Percentiles Explained: p50, p95 and p99 Latency

Quick answer: a percentile answers the question "how bad is it for the slowest few percent of requests?". If your p95 latency is 120 ms, 95 out of every 100 requests took 120 ms or less, and 5 took longer. p50 is the median (the middle request), p95 is the slow end of normal, and p99 is the very slow end. Averages hide these slow requests; percentiles expose them.

Why averages are not enough

Imagine ten requests: nine take 40 ms and one takes 800 ms. The average is (9 × 40 + 800) / 10 = 116 ms. That number describes none of the requests. Nine were fast; one was awful. A user who loads a page making 50 requests will very likely hit several slow ones, and the page feels slow even though the average looked acceptable.

Latency distributions are almost never bell-shaped. They have a tight cluster of normal values and a long tail of occasional slow ones (queueing, retransmissions, congestion, garbage collection on the server, a Wi-Fi retry). Statisticians call this a skewed or heavy-tailed distribution. For such data, a single average is misleading, and percentiles are the standard tool.

What is a percentile?

The n-th percentile is the value below which n% of the observations fall. Sort all measurements from smallest to largest; the p95 is the value at the position that has 95% of measurements at or before it.

Name	Also called	Meaning	Typical use
p0	Minimum	The fastest observation	The floor of the path
p50	Median	Half of requests are faster, half slower	A typical request
p90		9 in 10 requests are at or below	General quality
p95		19 in 20 are at or below	Common service-level target
p99		99 in 100 are at or below	Critical paths, tail behaviour
p99.9	Three nines	999 in 1000 are at or below	Very large services
p100	Maximum	The slowest observation	Worst case

How to calculate a percentile: the nearest-rank method

There are several definitions of percentile. GlobeLatency uses the nearest-rank method, which is simple and always returns a value that was actually measured.

1. Sort the N measurements from smallest to largest.
2. Compute the rank: rank = ceil(P / 100 × N), where P is the percentile you want.
3. The percentile is the value at that rank (counting from 1).

Worked example with 10 passes

Measured passes (ms): 38, 41, 40, 39, 44, 42, 40, 43, 41, 210. Sorted: 38, 39, 40, 40, 41, 41, 42, 43, 44, 210.

• p50: rank = ceil(0.50 × 10) = 5, value = 41 ms.
• p90: rank = ceil(0.90 × 10) = 9, value = 44 ms.
• p95: rank = ceil(0.95 × 10) = 10, value = 210 ms.
• p99: rank = ceil(0.99 × 10) = 10, value = 210 ms.
• Average: 57.8 ms. Median 41 ms. The average is pulled up by one spike; the median shows what is typical.

Worked example with 20 passes

With 20 sorted passes, p95 rank is ceil(0.95 × 20) = 19, so p95 is the 19th lowest value and only the single slowest pass is above it. p99 rank is ceil(0.99 × 20) = 20, the maximum. You need about 100 samples before p99 can differ from the maximum.

How many samples do you need?

Samples (N)	p95 rank	p99 rank	Meaning
3	3	3	Both equal the worst pass
10	10	10	Both equal the worst pass
20	19	20	p95 is distinct, p99 is the max
60	57	60	p95 skips the three slowest; p99 is the max
100	95	99	Both are distinct and meaningful

This is the honest limitation of short tests: with the default 3 passes the p95 and p99 you see equal the worst pass. They become informative with 20 passes, and p99 only really shines in long-running monitoring.

Percentiles in GlobeLatency

• Test mode: p95 and p99 are computed over the passes of each destination (1 to 20). Choose 20 passes when you want a useful p95.
• Map and Monitor: the "95th percentile" in the detail panel is computed over the last up to 60 readings of that destination (about 30 minutes at the default interval).
• The nearest-rank method is used in both.

See Reading your results and the Monitor guide for where these values appear.

How to interpret p50, p95 and p99 together

Pattern	What it means	Typical cause
p50 ≈ p95 ≈ p99	Very consistent connection	Wired, uncongested path
p95 about 1.5 to 2 times p50	Normal variation	Typical home or office network
p95 more than 3 times p50	Frequent spikes	Wi-Fi interference, congestion, bufferbloat
p99 far above p95	Rare but severe stalls	Retransmissions, route flaps, overloaded server
p50 high, spread small	Consistently slow, not unstable	Distance, slow access link, VPN

The "Unstable" flag in Test (worst pass more than 3 times the average) is a quick, short-test cousin of the second-last pattern. For the related idea of variation between consecutive readings, read jitter explained.

Percentiles and user experience

A page that makes 40 requests experiences the maximum of those requests in practice. If each request has a 1% chance of being slow, the chance that at least one of 40 is slow is 1 - 0.99^40 ≈ 33%. So a p99 problem for a single request becomes a "one in three page loads feels slow" problem for a whole page. This is why large services track tail percentiles and not averages. For websites and APIs see latency for websites and APIs; for real-time use see latency for gaming and video calls and VoIP.

Service level objectives (SLOs)

Teams often write goals such as "p95 under 200 ms and p99 under 500 ms over 30 days". Such statements are meaningful because they bound the slow tail, which average-based goals cannot. If you operate a site or API, measure from users' locations, not just from the data centre, because network latency adds to server time. A test from a few regions with custom destinations is a good start.

Why you cannot average percentiles

A common mistake: computing the average of p95 values from several servers or days to get an overall p95. That is mathematically invalid. Percentiles must be computed from the pooled raw samples (or from properly merged histograms). Averaging p95s usually understates the true tail.

Common mistakes and myths

• "p99 is the worst case." No, p100 (the maximum) is. p99 ignores the slowest 1%.
• "p95 from a three-pass test is reliable." It equals the worst pass.
• "The average is always fine to use." Not when the distribution has a tail, which is almost always the case with latency.
• "Lower percentiles mean lower averages." Not necessarily; the relationship depends on the shape of the distribution.
• Averaging percentiles across hosts or time periods.

What to do next

Compare central measures in mean, median, min and max, learn what causes spikes in jitter and bufferbloat, and then run a 20-pass test to see your own p95.