortnight. A link that jumped from thirty percent to seventy percent on a Tuesday is not growing at all; something happened on Tuesday. The shape of the curve tells you which story you are in.
Reading shapes is a skill, and a small one. Only six shapes matter on a monitoring graph. Each has a short list of likely causes, and each has a different correct response. Confusing them is expensive in both directions. Treat a step as a trend and you buy hardware to solve a configuration mistake. Treat a runaway as a trend and you schedule a purchase for a quarter you will not reach.
This guide walks through the six shapes with what each looks like, what usually causes it, and what to do. You will get a quarter-over-quarter test that separates linear from runaway growth without any statistics. You will also get a note on when a logarithmic scale helps and when it lies, and a table of shapes to keep beside the graph.
In this guide
- Why the shape matters more than the level
- The six shapes, one at a time
- Telling them apart: by eye and by a simple test
- When a logarithmic scale tells the truth, and when it hides it
- Matching each shape to a response
- The curve shapes table
- The step that was a backup job
- A shape-reading checklist
- Where to go next
Why the shape matters more than the level
A level tells you where a metric is; a shape tells you how it got there and where it is going. The previous guide, what degradation creep looks like, made the case that slopes break things and levels do not. This guide goes one step further: the slope is not a single number, because the line is rarely straight. Its shape over a year is the difference between "buy a bigger disk next spring" and "find out what started last Tuesday".
The shape also tells you whether you are looking at a trend at all. Many of the rising lines I have been asked to explain were not trends. They were one change, drawn as a cliff, with flat ground on both sides. Planning for a change as if it were a trend is how the network ends up with a new core switch and the same problem.
The six shapes, one at a time
You need the year view for this. The difference between a linear rise and an accelerating one only shows over several months, and a season needs the whole year to be visible at all. Set the window, fix the vertical axis so it starts at zero, and look.
Flat: healthy, or a dead collector
A flat line is a metric that has not changed. Most of the time that is the best news on the page: the disk is not filling, the errors are not climbing. But a perfectly flat line is also what a broken collector draws, because a value that is never updated never moves. Look at the texture. A healthy flat line has a little wobble in it, a daily rhythm, a wiggle at the backup. A dead one is a ruler-straight horizontal at the last value it managed to read. The monitoring the monitor pillar covers keeping collectors honest; here the rule is that a line with no noise at all deserves suspicion. (Nothing in IT is that calm.)
Linear: steady growth
A linear rise adds the same amount every week. Think of disk use on a file server, the row count in a database, the number of devices on the network, or mail store size. These things grow because people keep adding to them at a steady rate. Linear growth is the friendliest shape because you can plan it with a ruler. Lay a straight edge along the line, extend it to the threshold, and read the date off the axis. The capacity planning from history pillar turns that ruler into a purchase plan. Draw the ruler through a full year, not the quiet half of it.
Step: something changed on a date
A step is flat, then a vertical jump, then flat again at a new level. It is not growth; it is an event. A project went live, a site was migrated, a new office joined the WAN, a backup job was re-pointed, a firewall rule started logging every packet. Steps are the easiest shape to explain and the most often misread. On a short graph, a step that happened three months ago looks like a line that has always been high. The first job with a step is to find the date, and the second is to find out what changed on it. A step is a question, not a finding.
Seasonal: the calendar, drawn as a wave
A seasonal curve rises and falls with a rhythm that matches something in the real world. Think of term time at a school, month end in a finance team, the holiday quiet at a manufacturer, or the annual sale at a retailer. The guide on seasonal patterns explains the rhythms in detail. For trend reading, the rule is to compare like with like: this September against last September, not against August. A metric higher than it was in July may be growing, or it may simply be September. You cannot tell from one peak. You can tell from two.
Runaway: accelerating growth
A runaway curve gets steeper as it goes. Each month adds more than the month before, so the line bends upwards like the side of a bowl. This is the shape of a leak, a loop, or a runaway process. It might be a log file that logs its own errors about being too large. Or it might be a retry storm where each failure schedules two more attempts. Or a replication queue might fall further behind the further behind it gets. Runaway growth is the one shape where planning is the wrong response. Do not extend the line to the threshold and book a date. The line will get there before the date does. Find the cause and stop it.
Sawtooth: a cycle of fill and reset
A sawtooth climbs, drops sharply, and climbs again, over and over. The drop is a reset: a scheduled job that clears a queue, a log rotation, a nightly purge, a reboot. The teeth are usually harmless, and the interesting information is in their edges. If the peaks are climbing over the months, the reset is no longer keeping up. If the troughs are climbing, the reset is no longer clearing everything. Either is creep wearing a disguise. A memory leak on a device that reboots weekly draws a level sawtooth with a seven-day period. It looks healthy right up until the week the reboot does not happen.
Telling them apart: by eye and by a simple test
By eye, the tell for each shape is in a different place. Flat and linear are straight; a step has one vertical edge; seasonal repeats; runaway bends; sawtooth has many vertical edges at regular spacing. The two people confuse are linear and runaway, because over three months a gently bending line looks straight. The fix is to stop looking and start measuring.
The measurement is quarter-over-quarter growth, and it needs no statistics. Split the year into four quarters and take the average value of each. Then look at how the average changed from one quarter to the next, in two ways. Check the difference (this quarter minus last) and the ratio (this quarter divided by last). Linear growth has a roughly constant difference: plus 40 gigabytes each quarter. Runaway growth has a roughly constant ratio, or a difference that grows: plus 40, then plus 60, then plus 90. A step has one big difference and then differences near zero. Seasonal growth shows as a quarter that matches the same quarter last year rather than the quarter before it.
Here is the test as a short script, for a CSV export with a date column and a value column, one row per day.
# quarters.py - quarter-over-quarter growth from a CSV of date,value (one row per day)
import csv, sys
vals = [float(v) for _, v in csv.reader(open(sys.argv[1], newline=""))]
q = [sum(vals[i:i + 91]) / 91 for i in range(0, len(vals) - 90, 91)]
for n in range(1, len(q)):
diff = q[n] - q[n - 1]
ratio = q[n] / q[n - 1] if q[n - 1] else float("nan")
print(f"quarter {n}: mean {q[n]:.1f} change {diff:+.1f} ratio {ratio:.2f}")
Read the output down the change column. Steady changes mean linear; growing changes mean runaway; one large change followed by small ones means a step. In a spreadsheet the same test is four =AVERAGE() cells over the quarter ranges and two rows of subtraction and division beneath them. It takes five minutes. I have watched two experienced admins argue for twenty minutes about whether a line was bending before anyone divided one number by another.
Tip: the quarter test only works on a full year of data. With six months you get one comparison, which is a number, not a pattern. If your retention is shorter than a year, the guide on rollups and retention explains how to keep a daily summary for longer without keeping every sample.
When a logarithmic scale tells the truth, and when it hides it
A logarithmic scale is a vertical axis where each gridline is ten times the one below it: 10, 100, 1,000, 10,000. On a log scale, multiplying by the same factor each month draws a straight line. Adding the same amount each month draws a curve that flattens out. The log axis is the reverse of an ordinary, or linear, axis. And it is the right tool for exactly one job: showing a metric that spans several orders of magnitude on one graph.
The log scale tells the truth when you are looking for runaway growth in a counter that has grown from ten to a hundred thousand. On a linear axis, the first eleven months are a flat line along the bottom and the twelfth is a wall. You cannot see whether the growth was steady or accelerating. On a log axis, runaway growth is a straight line whose steepness you can read, and a change in the rate shows as a kink. Error counters, retry counts, and queue depths that explode are candidates.
The log scale hides the truth when the thing you care about is a modest linear creep in a large number. Disk use going from 60 to 75 percent is a clear rising line on a linear axis and a nearly flat one on a log axis. The ratio between 60 and 75 is small. Latency drifting from four milliseconds to nine is the same story. And a log axis cannot show zero at all. So a counter that was zero for six months simply vanishes from the left of the graph. Use a log scale to look for acceleration and a linear scale to look for creep. Never present a log-scale graph without saying so. It is an honest chart that is easy to misread, and the guide on misreading network metrics has company for it.
Matching each shape to a response
Each shape has one correct first move, and getting the first move right is most of the value of knowing the shape.
A linear rise gets a ruler and a date. Extend the line to the threshold, write down the date it arrives, and put that date in the capacity register with an owner. If the date is more than a year away, look again next quarter. If it is inside six months, start the purchase conversation now; the conversation takes longer than the disk does.
A step gets an investigation before it gets anything else. Find the date from the graph, then go to the change log, the ticket queue, and the calendar for that day. Ask what went live, what was migrated, what job was edited. If the new level is intended and permanent, it is the new normal. In that case, rebaseline the metric so your alerts stop arguing with it. The guide on rebaselining walks you through it. If the new level is a mistake, fix the mistake and watch the line step back down.
A seasonal wave gets compared with last year and planned for the peak. Capacity has to fit the busiest week of the year, not the average one. A seasonal metric that is growing shows it as this year's peak standing higher than last year's. Thresholds that are right in the quiet season will be wrong in the busy one; the guide on dynamic thresholds covers letting the line move with the calendar.
A runaway curve gets attention today. Not a register entry, not a review date. Find the process, the loop, or the leak, and stop it. If you genuinely cannot find it, restart the thing that is growing and note the time, because the restart buys you one more cycle in which to look. Runaway growth is the only shape here where the right answer is to interrupt someone's afternoon.
A sawtooth gets its edges checked. Identify the job or reboot that makes the drop, and confirm it is still scheduled. Then look at the peaks and the troughs across the year as two separate lines. If either is rising, treat that rising line as its own curve and read its shape. A flat line gets a glance at its texture and, if it has none, a check that the collector is alive.
The curve shapes table
This is the reference I keep next to the year-view page. It is short on purpose: name the shape in ten seconds and move on to the response.
| Shape | What it looks like | Likely cause | First check | Planning response |
|---|---|---|---|---|
| Flat | Horizontal, with or without noise | Healthy and stable, or a collector that stopped updating | Is there any wobble at all? When was the last real sample? | None if healthy; fix the collector if dead |
| Linear | Straight rise, same amount per week | Steady accumulation: files, rows, devices, mail | Does the quarter-over-quarter change stay constant? | Extend the line to the threshold, record the date and owner |
| Step | Flat, one vertical jump, flat again | A project, migration, new site, or changed job | What changed on the date of the jump? | Rebaseline if intended; revert if a mistake |
| Seasonal | Repeating wave over weeks or months | Calendar-driven load: term time, month end, holidays | Does this period match the same period last year? | Plan for the peak; compare year over year |
| Runaway | Rise that gets steeper each month | Leak, loop, retry storm, replication backlog | Does the quarter-over-quarter change keep growing? | Find and stop the cause now; do not schedule |
| Sawtooth | Climb, sharp drop, climb, at regular intervals | Scheduled job, log rotation, reboot cycle | Are the peaks or troughs rising across the year? | Confirm the reset still runs; read peaks and troughs as their own curves |
Two shapes can share a graph. A seasonal wave riding on a linear slope is common, and a sawtooth whose peaks form a runaway curve is the classic memory-leak-on-a-schedule picture. When a graph fits two rows of the table, read the slower shape first. The rhythm is the texture. The trend is the story.
The step that was a backup job
Meridian Parts watched its WAN link between the warehouse and head office climb from about thirty percent utilisation to about seventy percent, and stay there. The graph went into the quarterly capacity review as a growth trend. A link upgrade was costed. And the quotation went through two approval meetings. In the third meeting somebody set the graph to a year. They noticed the change was a single vertical edge on a Tuesday, and asked what had happened that day. The answer was a backup job whose repository had been re-pointed from the warehouse's local storage to a share at head office. So the entire store had crossed the WAN every night for eleven weeks. The target was corrected in four minutes, the line stepped back down the next morning, and the quotation was withdrawn. The upgrade had by then travelled further through the approval process than the backup job's change request ever did.
The review was not careless. It was reading a level on a graph too short to show the edge. Every step looks like a slope if you start the graph after the step. Widen the window before you name the shape, and ask what happened on the date before you ask what to buy.
A shape-reading checklist
Run this against each metric on your creep register, the list of slow-moving metrics the previous guide had you build. It takes a minute per metric once the graphs are on one page.
- Set the graph to twelve months with the vertical axis starting at zero.
- Name the shape: flat, linear, step, seasonal, runaway, or sawtooth. If two apply, name both.
- If flat, confirm the line has noise and the collector has a recent sample.
- If linear, extend the line to the threshold and write down the date it arrives.
- If step, find the date of the edge and identify the change made that day.
- If seasonal, compare this period with the same period last year, not with last month.
- If runaway, run the quarter test to confirm, then escalate the same day.
- If sawtooth, check that the peaks and the troughs are level across the year.
- Record the shape, the date, and the response in your findings log.
The findings log is the one-line-per-finding record that the guide on the weekly trend review gives you a template for. Recording the shape matters more than it seems, because next quarter's shape only means something compared with this quarter's. Nobody remembers what they thought a graph looked like three months ago. I certainly do not.
Where to go next
Six shapes, six responses, one test that separates the two you will confuse. With that, a rising line stops being a worry and becomes a sentence. "Linear, reaches the threshold in eleven months, owner assigned" or "step on the fourteenth, change identified, rebaselined." The next guide, early-warning metrics worth a weekly look, gives you the catalogue of metrics to read these shapes on. The guide to alerting on trends shows how to have the monitoring system name the shape for you. The rest of the Trend Analysis and Degradation Creep pillar is here.
Put this into practice with SysPeep Monitor
Reading a curve needs a year of history you can put on screen with the axis fixed at zero. You also need the thresholds you will move when a step becomes the new normal. A central Windows service covers the whole habit. It collects over SNMP, WMI and SSH and keeps long-term graphs. It gives you thresholds, alert rules, dashboards and reports from one web console. The free trial installs on one machine.
Download Free TrialLearn About SysPeep