Rotating proxy

Definition

A rotating proxy assigns a different exit IP to each outgoing request — or, in some configurations, at a fixed time interval (e.g., every 1 minute). The traffic distributes across the pool; no single target-side endpoint sees the same origin IP for more than one request.

When rotation is the right tool

AI workloads that benefit from rotation:

1. Bulk training-corpus scraping. Pulling millions of URLs across many hosts, where each request is independent and the aim is to distribute per-IP request rate below the target's limit.

2. Training-data diversity sampling. Explicitly wanting the sampled content to come from many different source origins, for corpus-level diversity.

3. Single-shot LLM evaluation. One prompt, one response, capture and move on. No need for session continuity.

4. Parallel independent tasks. 1000 independent tasks, each making one request. Rotation just means each task happens to use a different IP.

When rotation breaks things

Rotation is wrong for:

• Multi-turn agent evaluation (breaks state)
• Authenticated sessions (breaks login cookies)
• LFS-resumable dataset pulls (breaks resume logic)
• Any workflow that expects the target to treat sequential requests as from "the same user"

For those, use sticky sessions instead.

Rotation granularity

Not all rotation is per-request. Common granularities:

• Per-request: every outgoing request gets a new exit
• Per-minute, per-5-minutes, per-30-minutes: exit stays for a window, then changes
• Per-target-domain: same exit for all requests to the same hostname within a session
• Per-task: same exit for the lifetime of an agent task or similar logical unit

SquadProxy exposes these via X-Squad-Session — per-request, sticky-Xm, or a custom session identifier.

Related

• Sticky session glossary entry
• Residential proxy glossary entry
• Residential vs datacenter for AI