Node HandbookNODE HANDBOOK中文
Nodes · measurement

How to read a proxy speed test without being misled

A latency figure answers one narrow question. Treating it as “this node works” is the most common mistake, and it is why a node can look healthy while every page fails.

Node Handbook editorial team ·

Four conditions to fix before comparing

  1. Time. Record it, including the hour. Peak and off-peak are different networks in practice.
  2. Local network. Wi-Fi versus Ethernet, and mobile versus fixed line, are not comparable.
  3. Target. The same host and the same file for every node you compare.
  4. Client and rule set. The same client, the same rules. A rule change mid-test invalidates the comparison.

Without those four, you are measuring the difference between tests rather than the difference between nodes.

The six fields and what each one answers

FieldQuestion it answersWhat it cannot tell you
Latency (RTT)How long a probe took to returnThroughput, or whether a real request completes
JitterHow much that time variesWhether the variation matters for your specific use
Packet lossHow many probe packets never returnedAnything about HTTP or TLS outcomes
DownloadBytes per second for a specific file at a momentSustained bitrate for streaming, or behaviour under many connections
UploadThe same in the other directionWhether the target service accepts your traffic at all
Exit regionWhere the traffic appears to originateWhether the service treats that region as supported

Why “no loss” is not “available”

Many client latency displays are produced by a probe against an endpoint the provider chooses. That probe can succeed while an actual HTTPS request through the proxy fails, because the two use different paths, different TLS handling and sometimes different outbounds.

So record two separate things: the probe result, and the result of real requests. When you write it up, keep the names distinct — “1 of 10 requests failed” is a request failure rate; “10% packet loss” is a different measurement and you probably did not make it.

The correct reading order for a test image

  1. When and where. Time, network, client version.
  2. What was measured. Probe latency, or an actual request? Against which endpoint?
  3. Stability, not just the average. Look at the spread; one 800 ms spike matters more than a 20 ms improvement in the average for interactive use.
  4. Throughput at a stated file size. A 1 MB download does not describe sustained streaming.
  5. Exit region, verified independently. Two IP data sources, not the node label.
  6. Real target behaviour. The actual site or service, with the actual account, at the actual hour.

A screenshot that only shows a latency column answers step two and nothing else.

Peak-hour testing

Shared capacity means the same node can be excellent at 03:00 and unusable at 21:00. If your use case is evening streaming or video calls, the only meaningful test is at that hour. Testing off-peak and extrapolating produces confident, wrong conclusions.

Common misreadings

What people concludeWhat the data actually supports
“Low latency, so it is fast”The probe returned quickly; throughput is untested
“No loss, so it is stable”The probe packets returned; behaviour under load is untested
“Downloads at 80 Mbps, so streaming is fine”One file transferred quickly; sustained bitrate and platform checks are untested
“Works for me, so the node is fine”One path at one moment worked

Limits

This page describes how to read measurements; it does not publish measurements of any specific node or provider, and it makes no stability claim about free sources. A short test describes a short interval. Nothing here predicts long-term availability, privacy properties or whether a service will accept a given exit.