An average is a comfortable summary, but it is a poor guide to how bad things get. A percentile aggregate answers a different and often more operational question: what value is my process at or below most of the time, and how does the top of the distribution behave? The ninety-fifth percentile, or P95, tells you the level that only the worst five percent of readings exceed - a far more honest way to describe peak load or spiky behavior than an average that dilutes every spike into a sea of calm readings. Percentiles are how you see the tail an average smooths flat.
Percentile Aggregate in one line: A percentile aggregate returns the value below which a given fraction of samples fall in a window - P95 is the value 95 percent of readings stay under, P50 is the median. Unlike an average, percentiles describe the distribution and its tail, revealing peak loads and spikes that an average dilutes and hides. They are used when the extremes matter more than the typical.
A percentile is a position in the sorted distribution of values. To find the P95 of a window, you conceptually line up all the samples from lowest to highest and pick the value at the ninety-fifth-percent mark: ninety-five percent of the readings are at or below it, five percent are above. P50, the median, is the middle value, the one where half the readings fall on each side. P99 pushes further into the tail, the value only the top one percent exceed. Each percentile is a threshold that a known fraction of the data respects.
This is fundamentally different from an average, which is a computed center of mass rather than an actual observed position. An average can be dragged around by extreme values and does not correspond to any particular reading; a percentile is always an actual point in the distribution and describes where the bulk of the data sits relative to a chosen cutoff. The median in particular is a robust measure of the typical value because a few wild outliers barely move it, whereas they can pull an average noticeably off.
Percentiles come as a family precisely so you can probe different parts of the distribution. The median describes the center without being distorted by outliers. P90 and P95 describe the upper shoulder - the level that things routinely reach but rarely exceed. P99 and beyond describe the extreme tail, the rare worst case. Choosing which percentile to report is choosing which part of the behavior you want to characterize, from typical to nearly-worst.
The core weakness of an average is that it blends the calm and the extreme into a single number that represents neither. A pressure tag that sits quietly most of an hour and then spikes hard for a minute will have an average barely different from an entirely calm hour, because the spike is one small contribution to a long calm sum. Anyone reading only that average would conclude nothing unusual happened. P95, by contrast, reports the level the readings actually reached near their top, so the spike registers as the elevated tail it was.
This makes percentiles the right tool whenever the extremes are what matter. Peak load on a line, the pressure surges a vessel actually saw, the near-worst response of a slow control loop - all of these live in the tail of the distribution, exactly where the average is least informative. Reporting P95 or P99 answers questions like how high did it routinely get and how bad was the near-worst case, which are the questions safety, capacity, and reliability decisions actually turn on.
Percentiles are also the natural language for service-level and cycle-time expectations. Saying a cycle completes within some time on average tolerates a long tail of slow cycles hiding behind a good mean; saying the ninety-fifth-percentile cycle completes within a target holds the tail itself accountable. Because a percentile is defined by the fraction of the data that must respect it, it directly encodes a promise about consistency rather than about the middle, which is usually what an operational commitment really means.
In a cloud SCADA historian, computing a percentile aggregate over a window lets a platform like Merobix summarize spiky tags without letting the spikes vanish into an average. Trending P95 pressure per hour across a day draws the upper envelope of what the equipment actually experienced, which is a truer picture for anyone worried about excursions than a trend of hourly averages that flattens every surge. The same window can be summarized as median, P95, and P99 to show center, shoulder, and tail side by side.
The operational uses follow the questions each percentile answers. A reliability engineer sizing equipment or investigating stress wants the high percentiles, because those describe the loads the asset routinely approaches rather than the load it comfortably averages. An analyst wanting a robust sense of the normal operating level wants the median, which shrugs off the occasional bad reading that would tug an average. A team holding a cycle-time or response commitment wants a percentile threshold, because that is what actually encodes consistency.
Because percentiles describe distribution rather than magnitude, they also complement the value aggregates already in use. Pairing a median with a standard deviation characterizes both where the process sits and how spread out it is; pairing P95 with a plain average shows how far the tail runs above the typical. For a monitoring platform whose whole job is turning oceans of samples into decisions, offering percentile aggregates alongside averages means users can ask about the worst-case behavior directly, instead of hoping an average happened not to hide it.
P95 is the value that ninety-five percent of the samples in a window are at or below, and that only the top five percent exceed. It describes the upper shoulder of the distribution, so it captures how high readings routinely got without being thrown off by a single freak value. It is an actual observed position in the data, not a computed center like an average.
Whenever the extremes matter more than the typical. If you need to know how high a pressure surged, how bad the near-worst cycle time was, or the peak load an asset routinely approaches, a high percentile like P95 or P99 reports it while an average dilutes it into the calm readings around it. Use the median when you want a robust sense of the normal level that outliers cannot distort.
The median is the middle value of the sorted data, so a few extreme outliers barely move it - they change which value sits in the middle only slightly. An average, by contrast, is pulled toward extreme values because every reading contributes its full magnitude to the sum. When a tag has occasional wild readings, the median describes the typical level more faithfully than the average.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.