Skip to content

performance.md

Latest commit

 

History

History
80 lines (61 loc) · 3.35 KB

performance.md

File metadata and controls

80 lines (61 loc) · 3.35 KB
Raw

Performance benchmarks

missive uses Criterion benchmarks as local, non-blocking performance probes. They are intended to make regressions inspectable while the project is still adding features; the initial budgets below are soft guidance, not quality-gate failures.

Running benchmarks

Compile every benchmark target without measuring:

cargo bench --workspace --all-features --no-run

Run the current benchmark suite:

cargo bench -p missive-cli --bench performance --all-features

Run the same benchmark target in Criterion's fast smoke-test mode:

cargo bench -p missive-cli --bench performance --all-features -- --test

Criterion stores local reports and baselines under target/criterion/, which is build output and must not be committed. To inspect a local regression, save a baseline before the change and compare after the change:

cargo bench -p missive-cli --bench performance --all-features -- --save-baseline before
# make the candidate change
cargo bench -p missive-cli --bench performance --all-features -- --baseline before

Aggressive quality-gate mode compiles benchmarks with cargo bench --no-run when benchmark sources exist, but it does not run full measurements by default.

Current benchmark coverage

The performance Criterion target currently lives at crates/missive-cli/benches/performance.rs and covers:

  • configuration loading and redacted rendering for a multi-agent TOML config;
  • SQLite repository CRUD batches for agents, contexts, tasks, messages, artifacts, and events;
  • router planning for capability-match, weighted, broadcast, and quorum policies;
  • event-journal replay summary reconstruction;
  • A2A streaming event JSON parsing, including JSON-RPC unwrapping;
  • local group collective command paths for gather and deterministic summarise reduce over seeded profile state.

All benchmark fixtures use temporary in-memory stores or temporary MISSIVE_HOME directories and local loopback URLs only. They do not contact third-party agents.

Initial soft budgets

These budgets are deliberately loose so early development is not blocked by machine noise, debug packages, or still-evolving data models. Treat them as investigation thresholds on a typical developer laptop after a release build has already warmed dependencies.

Area Benchmark signal Soft investigation budget
Config load parse and validate a 32-agent TOML config Criterion upper estimate under 25 ms
Config redaction render the same config as redacted JSON Criterion upper estimate under 25 ms
Store operations one CRUD batch for 32 task/message/artifact/event rows Criterion upper estimate under 75 ms
Routing 128-candidate policy planning Criterion upper estimate under 20 ms per policy
Event replay replay 512 event records into summaries Criterion upper estimate under 50 ms
Stream parsing parse 256 A2A stream events Criterion upper estimate under 50 ms
Local collectives gather or local summarise reduce over 16 members Criterion upper estimate under 200 ms

When a result exceeds a soft budget, first rerun with a saved Criterion baseline and check whether the regression is repeatable. Only then decide whether to optimize, adjust the budget with a documented reason, or add a future automated performance gate.