P50

Anyone who wants to measure the speed of a computer service never gets a single number. They get thousands of measured values, because every request takes a different amount of time. P50 is a way of condensing this flood of data into one value. All measurements are sorted by size, and the value exactly in the middle is taken. Half of all requests were faster, half were slower. Mathematicians call this average value the median; in technology it is usually called P50, pronounced “P fifty”.

Why the average lies when measuring wait times

The obvious approach would be the average: add up all the values, divide by the number of values. But with wait times, this often leads you astray. That’s because individual requests sometimes take absurdly long, for instance because a server is restarting or a connection drops. Such outliers pull the average upward, even though almost no users notice anything of the sort.

An example makes this clear. Nine requests each take 100 milliseconds, one takes 10 seconds. The average comes out to around one second, but the P50 remains at 100 milliseconds. Here, the P50 describes everyday reality far more honestly. This is exactly why it appears in almost every technical report on response times.

Still, it’s important to note: P50 alone is never enough. It says nothing about how bad the bad cases are. That’s why it is almost always mentioned together with P95 or P99. P99 is the value below which 99 percent of all measurements fall. It shows what the unluckiest users experience.

How the value is derived from the measurement series

The calculation itself is simple. Over a certain period, all measured values are collected, for example all response times within an hour. This list is sorted from smallest to largest value. Then the element in the middle is taken. With 1000 measurements, that is the 500th value of the sorted list.

The “P” stands for percentile, meaning hundredth shares. P50 means: 50 percent of the values lie below it. P90, P95, and P99 work according to the same pattern. Any percentile can be formed, but in practice these four have become the standard.

In real systems, raw data is rarely stored in full, because that would consume too much storage. Instead, monitoring tools work with histograms, meaning counters for value ranges. The P50 is estimated from these and is therefore somewhat imprecise. Another common mistake is simply adding up P50 values. The median of two servers is not the average of their two medians.

P50 in model cards, cloud contracts, and weather reports

P50 is most commonly encountered in reports about AI services. When a provider writes that its language model responds in 300 milliseconds, this almost always refers to the P50. The time until the first word output by a chatbot is also given this way. Anyone comparing such figures should check whether both providers are citing the same percentile.

The term also appears in contracts. Cloud providers promise customers certain response times and fix percentiles for this purpose. A promise regarding the P99 is considerably more expensive than one regarding the P50, because the rare outliers are the hardest to control.

Outside of computer science, the same idea appears under other names. With median income, one half of people earn more, the other half less. Weather services and project planners also use P50 estimates for rainfall amounts or completion dates. The principle is always the same: the middle instead of the average, so that individual extreme values do not distort the picture.