Capacity Trending

Back to Server Setup · Resource Monitoring · Growth Projections · Capacity Planning

Tracking usage over time to flag when a server is approaching its limits well before it becomes urgent. The difference from resource monitoring is the tense. An alert tells you a threshold has been crossed, which is a problem you have now. A trend tells you when one will be, which is a purchase order you can raise calmly.

1. Produce a Date, Not a Percentage

“Disk is at 74%” is not actionable, because it does not say whether that is fine. “This filesystem fills in about seven months at the current rate” is actionable, and it is the same data with a line fitted through it.

So the useful alert is time to exhaustion: warn when any resource is projected to run out within the lead time needed to do something about it. On a cloud instance that might be a fortnight; where new hardware has to be ordered and approved, it could be three months. Same question, very different thresholds — and the lead time is what sets them, not a round number.

2. Trend the Things That Actually Run Out

• Disk space and inodes, per filesystem. The most common and most predictable exhaustion there is.
• Memory, watching for the slow climb that indicates a leak — a sawtooth that never quite returns to where it started is visible in a month of data and invisible in a day.
• CPU at peak, not average. Trend the daily peak, because the average can stay flat while peaks grow until they hit the utilisation bend described under workload requirements.
• Database size and table growth, which determine backup duration and restore time as much as storage cost.
• Connection counts and queue depths, which have hard configured ceilings and fail abruptly when reached.
• Backup window duration. The one that catches people: backups grow with the data while the window stays fixed, so the constraint arrives on a date you could have calculated a year earlier.

3. Fit the Trend Honestly

A straight line through recent points is often enough, and it is better than nothing by a wide margin. But a few things routinely make the naive projection wrong:

• Seasonality. A retailer's December is not their January. Compare like-for-like periods rather than fitting across them.
• Steps are not trends. A migration that added 200 GB overnight is an event, not a growth rate, and including it produces a nonsense projection.
• Compound growth curves upward. A linear fit to something growing by a percentage each month systematically predicts the exhaustion date too late — see the arithmetic under growth projections.
• Enough history. A week of data projects noise. A quarter projects a trend.
• Say how confident you are. “Between five and nine months” is more useful and more honest than a single date presented as fact.

4. Make It a Review, Not a Dashboard

Trending only works if someone looks. A short recurring review — monthly for most estates — that goes through the projections, notes what changed, and turns anything inside the lead time into a dated action with an owner. Fifteen minutes, and it is the difference between capacity being planned and capacity being an emergency.

It is also where the opposite finding shows up. Resources trending flat or downward on over-provisioned machines are money being spent for nothing, and the same review that prevents outages also funds itself — the right-sizing candidates from the infrastructure audit, kept current.

5. Connect It to the Triggers

Growth projections produced a list of triggers — when X crosses Y, do Z, and Z takes W weeks. Capacity trending is what watches them, and it closes the loop by comparing the original forecast against what actually happened. A forecast that was wrong is useful information: it says the assumptions need revisiting, which is a better outcome than discovering it at the ceiling.

What You Get

• Trend analysis on every resource that can be exhausted, with enough retained history to make it meaningful.
• Time-to-exhaustion alerts set against your real lead times rather than round percentages.
• Projections that account for seasonality and exclude step changes, with a stated confidence range.
• A monthly review producing dated actions with owners.
• Over-provisioning flagged alongside exhaustion, so the review works in both directions.