genson-rs

-- 🔥 Generate JSON Schema from Gigabytes of JSON data in seconds

genson-rs is a Rust rewrite of the GenSON Python library , which can be used to generate JSON schema (Draft-04 and after) from one or multiple JSON objects.

While not having full feature parity yet, genson-rs focuses on speed ⚡️. It offers MUCH better performance (25x ~ 75x faster) compared to the Python GenSON library, and is generally a lot faster than other open source schema inference tools as well. Its high performance makes it a viable choice for online schema inference for large JSON dataset at scale. Check out the benchmark section for performance benchmark comparisons.

Install

Installation via Cargo is the easiest. If you don't have it already, follow the link to set up Cargo (one simple command), then run:

cargo install genson-rs

Installing via brew will be supported soon.

Usage

genson-rs <OPTION> <FILE>

e.g. If you have a large JSON file full of request logs in JSON format

genson-rs request_logs.json

Additionally, if each request log is a JSON object in its own line, you can specify the delimiter which will slightly improve the performance

genson-rs --delimiter newline request_logs.json 

Benchmark

The following benchmarks are executed manually on my local 2023 Macbook Pro with the M2 Pro Chip (10 cores, 4 high-efficiency + 6 high-performance), 16GB RAM, running macOS 13.0. Each of the test JSON files is generated using the json_gen.py script inside of the tests/data folder, and each test was executed 3 times. The median was used out of the 3 runs.

As you can see, genson-rs is extremely fast, and might be the fastest schema inference engine out there based on my rudimentary benchmarks against other tools (that I'm aware of) as well.

Optimization Techniques

The genson-rs library leverages the following techniques to greatly speed up the schema generation process:

• ⚡️ Rust being blazingly fast itself -- without any GC or interpreter overhead, a 1-to-1 port in Rust running on a single CPU core runs 2x faster than the Python version already
• ⚡️ Parallel processing leveraging all available CPU cores -- whie Python has the limitation of the GIL that prevents it from leveraging multiple CPU cores efficiently, genson-rs parallelizes Map-Reduce type of workload whenever possible (e.g. when processing gigantic arrays), maxing out all the available CPU cores
• ⚡️ Extremely fast JSON parsing powered by SIMD instructions -- instead of fully parsing out the whole JSON dataset, we use the simd-json library (a Rust port of the C++ simdjson library) that leverages SIMD (Single Instruction/Multiple Data) instructions to only parse out the "tape" of the JSON dataset, which is sufficient enough to build the schema on top of without fully deserializing the whole dataset
• ⚡️ Efficient memory management using the MiMalloc allocator -- this is recommended by the simd-json library itself, genson-rs opts to use the MiMalloc allocator instead of the default global allocator which made the code run faster by a decent amount