mirror of
https://github.com/pnpm/pnpm.git
synced 2026-06-28 09:55:39 -04:00
65c9bef283
## What The pnpr install accelerator is a **remote** server, but the integrated benchmark ran it on **loopback** (RTT ≈ 0), which hides the round-trip cost that dominates a real install — and that pnpr exists to reduce. This injects network latency so the benchmark measures pnpr as the remote service it is in production. ## How A dependency-free, synchronous latency-injecting TCP proxy (`latency_proxy`) plus two knobs on `integrated-benchmark`: - **`--pnpr-latency-ms`** — fronts each `pnpr@<rev>` server, so the client↔server link pays the given round trip (half each direction). - **`--registry-latency-ms`** — fronts the registry for the direct (`pacquet`/`pnpm`/`--with-pnpm`) targets, so a direct install crosses the same network. `pnpr@<rev>` targets keep a **direct (fast) registry link** — that models a warm, colocated server, so pnpr's advantage shows up as **fewer round trips, not a faster backend**: ``` direct target: client → [latency proxy] → registry pnpr target: client → [latency proxy] → pnpr server → (direct) → registry ``` The workflow sets both equal (`50ms`) so the in-run pnpr-vs-direct ratio is fair and the `pnpr` Bencher testbed (pnpr@HEAD vs pnpr@main) becomes **sensitive to protocol round-trip-count changes** — which is what makes the upcoming protocol work (collapsing the 3-round-trip handshake/install/files flow) measurable on main. See #12165 for that plan. ## Notes - **Latency only, no bandwidth cap.** The public registry is CDN-backed and CI runners are fast, so install time is latency/round-trip bound, not throughput bound — a bandwidth cap would be overly pessimistic. A high-ceiling, opt-in bandwidth knob can follow if a slow-link scenario is ever wanted. - Both flags **default to `0`** (current behavior unchanged); the registry proxy is also skipped in `--registry=npm` mode (already remote). - The proxy is unit-tested (a round trip through it reflects the injected latency). `cargo check`/`clippy`/`fmt`/`dylint` clean. - One caveat the proxy does **not** model: TLS-handshake round trips and HTTP/2 multiplexing of a real CDN — it reproduces propagation delay, the dominant and relevant factor here, not a byte-exact replica of registry.npmjs.org.