jsonslicer - stream JSON parser

Overview

JsonSlicer performs a stream or iterative, pull JSON parsing, which means it does not load whole JSON into memory and is able to parse very large JSON files or streams. The module is written in C and uses YAJL JSON parsing library, so it's also quite fast.

JsonSlicer takes a path of JSON map keys or array indexes, and provides iterator interface which yields JSON data matching given path as complete Python objects.

Example

{
    "friends": [
        {"name": "John", "age": 31},
        {"name": "Ivan", "age": 26}
    ],
    "colleagues": {
        "manager": {"name": "Jack", "age": 33},
        "subordinate": {"name": "Lucy", "age": 21}
    }
}

from jsonslicer import JsonSlicer

# Extract specific elements:
with open('people.json') as data:
    ivans_age = next(JsonSlicer(data, ('friends', 1, 'age')))
    # 26

with open('people.json') as data:
    managers_name = next(JsonSlicer(data, ('colleagues', 'manager', 'name')))
    # 'Jack'

# Iterate over collection(s) by using wildcards in the path:
with open('people.json') as data:
    for person in JsonSlicer(data, ('friends', None)):
        print(person)
        # {'name': 'John', 'age': 31}
        # {'name': 'Ivan', 'age': 26}

# Iteration over both arrays and dicts is possible, even at the same time
with open('people.json') as data:
    for person in JsonSlicer(data, (None, None)):
        print(person)
        # {'name': 'John', 'age': 31}
        # {'name': 'Ivan', 'age': 26}
        # {'name': 'Jack', 'age': 33}
        # {'name': 'Lucy', 'age': 21}

# Map key of returned objects is available on demand...
with open('people.json') as data:
    for position, person in JsonSlicer(data, ('colleagues', None), path_mode='map_keys'):
        print(position, person)
        # 'manager' {'name': 'Jack', 'age': 33}
        # 'subordinate' {'name': 'Lucy', 'age': 21}

# ...as well as complete path information
with open('people.json') as data:
    for *path, person in JsonSlicer(data, (None, None), path_mode='full'):
        print(path, person)
        # ('friends', 0) {'name': 'John', 'age': 31})
        # ('friends', 1) {'name': 'Ivan', 'age': 26})
        # ('colleagues', 'manager') {'name': 'Jack', 'age': 33})
        # ('colleagues', 'subordinate') {'name': 'Lucy', 'age': 21})

# Extract all instances of deep nested field
with open('people.json') as data:
    age_sum = sum(JsonSlicer(data, (None, None, 'age')))
    # 111

API

jsonslicer.JsonSlicer(
    file,
    path_prefix,
    read_size=1024,
    path_mode=None,
    yajl_allow_comments=False,
    yajl_dont_validate_strings=False,
    yajl_allow_trailing_garbage=False,
    yajl_allow_multiple_values=False,
    yajl_allow_partial_values=False,
    yajl_verbose_errors=True,
    encoding=None,
    errors=None,
    binary=False,
)

Constructs iterative JSON parser. which reads JSON data from file (a .read()-supporting file-like object containing a JSON document).

file is a .read()-supporting file-like object containing a JSON document. Both binary and text files are supported, but binary ones are preferred, because the parser has to operate on binary data internally anyway, and using text input would require an unnecessary encoding/decoding which yields ~3% performance overhead. Note that JsonSlicer supports both unicode and binary output regardless of input format.

path_prefix is an iterable (usually a list or a tuple) specifying a path or a path pattern of objects which the parser should extract from JSON.

For instance, in the example above a path ('friends', 0, 'name') will yield string 'John', by descending from the root element into the dictionary element by key 'friends', then into the array element by index 0, then into the dictionary element by key 'name'. Note that integers only match array indexes and strings only match dictionary keys.

The path can be turned into a pattern by specifying None as a placeholder in some path positions. For instance, (None, None, 'name') will yield all four names from the example above, because it matches an item under 'name' key on the second nesting level of any arrays or map structure.

Both strings and byte objects are allowed in path, regardless of input and output encodings. are automatically converted to the format used internally.

read_size is a size of block read by the parser at a time.

path_mode is a string which specifies how a parser should return path information along with objects. The following modes are supported:

'ignore' (the default) - do not output any path information, just objects as is ('friends').

{'name': 'John', 'age': 31}
{'name': 'Ivan', 'age': 26}
{'name': 'Jack', 'age': 33}
{'name': 'Lucy', 'age': 21}

Common usage pattern for this mode is

for object in JsonSlicer(...)

'map_keys' - output objects as is when traversing arrays and tuples consisting of map key and object when traversing maps.
```
{'name': 'John', 'age': 31}
{'name': 'Ivan', 'age': 26}
('manager', {'name': 'Jack', 'age': 33})
('subordinate', {'name': 'Lucy', 'age': 21})
```
This format may seem inconsistent (and therefore it's not the default), however in practice only collection of a single type is iterated at a time and this type is known, so this format is likely the most useful as in most cases you do need dictionary keys.

Common usage pattern for this mode is
```
for object in JsonSlicer(...)  # when iterating arrays
for key object in JsonSlicer(...)  # when iterating maps
```

'full_paths' - output tuples consisting of all path components (both map keys and array indexes) and an object as the last element.

('friends', 0, {'name': 'John', 'age': 31})
('friends', 1, {'name': 'Ivan', 'age': 26})
('colleagues', 'manager', {'name': 'Jack', 'age': 33})
('colleagues', 'subordinate', {'name': 'Lucy', 'age': 21})

Common usage pattern for this mode is

for *path, object in JsonSlicer(...)

yajl_allow_comments enables corresponding YAJL flag, which is documented as follows:

Ignore javascript style comments present in JSON input. Non-standard, but rather fun

yajl_dont_validate_strings enables corresponding YAJL flag, which is documented as follows:

When set the parser will verify that all strings in JSON input are valid UTF8 and will emit a parse error if this is not so. When set, this option makes parsing slightly more expensive (~7% depending on processor and compiler in use)

yajl_allow_trailing_garbage enables corresponding YAJL flag, which is documented as follows:

By default, yajl will ensure the entire input text was consumed and will raise an error otherwise. Enabling this flag will cause yajl to disable this check. This can be useful when parsing json out of a that contains more than a single JSON document.

yajl_allow_multiple_values enables corresponding YAJL flag, which is documented as follows:

Allow multiple values to be parsed by a single handle. The entire text must be valid JSON, and values can be seperated by any kind of whitespace. This flag will change the behavior of the parser, and cause it continue parsing after a value is parsed, rather than transitioning into a complete state. This option can be useful when parsing multiple values from an input stream.

yajl_allow_partial_values enables corresponding YAJL flag, which is documented as follows:

When yajl_complete_parse() is called the parser will check that the top level value was completely consumed. I.E., if called whilst in the middle of parsing a value yajl will enter an error state (premature EOF). Setting this flag suppresses that check and the corresponding error.

yajl_verbose_errors enables verbose YAJL errors, with exception message including the JSON text where the error occured, along with an arrow pointing to the specific char.

encoding may be used to override output encoding, which is derived from the input file handle if possible, or otherwise set to the default one as Python builtin open() would use (usually 'UTF-8').

errors is an optional string that specifies how encoding and decoding errors are to be handled. Defaults to 'strict'

binary forces the output to be in form of bytes objects instead of str unicode strings.

The constructed object is as iterator. You may call next() to extract single element from it, iterate it via for loop, or use it in generator comprehensions or in any place where iterator is accepted.

Performance/competitors

The closest competitor is ijson, and JsonSlicer was written to be better. Namely,

It's up to 35x faster depending on ijson backend (starting with 3.0, ijson supports comparable performance via yajl2_c backend), close in performance to Python's native json module.
It supports more flexible paths/patterns specifying which objects to iterate over in JSON hierarchy and provides consistent interface for iteration over arrays and dictionaries

The results of bundled benchmark on Python 3.8.2 / clang 8.0.1 / -O2 -DNDEBUG / FreeBSD 12.1 amd64 / Core i7-6600U CPU @ 2.60GHz.

Facility	Type	Objects/sec
json.loads()	str	1147.6K
json.load(StringIO())	str	1139.3K
JsonSlicer (no paths, binary input, binary output)	bytes	1149.7K
JsonSlicer (no paths, unicode input, binary output)	bytes	1134.5K
JsonSlicer (no paths, binary input, unicode output)	str	1012.3K
JsonSlicer (no paths, unicode input, unicode output)	str	996.2K
JsonSlicer (full paths, binary output)	bytes	763.1K
JsonSlicer (full paths, unicode output)	str	567.2K
ijson.yajl2_c	bytes	1062.0K
ijson.yajl2_cffi	bytes	71.6K
ijson.yajl2	bytes	56.4K
ijson.python	str	32.0K

Status/TODO

JsonSlicer is currently in beta stage, used in production in Repology project. Testing foci are:

Edge cases with uncommon encoding (input/output) configurations
Absence of memory leaks

Requirements

Python 3.6+
yajl 2.0.3+ (older versions lack pkgconfig file)
pkg-config (build-time)
C++ compiler (build-time)

AMDmi3/jsonslicer