pysimdjson

Quick-n'dirty Python bindings for simdjson just to see if going down this path might yield some parse time improvements in real-world applications. So far, the results are promising, especially when only part of a document is of interest.

Bindings are currently tested on OS X, Linux, and Windows.

See the latest documentation at http://pysimdjson.tkte.ch.

Installation

There are binary wheels available for py3.6/py3.7 on OS X 10.12 & Windows. On other platforms you'll need a C++17-capable compiler.

pip install pysimdjson

If you're getting errors when installing from pip, there's probably no binary package available for your combination of platform & python version. As long as you have a C++17 compiler installed you can still use pip, you just need to provide a few extra compiler flags. The most common are:

• gcc/clang: CFLAGS="-march=native -std=c++17" pip install pysimdjson

• msvc (Visual Studio 2017):

  SET CL="/std:c++17 /arch:AVX2"
  pip install pysimdjson
  

or from git:

git clone https://github.com/TkTech/pysimdjson.git
cd pysimdjson
python setup.py install

Example

import simdjson

with open('sample.json', 'rb') as fin:
    doc = simdjson.loads(fin.read())

However, this doesn't really gain you that much over, say, ujson. You're still loading the entire document and converting the entire thing into a series of Python objects which is very expensive. You can instead use items() to pull only part of a document into Python.

Example document:

{
    "type": "search_results",
    "count": 2,
    "results": [
        {"username": "bob"},
        {"username": "tod"}
    ],
    "error": {
        "message": "All good captain"
    }
}

And now lets try some queries...

import simdjson

with open('sample.json', 'rb') as fin:
    # Calling ParsedJson with a document is a shortcut for
    # calling pj.allocate_capacity(<size>) and pj.parse(<doc>). If you're
    # parsing many JSON documents of similar sizes, you can allocate
    # a large buffer just once and keep re-using it instead.
    pj = simdjson.ParsedJson(fin.read())

    pj.items('.type') #> "search_results"
    pj.items('.count') #> 2
    pj.items('.results[].username') #> ["bob", "tod"]
    pj.items('.error.message') #> "All good captain"

AVX2

simdjson requires AVX2 support to function. Check to see if your OS/processor supports it:

• OS X: sysctl -a | grep machdep.cpu.leaf7_features
• Linux: grep avx2 /proc/cpuinfo

Low-level interface

You can use the low-level simdjson Iterator interface directly, just be aware that this interface can change any time. If you depend on it you should pin to a specific version of simdjson. You may need to use this interface if you're dealing with odd JSON, such as a document with repeated non-unique keys.

with open('sample.json', 'rb') as fin:
    pj = simdjson.ParsedJson(fin.read())
    iter = simdjson.Iterator(pj)
    if iter.is_object():
        if iter.down():
            print(iter.get_string())

Early Benchmark

Comparing the built-in json module loads on py3.7 to simdjson loads.

Getting subsets of the document is significantly faster. For canada.json getting .type using the naive approach and the items() appraoch, average over N=100.

This approach avoids creating Python objects for fields that aren't of interest. When you only care about a small part of the document, it will always be faster.