Sticky sessions are sold by duration. A provider offers sessions of 10, 30 or 60 minutes, and a buyer picks the one that covers their longest login, checkout or crawl. The number on the plan page reads like a promise. It is not one.
An advertised sticky session duration is the longest a session can last. How long it actually holds depends on the device behind the exit, and on residential proxies that device belongs to a real person. This guide explains why sessions end early, how to measure it, and how Atlas reports it.
What a sticky session is
A rotating residential proxy gives you a new exit IP for each request, or for each short burst. A sticky session asks the gateway to keep sending your requests through the same exit for a set time, usually by putting a session ID in the proxy username. If you need a refresher on the mechanism itself, the sticky session glossary entry covers it.
Sticky sessions exist because some targets tie state to an address. A login, a shopping cart, a multi-page form or a paginated search can break or trigger a security check if the IP changes halfway through. For those jobs the question is not “how long is the session?” but “how often does it last as long as I need?”.
Why the advertised duration is a ceiling
The gateway can only keep you on an exit while that exit exists. A residential exit is a device on a home or mobile connection, and those disappear all the time:
- The owner turns the device off, closes the app sharing its bandwidth, or puts it to sleep.
- A phone moves from Wi-Fi to cellular, or from one network to another.
- The home router restarts, or the ISP hands the household a new address.
- The gateway decides the exit is overloaded or unhealthy and moves you elsewhere.
None of these care about your session timer. A 30-minute session on a device that goes offline after 8 minutes ends at 8 minutes. The provider is not lying about the duration; it is quoting the setting, and the setting is an upper limit.
That is why the methodology page puts it bluntly: advertised TTLs are ceilings, and the gap between the ceiling and what you get is almost never disclosed.
The two ways a session ends
When an exit disappears, the gateway does one of two things.
It fails visibly. Your next request returns a proxy error, a connection reset or a timeout. Annoying, but your code knows something happened and can start a new session.
It swaps silently. The gateway routes your next request through a different exit and carries on, with the same session ID and no error. Your code thinks it is still on the same address. The target sees a new IP mid-session, and your login is logged out, your cart is emptied or your account is flagged.
The silent swap is the more damaging case, because nothing in the response tells you it happened. The only way to catch it is to check the exit IP yourself, on every request, against the one the session started with.
Hold rate: the figure that matters
Since the duration is a ceiling, the figure that describes a sticky product is its hold rate: the share of sessions that kept the same exit IP for the full advertised duration.
Say you open 100 sessions at 30 minutes each. If 85 keep their starting IP for the whole 30 minutes, the hold rate is 85%. The other 15 ended early, whether they failed loudly or swapped quietly. Two providers with the same advertised duration can have very different hold rates, and the hold rate is the one that predicts how many of your sessions you will have to restart.
Atlas tests sticky residential products with the same suite as rotating ones, plus TTL verification: it holds each session for its full advertised duration and records whether the exit IP the target saw survived. Session settings are left at each provider’s defaults, so the figure describes the product as sold. The result appears as “Sticky sessions held” in the side-by-side comparison. It comes from the latest full assessment, so it reads the same whichever time window you pick, and it shows N/A for a network where it has not been measured rather than counting it as zero.
Reading a hold rate
A hold rate needs a little context to be useful.
It depends on the duration you ask for. Longer sessions give devices more time to drop off, so a hold rate at 60 minutes will usually be lower than at 10 minutes on the same network. Compare like with like.
Short is often better than long. If your job needs 5 minutes, ask for a session not much longer than that. A long session buys nothing once the work is done, and it keeps you on an exit that may be getting less reliable.
Location narrows the pool. A session pinned to a small city has fewer devices to start from, and the gateway has fewer options when one drops. Expect a hold rate to vary by country.
It is not the same as rotation behaviour. Atlas also reports requests per exit IP for rotating products, which shows how long a “rotating” exit really stays in place. That is a separate question: whether rotation happens when it should, rather than whether a session holds when it should.
Designing jobs that survive early endings
Even a high hold rate means some sessions end early, so build for it:
- Check the exit IP on every request. Compare it with the IP the session started on. A change is a session ending, whatever the status code says.
- Keep each session’s work short. Split long flows into steps that can resume. The shorter the step, the less a dropped session costs.
- Save state as you go. Cookies, tokens and the current page should survive the session, so a new session can pick up where the old one stopped.
- Ask for the duration you need, not the maximum. It keeps you on fresher exits and makes your hold rate higher in practice.
- Restart cleanly. When a session ends, open a new session ID rather than retrying on the old one.
For login-heavy work, scraping sites that require login goes deeper on pinning accounts to sessions and keeping cookies alive. If you are still deciding whether a job needs sticky sessions at all, sticky vs rotating residential proxies covers the trade-off.
Testing a provider’s sticky sessions yourself
- Open at least 50 sticky sessions at the duration you plan to use, spread over a day so you see busy and quiet hours.
- In each session, send a light request every minute to an endpoint that echoes the IP it saw.
- Record the starting IP, and the minute at which the echoed IP first changed or a request first failed.
- Count the sessions that kept the starting IP for the full duration. That share is your hold rate.
- Plot how many sessions survived at each minute. The shape shows whether sessions fail steadily or fall off at a particular point.
Then set your result beside Atlas’s figure for the same provider. If they disagree, the duration, the location or the time of day you tested is usually the reason.