The async iterator API is quite ergonomic, and allows us to send paths
in batches. Sending paths in batches means that we can start processing
the paths before we've finished searching, and also makes passing
messages between rust and nodejs much faster.
I tried quite a few variants, like using a Buffer and manually
reconstructing strings, callbacks, etc.
The fastest way to pass data between rust and nodejs is by encoding /
decoding JSON. It's pretty unintuitive, but this is mainly because rust
and C++ (native JSON.parse) are way faster than constructing strings
manually with the v8 engine.
Both callbacks and iterators are quite fast, and have their own
trade-offs. The API for iterators is much more ergonomic however, so
that's what I pursued.
The performance improvement compared to the current implementation is
huge. Up to 2-3x in some cases.
❯ hyperfine 'node example/old.ts' 'node example/stream.ts' 'node example/main.ts'
Benchmark 1: node example/old.ts
Time (mean ± σ): 2.007 s ± 0.478 s [User: 2.572 s, System: 2.357 s]
Range (min … max): 1.619 s … 3.326 s 10 runs
Benchmark 2: node example/stream.ts
Time (mean ± σ): 1.234 s ± 0.258 s [User: 2.976 s, System: 2.811 s]
Range (min … max): 0.913 s … 1.735 s 10 runs
Benchmark 3: node example/main.ts
Time (mean ± σ): 1.117 s ± 0.267 s [User: 2.425 s, System: 2.484 s]
Range (min … max): 0.809 s … 1.680 s 10 runs
Summary
node example/main.ts ran
1.10 ± 0.35 times faster than node example/stream.ts
1.80 ± 0.61 times faster than node example/old.ts
Mert
Jonathan Jogenfors
Thomas