Skip to content

Capacity & Correctness

What Is the Saturation Point?

The saturation point is the load at which throughput stops rising even as more users arrive — the moment a resource runs out and queueing takes over.

Also known as: Knee Point, Saturation

How to spot it on a graph

Plot achieved throughput and response time against active virtual users on the same time axis. Before saturation, throughput climbs in step with users and response time stays roughly flat. At saturation the two lines split: throughput flattens, and response time starts climbing.

The split is the finding. Past that point every additional user contributes waiting rather than work, and load beyond it tells you nothing new about capacity — only about how the system degrades.

Something specific ran out

Saturation is never abstract. Some single resource hit 100% utilisation and everything behind it queued. Finding which one turns an observation into an action.

  • CPU on the application tier — the easiest case, and usually the cheapest to fix by scaling out.
  • A connection pool, where the tell is rising response time with idle CPU everywhere.
  • Database contention — lock waits, a hot row, or a query whose plan flips once the working set stops fitting in memory.
  • Disk or network I/O bandwidth.
  • A downstream quota or rate limit, in which case the ceiling is not yours to raise.

Saturation is not the breaking point

A saturated system is still serving traffic, just more slowly. The breaking point — where errors spike and the service stops being usable — is normally some distance beyond it, and a breakpoint test is what measures the gap.

That gap is your safety margin, and it is worth knowing. A system that degrades gently between saturation and failure gives operators time to react; one that goes from saturated to dead in seconds does not, and needs load shedding rather than more capacity.

Run below it, not at it

Saturation is a measurement, not a target. Systems run at their saturation point have no absorption capacity: an ordinary traffic fluctuation, a slow upstream, or one restarted instance pushes them straight into queueing, and response times climb far out of proportion to the extra load.

The usual guidance is to size for a steady state well below the measured ceiling — commonly around 60 to 70 percent — so that normal variance stays in the flat part of the curve. The saturation point tells you where the wall is; the operating point should be some distance in front of it.

Related terms

Ready to push your limits?

Start with 500 VUs, real cloud infrastructure, and real metrics — no credit card required. Upgrade only when you outgrow it.

Free forever — 500 VUs, 10-minute runs, no credit card