ijson 3.5.0

Iterative JSON parser with standard Python iterator interfaces

ijson

Ijson is an iterative JSON parser with standard Python iterator interfaces.

• Installation

• Usage

  • High-level interfaces

  • Lower-level interfaces

  • Command line

  • Benchmarking

  • bytes/str support

  • Iterator support

  • asyncio support

  • Intercepting events

  • Push interfaces

• Options

• Events

• Prefix

• Backends

  • Capabilities

• Performance tips

• FAQ

• Acknowledgements

Installation

Ijson is hosted in PyPI, so you should be able to install it via pip:

pip install ijson

Binary wheels are provided for major platforms and python versions. These are built and published automatically using cibuildwheel via GitHub Actions.

Usage

All usage example will be using a JSON document describing geographical objects:

{
  "earth": {
    "europe": [
      {"name": "Paris", "type": "city", "info": { ... }},
      {"name": "Thames", "type": "river", "info": { ... }},
      // ...
    ],
    "america": [
      {"name": "Texas", "type": "state", "info": { ... }},
      // ...
    ]
  }
}

High-level interfaces

ijson works by continuously reading data from a JSON stream provided by the user. This is presented as a file-like object. In particular it must provide a read(size) method returning either bytes (preferably) or str. Example file-like objects are files opened with open, HTTP/HTTPS requests made using urllib.request.urlopen, socket.socket objects, and more.

The most common usage of ijson is to yield native Python objects located under a prefix. This is done using the items function. Here’s how to process all European cities:

import ijson

f = urlopen('http://.../')
objects = ijson.items(f, 'earth.europe.item')
cities = (o for o in objects if o['type'] == 'city')
for city in cities:
    do_something_with(city)

For how to build a prefix see the prefix section below.

Other times it might be useful to iterate over object members rather than objects themselves (e.g., when objects are too big). In that case one can use the kvitems function instead:

import ijson

f = urlopen('http://.../')
european_places = ijson.kvitems(f, 'earth.europe.item')
names = (v for k, v in european_places if k == 'name')
for name in names:
    do_something_with(name)

Lower-level interfaces

Sometimes when dealing with a particularly large JSON payload it may worth to not even construct individual Python objects and react on individual events immediately producing some result. This is achieved using the parse function:

import ijson

parser = ijson.parse(urlopen('http://.../'))
stream.write('<geo>')
for prefix, event, value in parser:
    if (prefix, event) == ('earth', 'map_key'):
        stream.write('<%s>' % value)
        continent = value
    elif prefix.endswith('.name'):
        stream.write('<object name="%s"/>' % value)
    elif (prefix, event) == ('earth.%s' % continent, 'end_map'):
        stream.write('</%s>' % continent)
stream.write('</geo>')

Even more bare-bones is the ability to react on individual events without even calculating a prefix using the basic_parse function:

import ijson

events = ijson.basic_parse(urlopen('http://.../'))
num_names = sum(1 for event, value in events
                if event == 'map_key' and value == 'name')

Command line

A command line utility is included with ijson to help visualise the output of each of the routines above. It reads JSON from the standard input, and it prints the results of the parsing method chosen by the user to the standard output.

The tool is available by running the ijson.dump module. For example:

$> echo '{"A": 0, "B": [1, 2, 3, 4]}' | python -m ijson.dump -m parse
#: path, name, value
--------------------
0: , start_map, None
1: , map_key, A
2: A, number, 0
3: , map_key, B
4: B, start_array, None
5: B.item, number, 1
6: B.item, number, 2
7: B.item, number, 3
8: B.item, number, 4
9: B, end_array, None
10: , end_map, None

Using -h/--help will show all available options.

Benchmarking

A command line utility is included with ijson to help benchmarking the different methods offered by the package. It offers some built-in example inputs that try to mimic different scenarios, but more importantly it also supports user-provided inputs. You can also specify which backends to time, number of iterations, and more.

The tool is available by running the ijson.benchmark module. For example:

$> python -m ijson.benchmark my/json/file.json -m items -p values.item

Using -h/--help will show all available options.

bytes/str support

Although not usually how they are meant to be run, all the functions above also accept bytes and str objects directly as inputs. These are then internally wrapped into a file object, and further processed. This is useful for testing and prototyping, but probably not extremely useful in real-life scenarios.

Iterator support

In many situations the direct input users want to pass to ijson is an iterator (e.g., a generator) rather than a file-like object. ijson provides a built-in adapter to bridge this gap:

• ijson.from_iter(iterable_or_async_iterable_of_bytes)

asyncio support

All of the methods above work also on file-like asynchronous objects, so they can be iterated asynchronously. In other words, something like this:

import asyncio
import ijson

async def run():
   f = await async_urlopen('http://..../')
   async for object in ijson.items(f, 'earth.europe.item'):
      if object['type'] == 'city':
         do_something_with(city)
asyncio.run(run())

An explicit set of *_async functions also exists offering the same functionality, except they will fail if anything other than a file-like asynchronous object is given to them. (so the example above can also be written using ijson.items_async). In fact in ijson version 3.0 this was the only way to access the asyncio support.

Intercepting events

The four routines shown above internally chain against each other: tuples generated by basic_parse are the input for parse, whose results are the input to kvitems and items.

Normally users don’t see this interaction, as they only care about the final output of the function they invoked, but there are occasions when tapping into this invocation chain this could be handy. This is supported by passing the output of one function (i.e., an iterable of events, usually a generator) as the input of another, opening the door for user event filtering or injection.

For instance if one wants to skip some content before full item parsing:

import io
import ijson

parse_events = ijson.parse(io.BytesIO(b'["skip", {"a": 1}, {"b": 2}, {"c": 3}]'))
while True:
    prefix, event, value = next(parse_events)
    if value == "skip":
        break
for obj in ijson.items(parse_events, 'item'):
    print(obj)

Note that this interception only makes sense for the basic_parse -> parse, parse -> items and parse -> kvitems interactions.

Note also that event interception is currently not supported by the async functions.

Push interfaces

All examples above use a file-like object as the data input (both the normal case, and for asyncio support), and hence are “pull” interfaces, with the library reading data as necessary. If for whatever reason it’s not possible to use such method, you can still push data through yet a different interface: coroutines (via generators, not asyncio coroutines). Coroutines effectively allow users to send data to them at any point in time, with a final target coroutine-like object receiving the results.

In the following example the user is doing the reading instead of letting the library do it:

import ijson

@ijson.coroutine
def print_cities():
   while True:
      obj = (yield)
      if obj['type'] != 'city':
         continue
      print(obj)

coro = ijson.items_coro(print_cities(), 'earth.europe.item')
f = urlopen('http://.../')
for chunk in iter(functools.partial(f.read, buf_size)):
   coro.send(chunk)
coro.close()

All four ijson iterators have a *_coro counterpart that work by pushing data into them. Instead of receiving a file-like object and option buffer size as arguments, they receive a single target argument, which should be a coroutine-like object (anything implementing a send method) through which results will be published.

An alternative to providing a coroutine is to use ijson.sendable_list to accumulate results, providing the list is cleared after each parsing iteration, like this:

import ijson

events = ijson.sendable_list()
coro = ijson.items_coro(events, 'earth.europe.item')
f = urlopen('http://.../')
for chunk in iter(functools.partial(f.read, buf_size)):
   coro.send(chunk)
   process_accumulated_events(events)
   del events[:]
coro.close()
process_accumulated_events(events)

Options

Additional options are supported by all ijson functions to give users more fine-grained control over certain operations:

• The use_float option (defaults to False) controls how non-integer values are returned to the user. If set to True users receive float() values; otherwise Decimal values are constructed. Note that building float values is usually faster, but on the other hand there might be loss of precision (which most applications will not care about) and will raise an exception when overflow occurs (e.g., if 1e400 is encountered). This option also has the side-effect that integer numbers bigger than 2^64 (but sometimes 2^32, see capabilities) will also raise an overflow error, due to similar reasons. Future versions of ijson might change the default value of this option to True.

• The multiple_values option (defaults to False) controls whether multiple top-level values are supported. JSON content should contain a single top-level value (see the JSON Grammar). However there are plenty of JSON files out in the wild that contain multiple top-level values, often separated by newlines. By default ijson will fail to process these with a parse error: trailing garbage error unless multiple_values=True is specified.

• Similarly the allow_comments option (defaults to False) controls whether C-style comments (e.g., /* a comment */), which are not supported by the JSON standard, are allowed in the content or not.

• For functions taking a file-like object, an additional buf_size option (defaults to 65536 or 64KB) specifies the amount of bytes the library should attempt to read each time.

• The items and kvitems functions, and all their variants, have an optional map_type argument (defaults to dict) used to construct objects from the JSON stream. This should be a dict-like type supporting item assignment.

Events

When using the lower-level ijson.parse function, three-element tuples are generated containing a prefix, an event name, and a value. Events will be one of the following:

• start_map and end_map indicate the beginning and end of a JSON object, respectively. They carry a None as their value.

• start_array and end_array indicate the beginning and end of a JSON array, respectively. They also carry a None as their value.

• map_key indicates the name of a field in a JSON object. Its associated value is the name itself.

• null, boolean, integer, double, number and string all indicate actual content, which is stored in the associated value.

Prefix

A prefix represents the context within a JSON document where an event originates at. It works as follows:

• It starts as an empty string.

• A <name> part is appended when the parser starts parsing the contents of a JSON object member called name, and removed once the content finishes.

• A literal item part is appended when the parser is parsing elements of a JSON array, and removed when the array ends.

• Parts are separated by ..

When using the ijson.items function, the prefix works as the selection for which objects should be automatically built and returned by ijson.

Backends

Ijson provides several implementations of the actual parsing in the form of backends located in ijson/backends:

• yajl2_c: a C extension using YAJL 2.x. This is the fastest, but might require a compiler and the YAJL development files to be present when installing this package. Binary wheel distributions exist for major platforms/architectures to spare users from having to compile the package.

• yajl2_cffi: wrapper around YAJL 2.x using CFFI.

• yajl2: wrapper around YAJL 2.x using ctypes, for when you can’t use CFFI for some reason.

• yajl: deprecated YAJL 1.x + ctypes wrapper, for even older systems.

• python: pure Python parser, good to use with PyPy

This list of backend names is available under the ijson.ALL_BACKENDS constant.

You can import a specific backend and use it in the same way as the top level library:

import ijson.backends.yajl2_cffi as ijson

for item in ijson.items(...):
    # ...

Importing the top level library as import ijson uses the first available backend in the same order of the list above, and its name is recorded under ijson.backend. If the IJSON_BACKEND environment variable is set its value takes precedence and is used to select the default backend.

You can also use the ijson.get_backend function to get a specific backend based on a name:

backend = ijson.get_backend('yajl2_c')
for item in backend.items(...):
    # ...

Capabilities

Apart from their performance, all backends are designed to support the same capabilities. There are however some small known differences, all of which can be queried by inspecting the capabilities module constant. It contains the following members:

• c_comments: C-style comments are supported.

• multiple_values: multiple top-level JSON values are supported.

• detects_invalid_leading_zeros: numbers with leading zeroes are reported as invalid (as they should, as pert the JSON standard), raising a ValueError.

• detects_incomplete_json_tokens: detects incomplete JSON tokens at the end of an incomplete document (e.g., {"a": fals), raising an IncompleteJSONError.

• int64: when using use_float=True,

  values greater than or equal to 2^32 are correctly returned.

These capabilities are supported by all backends, with the following exceptions:

• The yajl backend doesn’t support multiple_values, detects_invalid_leading_zeros and detects_incomplete_json_tokens. It also doesn’t support int64 in platforms with a 32-bit C long type.

• The python backend doesn’t support c_comments.

Performance tips

In more-or-less decreasing order, these are the most common actions you can take to ensure you get most of the performance out of ijson:

• Make sure you use the fastest backend available. See backends for details.

• If you know your JSON data contains only numbers that are “well behaved” consider turning on the use_float option. See options for details.

• Make sure you feed ijson with binary data instead of text data. See faq #1 for details.

• Play with the buf_size option, as depending on your data source and your system a value different from the default might show better performance. See options for details.

The benchmarking tool should help with trying some of these options and observing their effect on your input files.

FAQ

1. Q: Does ijson work with bytes or str values?

   A: In short: both are accepted as input, outputs are only str.

   All ijson functions expecting a file-like object should ideally be given one that is opened in binary mode (i.e., its read function returns bytes objects, not str). However if a text-mode file object is given then the library will automatically encode the strings into UTF-8 bytes. A warning is currently issued (but not visible by default) alerting users about this automatic conversion.

   On the other hand ijson always returns text data (JSON string values, object member names, event names, etc) as str objects. This mimics the behavior of the system json module.

2. Q: How are numbers dealt with?

   A: ijson returns int values for integers and decimal.Decimal values for floating-point numbers. This is mostly because of historical reasons. Since 3.1 a new use_float option (defaults to False) is available to return float values instead. See the options section for details.

3. Q: I’m getting an UnicodeDecodeError, or an IncompleteJSONError with no message

   A: This error is caused by byte sequences that are not valid in UTF-8. In other words, the data given to ijson is not really UTF-8 encoded, or at least not properly.

   Depending on where the data comes from you have different options:

   • If you have control over the source of the data, fix it.

   • If you have a way to intercept the data flow, do so and pass it through a “byte corrector”. For instance, if you have a shell pipeline feeding data through stdin into your process you can add something like ... | iconv -f utf8 -t utf8 -c | ... in between to correct invalid byte sequences.

   • If you are working purely in python, you can create a UTF-8 decoder using codecs’ incrementaldecoder to leniently decode your bytes into strings, and feed those strings (using a file-like class) into ijson (see our string_reader_async internal class for some inspiration).

   In the future ijson might offer something out of the box to deal with invalid UTF-8 byte sequences.

4. Q: I’m getting parse error: trailing garbage or Additional data found errors

   A: This error signals that the input contains more data than the top-level JSON value it’s meant to contain. This is usually caused by JSON data sources containing multiple values, and is usually solved by passing the multiple_values=True to the ijson function in use. See the options section for details.

5. Q: How do I use ijson with requests or httpx

   A: The requests library downloads the body of the HTTP response immediately by default. To stream JSON into ijson, pass stream=True and adapt the byte iterator:

   import requests
   import ijson
   
   with requests.get('https://..', stream=True) as resp:
       resp.raise_for_status()
       f = ijson.from_iter(resp.iter_content(chunk_size=64*1024))
       objects = ijson.items(f, 'earth.europe.item')
       cities = (o for o in objects if o['type'] == 'city')
       for city in cities:
         do_something_with(city)

   You can also pass Response.raw directly (it’s a file-like object), but using iter_content is preferred because requests will transparently handle HTTP transfer encodings (e.g., gzip, chunked).

   For async usage with httpx:

   import httpx
   import ijson
   
   async with httpx.AsyncClient() as client:
       async with client.stream('GET', 'https://..') as resp:
           resp.raise_for_status()
           f = ijson.from_iter(resp.aiter_bytes())
           objects = ijson.items(f, 'earth.europe.item')
           cities = (o async for o in objects if o['type'] == 'city')
           async for city in cities:
             do_something_with(city)

Acknowledgements

ijson was originally developed and actively maintained until 2016 by Ivan Sagalaev. In 2019 he handed over the maintenance of the project and the PyPI ownership.

Python parser in ijson is relatively simple thanks to Douglas Crockford who invented a strict, easy to parse syntax.

The YAJL library by Lloyd Hilaiel is the most popular and efficient way to parse JSON in an iterative fashion. When building the library ourselves, we use our own fork that contains fixes for all known CVEs.

Ijson was inspired by yajl-py wrapper by Hatem Nassrat. Though ijson borrows almost nothing from the actual yajl-py code it was used as an example of integration with yajl using ctypes.