#fast #json #python


A hyper-fast Python module for reading/writing JSON data

4 releases

0.2.4 Jan 31, 2020
0.2.3 Jan 29, 2020
0.2.2 Sep 21, 2019
0.2.1 Sep 19, 2019

#209 in Encoding


893 lines


Build Status

A hyper-fast, safe Python module to read and write JSON data. Works as a drop-in replacement for Python's built-in json module. This is alpha software and there will be bugs, so maybe don't deploy to production just yet. 😉


pip install hyperjson


hyperjson is meant as a drop-in replacement for Python's json module:

>>> import hyperjson
>>> hyperjson.dumps([{"key": "value"}, 81, True])
>>> hyperjson.loads("""[{"key": "value"}, 81, true]""")
[{u'key': u'value'}, 81, True]


Parsing JSON is a solved problem; so, no need to reinvent the wheel, right?
Well, unless you care about performance and safety.

Turns out, parsing JSON correctly is a hard problem. Thanks to Rust however, we can minimize the risk of running into stack overflows or segmentation faults however.

hyperjson is a thin wrapper around Rust's serde-json and pyo3. It is compatible with Python 3 (and 2 on a best-effort basis).

For a more in-depth discussion, watch the talk about this project recorded at the Rust Cologne Meetup in August 2018.


  • Compatibility: Support the full feature-set of Python's json module.
  • Safety: No segfaults, panics, or overflows.
  • Performance: Significantly faster than json and as fast as ujson (both written in C).


  • Support ujson and simplejson extensions:
    Custom extensions like encode(), __json__(), or toDict() are not supported. The reason is, that they go against PEP8 (e.g. dunder methods are restricted to the standard library, camelCase is not Pythonic) and are not available in Python's json module.
  • Whitespace preservation: Whitespace in JSON strings is not preserved. Mainly because JSON is a whitespace-agnostic format and serde-json strips them out by default. In practice this should not be a problem, since your application must not depend on whitespace padding, but it's something to be aware of.


We are not fast yet. That said, we haven't done any big optimizations. In the long-term we might explore features of newer CPUs like multi-core and SIMD. That's one area other (C-based) JSON extensions haven't touched yet, because it might make code harder to debug and prone to race-conditions. In Rust, this is feasible due to crates like faster or rayon.

So there's a chance that the following measurements might improve soon.
If you want to help, check the instructions in the Development Environment section below.

Test machine:
MacBook Pro 15 inch, Mid 2015 (2,2 GHz Intel Core i7, 16 GB RAM) Darwin 17.6.18

Serialization benchmarks Deserialization benchmarks

Contributions welcome!

If you would like to hack on hyperjson, here's what needs to be done:

Just pick one of the open tickets. We can provide mentorship if you like. 😃

Developer guide

This project uses pipenv for managing the development environment. If you don't have it installed, run

pip install poetry

The project requires the nightly version of Rust.

Install it via rustup:

rustup install nightly

If you have already installed the nightly version, make sure it is up-to-date:

rustup update nightly

After that, you can compile the current version of hyperjson and execute all tests and benchmarks with the following commands:

make install
make test
make bench

🤫 Pssst!... run make help to learn more.

Drawing pretty diagrams

In order to recreate the benchmark histograms, you first need a few additional prerequisites:

On macOS, please also add the following to your ~/.matplotlib/matplotlibrc (reference):

backend: TkAgg

After that, run the following:

make plot


hyperjson is licensed under either of

at your option.


Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in hyperjson by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.