perf(benchmarks): measure actor wake with native Rust SQLite - #5772

Stack for rivet-dev/actors

• #5772 perf(benchmarks): measure actor wake with native Rust SQLite 👈
• #5777 perf(guard): deliver waiting HTTP requests with actor Start
• #5774 perf(actors): preload bounded SQLite pages with actor Start
• #5776 feat(envoy-protocol): add versioned actor startup payloads
• #5778 perf(rivetkit): skip unchanged bootstrap writes on actor wak …
• #5773 perf(envoy-client): disable Nagle for native WebSocket traff …
• #5775 feat(pegboard): prototype Guard-owned actor leases
• main

Get stack: forklift get 5772
Push local edits: forklift submit
Merge when ready: forklift merge 5772

_{change kmlxvnqm}

Review: perf(benchmarks): measure actor wake with native Rust SQLite

This adds a new opt-in benchmarks/guard-leasing crate: a Rust actor + HTTP harness for measuring cold-start/wake latency for the Guard leasing prototype. Overall this is a self-contained, low-risk addition (new workspace member only, no production code paths touched). A few notes:

Code quality

• Cargo.toml:4 - the new member entry uses a tab for indentation (\t"benchmarks/guard-leasing",) while every other entry in members = [...] uses two spaces. Worth fixing for consistency (mixed whitespace in this list).
• benchmarks/guard-leasing/src/main.rs:134 - Api::generation checks if let Some(generation) = body.as_u64() before falling through to the object-shaped parse below. Looking at ColdStart::on_fetch (around line 35), the actor always responds with json!({"generation": ..., "startup": ...}), an object, never a bare number, so this branch looks unreachable/dead. If it is a leftover from an earlier response format, worth deleting so it does not get mistaken for a live code path.
• main.rs:297 - args.get(1).is_none_or(|s| s != "warm") re-derives the cold bool from raw args inside the "cold-start" | "bench" | "warm" arm, duplicating the branch selection the outer match already encodes. Deriving cold directly per match arm, rather than re-parsing args a second time with slightly different logic, would remove the duplication.

Bugs / correctness

• measure's percentile closure (times[(times.len() as f64 * p).ceil() as usize - 1]) would underflow if ever called with times empty or p <= 0, but today it is only invoked with the hardcoded 0.50/0.95/0.99 against a samples-sized vec that is already ensure!'d to be greater than 0, so this is not currently reachable. Worth keeping in mind if the percentiles become configurable later.
• Cleanup ordering in the "cold-start"|"bench"|"warm" arm (result?; cleanup?;) runs the DELETE regardless of whether measure failed (good, avoids leaking the actor), though if both measure and the cleanup request fail, only the measure error is surfaced. Fine for a benchmarking CLI.

Performance / security

No concerns. This only runs as an explicit local opt-in binary against a dev Engine/namespace, driven by env vars (RIVET_ENDPOINT, RIVET_TOKEN, etc.), and is not wired into any production code path.

Test coverage

No automated tests, but that matches this repo's convention for benchmark harnesses (manual, opt-in, requires a running local Engine) rather than something exercised in CI.

Nice README with a clear stage-by-stage comparison table for the rollout flags, which makes the benchmark easy to reproduce.