More unique IPs than any network on the planet, in the world's largest residential pool. That's Shifter.More unique IPs than any network, in the world's largest pool. That's Shifter.

See the numbers

Proxy Latency: How to Read a Latency Chart

What p50, p90 and p99 mean on a proxy latency chart, why the median hides the slow requests that stall a scraper, and which questions to ask before you trust any latency figure.

Elena Petrova Data Science · Updated November 12, 2026 · 7 min read
p50p99
Requests by response time. The amber tail is the slow part.
Performance
On this page
  1. What a latency chart actually shows
  2. Why the median hides the tail
  3. Percentiles cannot be averaged
  4. Which latency are you looking at?
  5. Reading the shape
  6. What latency does not tell you
  7. A checklist for any latency figure
  8. Frequently asked questions

Most proxy latency figures are a single number, and a single number is the least useful way to describe how long requests take. Latency is a distribution. Most requests land in a narrow band, a few take much longer, and the few are what decide whether your job finishes on time.

Residential exits make the spread wider than most people expect. A request through residential proxies travels from your client to the provider’s gateway, out to a consumer device on a home connection, on to the target, and all the way back. Any of those hops can have a bad second. This guide explains how to read a chart that shows the whole spread, and what to ask before you trust one.

What a latency chart actually shows

A latency chart is usually a histogram: response times along the bottom, the number of requests in each time bucket up the side. For a residential network it has a recognisable shape. A tall peak on the left, where most requests finish, and a long thin tail stretching to the right, where the slow ones sit.

Percentiles are markers on that shape:

  • p50, the median. Half of requests finished faster than this, half slower.
  • p90. Nine in ten requests finished faster. One in ten took at least this long.
  • p99. Ninety-nine in a hundred finished faster. This is where the slowest one percent starts.

The gap between p50 and p99 is the most informative part of the chart. A narrow gap means a network behaves the same on its good and bad minutes. A wide gap means the median is describing a network you will only meet some of the time.

Why the median hides the tail

The median is a fine description of one request. It is a poor description of a job, because jobs are made of many requests and the slow ones add up.

Take a made-up example. Say a network’s median is 600 ms and its p99 is 6 seconds. One request in a hundred lands in that tail. If your scraper fetches 20 pages in sequence for each record, the chance that a given record hits at least one tail request is about 1 in 5 (one minus 0.99 to the power of 20). A fifth of your records now take seconds longer than the median suggests, and the median has not moved at all.

Volume makes it worse. At 500,000 requests a day, one percent is 5,000 slow requests. If each one holds a worker or a browser tab open, your throughput is set by the tail, not the peak. This is why two networks with the same median can produce very different job times.

Percentiles cannot be averaged

A common mistake in comparison tables is averaging percentiles: taking the p99 for each country or each hour and averaging them into one p99. The result is not a p99 of anything.

Percentiles come from the shape of the whole distribution. To combine two sets of measurements correctly you merge the underlying histograms first, then read the percentile off the merged shape. Atlas does it this way: each time slot’s histogram is merged into the window at equal weight, so a quiet overnight hour counts as much as a busy afternoon, and no figure is an average of averages. The latency definitions on the methodology page give the exact rules.

If a published chart does not say how its percentiles were combined, assume they were averaged and treat the tail figure with suspicion.

Which latency are you looking at?

“Latency” covers at least three different measurements, and providers tend to quote the smallest.

Connect time. How long it takes to open a connection to the gateway. On an idle connection to a nearby gateway this is fast, and it says nothing about the residential exit or the target.

Time to first byte. From sending the request to the first byte of the response. It includes the trip through the exit and the target’s own processing time.

Total time. From request start to the last byte. This is what your code waits for, and on a heavy page it can be much longer than the first byte.

Atlas measures total time under load, through the proxy, on a plain-HTTP target. Plain HTTP is a deliberate choice. HTTPS adds a CONNECT and a TLS handshake, roughly doubling the round trips, and mixing HTTP and HTTPS targets would make a chart reflect which targets were tested rather than how the network performs. Other targets still report success and blocks; they just do not feed the latency figure.

Load matters too. A network answering a handful of requests can look very different from the same network at real concurrency. Atlas records light-load latency and measures how much each region slows down at census concurrency, then shows both. If a chart does not say how much traffic was running while it was drawn, it is probably the light-load picture.

Reading the shape

Once you know what was measured, the shape of the chart tells you more than any single percentile.

A tight peak with a short tail. Consistent exits and a gateway that is not struggling. The best case.

A normal peak with a long, flat tail. Most exits are fine, but some fraction of requests hit slow devices, distant exits or retries inside the gateway. Watch the p99 over time; if it moves while the median holds still, the tail is load or time-of-day dependent.

Two peaks. Two populations of requests. Often a mix of nearby and far-away exits, or two kinds of peer devices with different connections. Split by country or by exit network before drawing conclusions.

A spike at the timeout. A cluster of requests piled up at exactly the client’s time limit. Those are not slow responses, they are failures, and they should be counted as failures, not as latency.

What latency does not tell you

A latency chart only describes requests that came back. It says nothing about whether what came back was useful.

A block page or a CAPTCHA can return faster than the real page, because the target did less work. A network that is quick because it is being refused will look excellent on a latency chart. Always read latency next to the success and challenge rates for the same window and the same targets.

If you want to cut the latency you do control, connection reuse, sensible geo filters and fail-fast timeouts help more than switching providers; reducing latency with residential proxies walks through each one. For how latency feeds into a ranking, see the companion guide on judging the fastest residential proxies.

A checklist for any latency figure

Before you trust a latency number, from a provider, a review site or your own test, check:

  1. Which percentiles? A median alone hides the tail. Ask for p90 and p99 too.
  2. Which time? Connect, first byte or last byte. Only the last is what you wait for.
  3. Under what load? Idle traffic flatters every network.
  4. Over what window? One quiet hour is not a week of traffic.
  5. Combined how? Merged histograms, or averaged percentiles.
  6. Which target? HTTP and HTTPS, light and heavy pages, are not comparable.
  7. With what success rate? Latency without the failure share beside it can mislead.

To run your own measurement as a distribution rather than a single number, the walkthrough on testing residential proxy speed and success rate has scripts you can adapt. Then put your p99 next to the side-by-side latency rows in Atlas and see whether the tail you measured matches the one published.

Frequently asked questions

What is a good latency for residential proxies?

There is no single good figure: it depends on the target, the exit country and the load you run. Compare networks on the same target and country, use p50 for typical requests and p99 for the slow tail, and read latency next to success rate, because a fast block is still a failure.

What is the difference between proxy latency and response time?

The words are used loosely. In Atlas, latency is the total time from request start to the last byte, through the proxy, on a plain-HTTP target. Other tools may quote connect time or time to first byte, which are shorter, so check what a figure measures before comparing it.

Why does proxy latency vary between different locations?

Because of the physical distance between the exit and the target, the routing paths in between, the home connection behind the exit and congestion along the way. An exit closer to the target server is often faster, but routing and network conditions can change that.

Get started

Watch your own tail not just the median

Start a free trial of Shifter's residential network, log the time of every request you send, and read the p99 before you decide anything.

3-day money-back guarantee. Cancel anytime.