JSON Deserialization Speed Test

This repo compares how fast Python, C++, and Rust can parse JSON using various libraries. Same dataset, same schema, different languages—let’s see who wins!

What’s Tested?

We’re parsing a simple sensor data schema (details in the code) from an NDJSON file (messages.ndjson) with 200,000 entries by default.

Libraries

• Python: json, orjson, msgspec, rapidjson
• C++: nlohmann/json, RapidJSON
• Rust: serde_json, simd-json

Only deserialization (JSON to objects) is measured.

Repo Structure

xlang-json-bench/
├─ data/generate.py  # Makes the dataset
├─ python/bench.py   # Python tests
├─ cpp/bench.cpp     # C++ tests
├─ rust/             # Rust tests (Cargo.toml, src/main.rs)
├─ Dockerfile.*      # One for each language

Dataset

Generated by data/generate.py. Tweak size with N env var:

python data/generate.py --n 500000 --out messages.ndjson

Running It

Each language uses its own Docker setup:

# Python
docker build -f Dockerfile.python -t bench-python .
docker run --rm -e N=400000 bench-python

# C++
docker build -f Dockerfile.cpp -t bench-cpp .
docker run --rm -e N=400000 bench-cpp

# Rust
docker build -f Dockerfile.rust -t bench-rust .
docker run --rm -e N=400000 bench-rust

Results

Throughput (items/second) for 400,000 items:

cpp nlohmann/json parse    712,808.0 items/s
cpp RapidJSON parse       3,207,240.0 items/s
py json.loads             704,572.4 items/s
py orjson.loads         2,865,668.9 items/s
py msgspec.json decode  2,835,611.0 items/s
py rapidjson.loads       932,406.5 items/s
rs serde_json parse     2,355,698.7 items/s
rs simd-json parse      2,522,190.0 items/s

Notes

• Dataset auto-generates on first run.
• Use -e N=1000000 to change dataset size.
• I/O is excluded for fair timing.

License

MIT License