Splitting of (XML, JSON) objects from a continuous stream

pip install splitstream==1.2.2


splitstream - Stream document splitter

Travis CI status License PyPi version Bindings

###{"j":"s"}["o","n"]{"j":"s"}, ["o","n"] ###<xml1/><xml2/><xml1/>, <xml2/>

This is a C library with Python bindings that will help you split continuous streams of non-delimited XML documents or JSON and UBJSON objects/arrays. Instead of using regular expressions or other forms of string matching, the approach is to implement a basic tokenizer and element depth parser to detect where the document ends, without relying on the existence of specific elements, delimiters or length prefix.

The library is written in C and provides bindings for Python (supports Python 2.6+ and Python 3 as well as PyPy).

Developing for OS X or iOS? Take a look at IQSerialization, which provides Obj-C (and Swift) bindings for this library (and much more).

Use cases

  • Parsing XML/JSON log files as they are written. In this case, the log file is usually not wellformed since the end tag for the root element is missing, or each log entry is its own root. This library handles both cases.
  • Parsing huge XML files as objects. It may not be feasible to feed the entire file into a deserializer or DOM parser, and using an event based parser can add complexity to your code. By pre-processing with this library you can parse chunks of the stream instead.
  • Parsing objects from a raw TCP stream. Parse the objects as they appear on the raw stream instead of wrapping in HTTP requests or your own protocol.


  • Understands XML, JSON and UBJSON.
  • Tokenizer will correctly handle complex documents (e.g. xml within comments or CDATA, escape sequences, processing instructions, etc).
  • Comprehensive and growing test suite.
  • Written in clean C with no dependencies except the standard C library.
  • Possible to add your own tokenizers (in C) to add more object formats.
  • Python bindings (Python 2.x and 3.x, tested with PyPy).
  • Set initial parse depth to allow for e.g. parsing a log file with an "infinite" (unclosed) root element.
  • Quite good performance and low memory footprint (constant with stream size), even from Python.
  • Convenience API for working with FILE* pointers (C) and file objects (Python).


  • Will not actually parse the objects (this is very much a design decision) - you need to feed the output from splitstream into your parser of choice.
  • Only JSON arrays and objects are supported as root objects.
  • Limited error handling. May not recover from malformed objects in many cases.
  • Limited documentation (this page and the unit tests are a good start for now).
  • UBJSON support is relatively untested.
  • Python 2.x is not supported on the Microsoft Windows platform, as it is forced to use and old version of the C compiler that does not support the C99 syntax that this library uses.

The C interface

Tokenization context

Allocate a SplitstreamState on the stack or heap and initialize it:

void SplitstreamInit(SplitstreamState* state);
void SplitstreamInitDepth(SplitstreamState* state, int startDepth);

The SplitstreamInitDepth(...) function also allows specifying a start depth that helps parsing subtrees of "infinite" XML documents, such as

<?xml version="1.0" encoding="utf-8">
  <logEntry id="1000" type="info">Hello</logEntry>
  <logEntry id="1001" type="info">Hello</logEntry>
  <logEntry id="1002" type="error">problem</logEntry>

When done tokenizing, free the context:

void SplitstreamFree(SplitstreamState* state);

Note: This will free any buffers previously returned to the caller, so make sure not to use them after this function is called.

Once a context is specified, you need to drive the tokenizer

SplitstreamDocument SplitstreamGetNextDocument(
   SplitstreamState* state, 
   size_t max, 
   const char* buf, 
   size_t len, 
   SplitstreamScanner scanner);
SplitstreamDocument SplitstreamGetNextDocumentFromFile(
   SplitstreamState* s, 
   char* buf, 
   size_t bufferSize, 
   size_t max, 
   FILE* file, 
   SplitstreamScanner scanner);

The SplitstreamGetNextDocument function lets you feed data from a memory buffer into the tokenizer. The SplitstreamGetNextDocumentFromFile is a helper function for the common case that you have a FILE* object.

The state parameter is a state context that was previously initialized using SplitstreamInit or SplitstreamInitDepth.

The max parameter is the maximum allowable document length. If the internal buffer exceeds this size, tokenization restarts at whatever the current position was (this may cause the next document to be invalid as well).

The buf parameter is a pointer to the input data (SplitstreamGetNextDocument) or a preallocated read buffer (SplitstreamGetNextDocumentFromFile). The len parameter is the number of bytes of valid data, bufferSize is similarly the size of the buffer.

The file parameter is a pointer to an open file or stream. The file needs to be readable, but not seekable.

The scanner parameter is one of SplitstreamXMLScanner, SplitstreamJSONScanner or SplitstreamUBJSONScanner, or your own tokenizer implementing the following prototype:

typedef size_t (*SplitstreamScanner)(
   SplitstreamState* state, 
   const char* buf, 
   size_t len, 
   size_t* start);

The tokenization pattern

Important! The C API is a low level interface to the tokenizer and you need to follow the following pattern to ensure that no document is missed:

while(get_data(buf, &len)) {
   /* Feed new data and extract the first document, if any */
	doc = SplitstreamGetNextDocument(s, max, buf, len, scan);
	while(doc.buffer) {
	   /* Extract any documents remaining in the internal buffer */
	   doc = SplitstreamGetNextDocument(s, max, NULL, 0, scan);

Similarly, for the FILE* version

size_t len = 4096; /* Read buffer size */
char* buf = malloc(len);
while(!feof(f)) {
   /* Feed new data and extract the first document, if any */
	doc = SplitstreamGetNextDocumentFromFile(s, buf, len, 
	         max, file, scan);
	if(doc.buffer) {
do {
	/* Extract any documents remaining in the internal buffer */
	doc = SplitstreamGetNextDocumentFromFile(s, buf, len, 
	         max, NULL, scan);
	if(doc.buffer) {
} while(doc.buffer);

The Python interface


The easiest way to use splitstream is to create a requirements.txt for your project:

echo 'splitstream>=1.0.2' >> requirements.txt

or install manually

pip install splitstream

It is also very easy to clone this repo and build from source, see the Building section.


There is only one function in the Python interface:

splitfile(file, format[, callback[, startdepth
	[, bufsize[, maxdocsize[, preamble]]]]])

The file argument is a file-like object (e.g. open file or StringIO, or anything with a read([n]) method).

format is either "xml", "json" or "ubjson" and specifies the document type to split on.

startdepth helps parsing subtrees of "infinite" XML documents, such as

<?xml version="1.0" encoding="utf-8">
  <logEntry id="1000" type="info">Hello</logEntry>
  <logEntry id="1001" type="info">Hello</logEntry>
  <logEntry id="1002" type="error">problem</logEntry>

In this case, setting startdepth to 1 skips the <log> element and even allows parsing the logEntry elements before the document is finished (such as a current log file).

The callback argument is used to specify a callback function to be called with each document found in the stream. The splitfile function can operate in two modes:

  • Callback mode (specifying a callback). The documents are passed to the callback.
  • Generator mode (leaving callback unset or None). The documents are returned as a generator.

bufsize specifies the buffer size. It may make sense to increase this size when it is expected that the documents are large. Usually you should leave this as default.

maxdocsize specifies the maximum size for a document. Even if it is set, there is a maximum document size to prevent running out of memory in the case of oversized or malformed documents. If the internal buffer exceeds this size, tokenization restarts at whatever the current position was (this may cause the next document to be invalid as well).

preamble is an optional string that should be parsed before reading the file. By combining preamble with seeking the file, the header can be rewritten without filtering all subsequent reads. Another useful application is when reading the first few bytes to detect the file format (magic bytes) or when chaining stream splitters.


from splitstream import splitfile

with file("myfile.xmls") as f:
	for xml in splitfile(f, format="xml")):
   	    print xml


The C library

Currently there is no build script to make a shared or static library for use with C or compatible languages. The setup script will only build the Python module, but it should be trivial to add the source file to your project with minimal overhead.

The reason for this is that there are so many different build systems for C out there and there is no point choosing one or maintaining multiple build scripts. Adding the files to your Xcode, CMake or Visual Studio project is just a matter of adding the source and header files.

This library only depends on the standard C library.

The Python module

The Python module uses the standard distutils build. To build, use:

python setup.py build


python setup.py test

to run the unit tests. To install, use

python setup.py install