Benchmarks

Headline numbers

Measured with Criterion against crates/storage's cargo bench -p cairn-storage:

• ~552µs per 1,000 sequential puts — roughly 1.8M puts/sec on a single node, single thread, values held entirely in memory before a flush.
• ~6.1µs cold point-read latency — a get against a key that's been flushed and compacted to disk (i.e., not sitting in the memtable), so the measurement includes a bloom-filter check and an SSTable lookup, not just a BTreeMap hit.

What these numbers are, precisely

• Microbenchmarks, not a production SLA. They measure the storage engine in isolation — no network, no Raft replication, no concurrent client load.
• Single-node, single-thread. There is no concurrency story yet; the engine's public API takes &mut self for writes.
• Criterion's statistical methodology (warm-up, multiple samples, outlier detection) — these are stable estimates on the benchmark machine, not a guarantee that will hold on different hardware.

They're published because they're the honest baseline the roadmap's later phases (Raft replication, MVCC, sharding) will each be measured against — the point of a "before/after" number is having a real "before."

Test coverage by module

Correctness, not speed, is what actually gates each phase. 32 tests pass across the engine's modules, plus a property-based test that runs hundreds of randomized Put/Delete/Flush/Compact sequences against a BTreeMap reference model and asserts every Get agrees:

Reproducing these numbers

git clone https://github.com/uptonm/cairn
cd cairn
cargo test -p cairn-storage    # 32 tests
cargo bench -p cairn-storage   # criterion benchmarks