.bp-clean-small-figure{max-width:540px!important;width:100%!important;margin:22px auto!important;padding:10px!important;border-radius:20px!important;background:#fff!important;box-shadow:0 10px 28px rgba(7,18,5,.10)!important}.bp-clean-small-figure img,.bp-clean-small-image{display:block!important;width:100%!important;height:auto!important;border-radius:16px!important}.bp-clean-small-figure figcaption{font-size:14px!important;line-height:1.45!important;color:#4b5563!important;margin-top:10px!important;text-align:center!important}
Proxy capacity planning
Proxy Bandwidth, Threads and Concurrent Connections Explained
Proxy capacity is easy to misunderstand because providers may sell by IP, by gigabyte, by port or by concurrent connection. The right number depends on page weight, retry behavior and how much work happens at the same moment.
Separate the terms
Bandwidth, traffic and concurrent connections are not interchangeable
Bandwidth is often used loosely. It can mean a transfer allowance, such as 10 GB per month, or a transfer rate, such as megabits per second. Read the plan wording. “Unlimited bandwidth” may mean no fixed monthly byte quota while speed, ports, concurrent connections or fair-use rules still apply.
A concurrent connection is one connection in progress at a given moment. A thread is an application execution unit that may open one connection, several connections or none while it waits. Ten threads do not always equal ten proxy connections, especially when software reuses connections or downloads several resources in parallel.
Requests per second measures completed or attempted application requests over time. It is influenced by latency, target speed, object size, retries and concurrency. Raising the thread count can increase throughput until a real bottleneck appears; after that point it mostly creates waiting and failures.
Estimate traffic from a real sample
The safest calculation begins with one representative task. Measure the bytes transferred through the proxy, including images, scripts, API responses and retries. Then multiply by the expected number of tasks and add a sensible margin.
Do not base the estimate only on HTML size. Modern pages can load far more JavaScript, images, fonts and tracking resources than the initial document. A browser session is usually heavier than a focused API request. Conversely, blocking unnecessary images during a permitted automated QA job can cut traffic dramatically.
Compression and caching also matter. If the client accepts compressed responses and reuses cacheable assets, repeated visits can use less traffic. But a clean browser profile or proxy rotation strategy may reduce cache reuse. Measure the same conditions you plan to run.
Threads, connections and why more can be slower
| Metric | What it controls | Common mistake |
|---|---|---|
| Threads/workers | How much application work can proceed in parallel | Assuming each thread maps to exactly one connection |
| Concurrent connections | How many network connections are active at once | Ignoring provider, OS or target limits |
| Requests per second | Rate of attempted or completed requests | Counting retries as useful throughput |
| Latency | Time spent waiting for a response | Trying to solve long routes only by adding workers |
| Traffic allowance | Total data transferable in a billing period | Forgetting images, failed downloads and retries |
At low concurrency, adding workers can keep the connection busy while other requests wait. At high concurrency, queues grow, sockets compete, the destination rate-limits traffic and retry storms consume the remaining capacity. The useful number is successful tasks per minute, not the largest thread count the interface accepts.
Questions to ask before choosing a proxy plan
- Is pricing per IP, port, gigabyte or billing period?
- Does the allowance count both uploaded and downloaded traffic?
- Are simultaneous connections limited per proxy or per account?
- Is the advertised speed shared, capped or subject to fair use?
- What happens at the limit: throttling, rejection or extra charges?
- Are replacements and failed proxy traffic treated differently?
- Does the software open HTTP, HTTPS CONNECT or SOCKS connections?
For static datacenter proxies sold per IP, the practical constraint may be concurrency and target acceptance rather than a small bandwidth quota. For rotating products sold per GB, page weight becomes a direct cost. Compare the commercial model with the workload instead of comparing two headline prices that measure different things.
The dedicated versus semi-dedicated guide explains how sharing affects consistency. The datacenter versus residential guide covers the network-source trade-off.
Reduce wasted proxy traffic before buying more
Download only what the task needs
A visual QA browser needs the full page. An API integration may need only JSON. A health check may need a small endpoint and expected text. Match the request to the question; do not load an entire media-heavy homepage to prove that a proxy can connect.
Use bounded retries
Retrying every failure immediately can double or triple traffic while making the incident worse. Classify failures first. A 407 needs corrected authentication, a 429 needs delay, and a permanent 403 should not be retried twenty times without a change.
Reuse safe connections
Connection reuse can reduce handshake overhead for repeated requests to the same destination. Respect the behavior of the proxy and client library, especially when session isolation is required. Do not reuse cookies or authenticated sessions across unrelated accounts simply to save a small amount of traffic.
Track successful output
Report bytes per successful task, not only total bytes. A job that transfers 50 GB and produces mostly failures is not efficient, even if the proxy plan technically allows 500 GB.
A small capacity test before scaling
- Choose a representative target. Include the assets and response size the real job uses.
- Run one worker. Record latency, bytes and success rate for a clean baseline.
- Increase gradually. Try two, five and ten concurrent tasks instead of jumping to one hundred.
- Watch useful throughput. Stop increasing when completed work flattens or errors rise sharply.
- Repeat by region. Network distance can change both latency and the concurrency needed to keep a pipeline busy.
- Add a margin. Production has variance, but an enormous margin can hide inefficient software.
Use the Proxy Region Speed Test to compare target latency and the Proxy Tester for basic health. Neither replaces a controlled load test against a target you are authorized to use.
Keep the test responsible. Provider limits and destination rules still apply, and a capacity exercise should never become an accidental denial-of-service test.
Capacity notes
Why real traffic is usually higher than the spreadsheet estimate
Early estimates often count only successful response bodies. Production also transfers request headers, TLS handshakes, redirects, failed responses, retries and assets loaded by the browser after the main document. The gap can be small for a compact API and large for a modern website. Measure at the proxy or network layer when billing is based on transferred bytes.
Upload traffic can matter
Most browsing is download-heavy, but form submissions, file uploads and API requests can send substantial data. Ask whether the provider counts both directions. If a workflow uploads screenshots or documents, include that volume and the possibility of retrying a partially completed transfer.
Latency changes the useful concurrency
A worker waiting 800 ms can complete fewer sequential tasks than one waiting 80 ms. Some additional concurrency can keep the pipeline productive, but the right number depends on response time. That is why a setting copied from a low-latency local test may perform badly against a distant target.
Connection reuse changes the math
HTTP keep-alive and multiplexing can reduce repeated connection setup, depending on the client, proxy and destination. Browser developer tools may show several resource requests sharing fewer underlying connections. Count what the system actually opens instead of multiplying threads by pages.
Retries consume capacity twice
A timeout can transfer part of a response before the client gives up, and the retry starts again from zero. When error rates rise, bandwidth per completed task often rises too. Set maximum attempts, use backoff and fix persistent authentication or content errors rather than paying to repeat them.
Plan for the busy hour
A monthly average can hide a short peak where every scheduled job starts together. Stagger non-urgent work and measure the highest normal concurrency. Capacity planning should cover realistic peaks without assuming that every theoretical task must run at the same second.
Technical references
Parallel transfers need measured limits
The Everything cURL guide to parallel transfers shows how a client controls simultaneous work, while MDN connection-management documentation explains persistent connections and their effect on repeated requests. Neither number replaces a target-specific load test, but both help interpret what the client is doing.
Quick answers
Proxy bandwidth FAQ
What does proxy bandwidth mean?
It may refer to total traffic allowed in a billing period or to transfer speed. Read the plan definition because unlimited traffic can still have speed, connection or fair-use limits.
Do ten threads use ten proxy connections?
Not necessarily. A thread may open several connections, reuse one or wait without an active connection. Measure connections at runtime.
How much proxy traffic do I need?
Measure transferred bytes for one representative completed task, multiply by task count and include a realistic retry and growth margin.
Will more concurrent connections make the proxy faster?
Only until another bottleneck appears. Excess concurrency can increase queues, rate limits, timeouts and retries.
How can I reduce proxy bandwidth usage?
Request only necessary resources, enable appropriate compression and caching, classify failures before retrying and track bytes per successful task.
Choose capacity from a measured sample
Start with a realistic target and a small proxy plan. Track transferred bytes, completion rate and latency, then scale from evidence instead of a guessed thread count.
API and CI workloads should pair these connection limits with the API proxy testing workflow, including bounded retries and timing budgets.
Turn connection limits into a buying estimate with the proxy pool size calculator guide, using peak work, service time, per-proxy concurrency and reserve.
Related BuyProxies guides
These related pages help connect the setup, troubleshooting and buying decisions instead of treating each proxy problem in isolation.
Proxy authentication failed
Use this guide as the next step when you need a more specific proxy workflow.
HTTP 407 proxy authentication required
Use this guide as the next step when you need a more specific proxy workflow.
Proxy timeout errors
Use this guide as the next step when you need a more specific proxy workflow.
Best proxies for scraping
Use this guide as the next step when you need a more specific proxy workflow.
How many proxies for scraping
Use this guide as the next step when you need a more specific proxy workflow.
How many proxies for multiple accounts
Use this guide as the next step when you need a more specific proxy workflow.