Latency
The delay between sending a request through a proxy and getting a response, driven largely by physical distance to the exit.
Latency is the waiting: the round trip from your request leaving to the answer arriving, measured in milliseconds. Some of it is negotiable and some of it is physics, and knowing which is which saves a lot of pointless tuning.
The physics part sets a floor nobody can go under. Light in optical fibre travels at roughly 200,000 kilometres per second, about two thirds of its speed in vacuum, which works out to about 5 milliseconds per 1,000 kilometres one way and 10 milliseconds per 1,000 kilometres for a round trip. London to New York is around 5,600 kilometres, so the theoretical minimum round trip is roughly 56 milliseconds. Real measurements land nearer 70 to 80, because cables do not follow great circles and every router adds a little. No provider, plan or protocol beats that floor.
Which is why a proxy's location is a performance decision as much as a targeting one, and why the geometry matters more than people expect. Your traffic no longer goes to the destination, it goes to the exit and then to the destination. If you are in Europe, your exit is in New York and your target is in Europe, you have bought two ocean crossings for a request that would have travelled almost no distance directly. The worst latency setups are usually not slow proxies, they are unnecessary triangles.
Address type contributes too. A datacenter proxy sits on well-connected infrastructure with short paths to backbone routes. A residential address is behind consumer broadband, and a mobile address is behind a radio link whose latency varies with signal and cell congestion in ways nobody controls. That ordering, datacenter fastest and mobile most variable, is structural rather than a quality difference between providers.
Whether any of it matters depends entirely on the shape of your work, and this is where most of the wasted effort happens. Bulk collection is governed by throughput across many parallel connections, so a hundred requests each taking 300 milliseconds finish in about the same wall-clock time as a hundred taking 100 milliseconds, provided you run them concurrently. Latency dominates only when steps are sequential and each one must finish before the next begins, which is the case for logins, checkouts and any paginated crawl following a server-side cursor.
So the practical rule is to fix the geometry first and the concurrency second, and to stop optimising latency for a workload that is not sensitive to it. Put the exit near the target, not near you, because the leg that repeats is the one between the exit and the destination.
How HProxy handles it
We only quote a representative latency figure on a country page when our own vantage point can time that country's pool fairly. For distant countries we report composition and uptime instead, because a number measured across an ocean says more about where we were standing than about the pool.
Frequently asked questions
Why is my proxy so slow?
Check the geometry before blaming the proxy. If you are in one region, your exit is in another and your target is back near you, every request pays for two long crossings that a direct connection would not. After that, consider address type, since residential and mobile are structurally slower than datacenter, and then whether the delay is really latency or actually bandwidth, which is a different problem with a different fix.
Should the proxy be close to me or close to the target?
Close to the target, in almost every case. Your connection to the exit happens once per request, but the exit's conversation with the destination is where the work happens, including TLS negotiation and any redirects. Putting the exit near the destination shortens the leg that repeats.
Does latency matter for web scraping?
Much less than people assume, as long as you run requests concurrently. Bulk collection is bound by throughput, so latency per request mostly disappears into parallelism. It matters when steps are sequential and each depends on the last, such as logging in, walking a checkout, or following a paginated cursor, where every millisecond is paid once per step.
What is a good latency for a proxy?
There is no absolute number, only one relative to the distance involved. Compare against the physical floor for that route: about 10 milliseconds of round trip per 1,000 kilometres. If a nearby exit is returning several hundred milliseconds, something is wrong beyond geography. If a transcontinental exit is returning 80, that is close to the best the route allows.
Back to the full glossary.