Streaming — Constant-Memory Parsing for Big XML¶

Everything covered so far (Objectify — Dotted Navigation, Dictify — XML to Dict, XPath) is built on pugixml’s in-memory DOM — fast, but the whole document has to fit in RAM. pygixml’s streaming layer is a second, independent engine: a self-contained, inlined yxml push parser that reads an XML file (or string, or file-like object) one chunk at a time and lets you process matching elements without ever holding the full document in memory. This is the layer to reach for once a file is too big — or you simply don’t want — to load whole.

import pygixml

for record in pygixml.iterfind("big.xml", "record"):
    print(record.tag, record.get("id"), record.find("name").text)
    record.clear()      # drop this element's memory before the next one

Three layers build on top of each other, from lowest- to highest-level:

Function	What it gives you
`pygixml.iterparse()`	`(event, elem)` pairs, `ElementTree`-style — full control
`pygixml.iterfind()`	Just the matched `StreamElement` objects
`pygixml.dictify.iterdict()` / `pygixml.jsonify.iterjsonl()`	Each match already converted to a `dict` / JSON `str`

And for the common “convert the whole file” case, pygixml.jsonify adds two endpoints that skip Python objects entirely and write straight to disk — see Jsonify — XML to JSON.

`iterparse` / `iterfind`¶

pygixml.iterparse(source, events=('end',), tag=None, stack_size=4096, chunk_size=65536)

An incremental, ElementTree-style parser. Reads source in chunk_size-byte chunks and yields (event, elem) tuples as elements start and/or end, without ever building a full document tree.

Parameters:

source (str or os.PathLike or bytes or bytearray or file-like) – A path, bytes/bytearray of XML content, or any file-like object with .read().
events (tuple[str, ...]) – Which events to yield: "start", "end", or both. Only "end" events carry a fully-populated element (children, text, attributes); a "start" event’s element has its tag and attributes but no children or text yet.
tag (str or None) – If given, only elements with this tag name produce events — everything else is skipped without allocating a StreamElement for it.
stack_size (int) – Size (bytes) of yxml’s internal element/attribute name stack. Increase this only if you hit a “stack too small” parse error on documents with unusually deep nesting or very long tag/attribute names.
chunk_size (int) – Bytes read per I/O operation from source.

Returns:

A generator of (event, elem) tuples.

Return type:

Iterator[tuple[str, pygixml.StreamElement]]

Raises:

PygiXMLError – On malformed XML.

import pygixml

for event, elem in pygixml.iterparse("big.xml", events=("start", "end")):
    if event == "start" and elem.tag == "record":
        print("entering record", elem.get("id"))
    elif event == "end" and elem.tag == "record":
        handle(elem)
        elem.clear()

pygixml.iterfind(source, tag, stack_size=4096, chunk_size=65536)

Shortcut for iterparse(source, events=("end",), tag=tag) that yields StreamElement objects directly — no (event, elem) tuple to unpack.

Parameters:

source – Same as iterparse().
tag (str) – Tag name of the elements to yield. Matches at any depth, including nested occurrences of the same tag.

Returns:

A generator of matched elements.

Return type:

Iterator[pygixml.StreamElement]

for record in pygixml.iterfind("big.xml", "record"):
    handle(record)
    record.clear()

`StreamElement`¶

class pygixml.StreamElement

A small, ElementTree-like element produced while streaming. It is not connected to a pugixml document — it’s a standalone tree of plain Python objects (built once, for this one match, then thrown away), with a tag, an attrib dict, optional text/tail strings, and child StreamElement nodes.

tag: str

attrib: dict

text: str or None

tail: str or None

children: list[pygixml.StreamElement]: Direct children. Also available via iteration (for child in elem), indexing (elem[0]), and len(elem).

get(key, default=None): attrib.get(key, default).

find(path): First descendant matching path, or None. path supports "tag", "a/b/c" (direct-child traversal), "*" (any child), and ".//tag" (any descendant).

findall(path): All descendants matching path (same syntax as find()). Always returns a list, possibly empty.

findtext(path, default=None): .text of the first match of path, or default.

iter(tag=None): Depth-first iterator over this element and all its descendants, optionally restricted to tag.

clear(): Drop this element’s attributes, text, tail, and children, freeing the memory they hold. Call this after processing each element yielded by iterfind() — it’s what keeps peak memory flat across millions of elements.

to_dict(attr_prefix='@', cdata_key='#text', force_list=None): Convert this element (and its subtree) to a plain dict, using the exact same conventions as pygixml.dictify.parse() (@-prefixed attributes, #text for mixed content, repeated siblings collapsed into a list).

to_json(attr_prefix='@', cdata_key='#text', force_list=None): Same conversion as to_dict(), returned as a JSON str instead of a dict.

for record in pygixml.iterfind("big.xml", "record"):
    d = record.to_dict()              # {'@id': '1', 'name': 'Ali', ...}
    line = record.to_json()           # '{"@id": "1", "name": "Ali", ...}'
    record.clear()

Streaming straight to `dict` or JSON¶

Wrapping every loop in elem.to_dict() / elem.to_json() / elem.clear() is common enough to have its own generators:

pygixml.dictify.iterdict(source, tag, attr_prefix='@', cdata_key='#text', force_list=None, stack_size=4096, chunk_size=65536)

Generator yielding to_dict() for every element matching tag, clearing each one automatically once converted. Identical to looping over iterfind() yourself, just shorter.

from pygixml import dictify

for record in dictify.iterdict("big.xml", "record"):
    print(record["@id"], record["name"])     # plain dict, no XML API

pygixml.jsonify.iterjsonl(source, tag, attr_prefix='@', cdata_key='#text', force_list=None, stack_size=4096, chunk_size=65536)

Generator yielding to_json() (one JSON object string per match) for every element matching tag. Each yielded line is independently parseable JSON — write them to a .jsonl file yourself, forward them over a socket, push them onto a queue, whatever fits:

from pygixml import jsonify

with open("big.jsonl", "w") as f:
    for line in jsonify.iterjsonl("big.xml", "record"):
        f.write(line + "\n")

If the destination really is just a .jsonl file and you don’t need the records in Python at all, pygixml.jsonify.stream_jsonl() does the same job without creating a single Python object per element — see Jsonify — XML to JSON.

Sources accepted everywhere¶

iterparse(), iterfind(), iterdict(), and iterjsonl() all accept the same set of source types:

Type	Behavior
`str` / `os.PathLike`	Treated as a filesystem path and opened for reading.
`bytes` / `bytearray`	Treated as XML content (not a path) and read from memory.
File-like object	Anything with a `.read()` method — sockets, `io.BytesIO`, already-open file handles, decompression streams, etc.

Note

A plain str is always treated as a path, never as XML content — pass bytes (e.g. xml.encode()) if you have an XML string in memory and want to stream it without writing it to a file first.

Memory model¶

Peak memory while streaming is bounded by the size of one matched element’s own subtree, not the document. A 10 GB XML file with millions of small, flat <record> elements streams in roughly the same peak memory as a 10 KB one — only the time scales with the file size, not the memory.

import pygixml

n, total_score = 0, 0
for record in pygixml.iterfind("huge_export.xml", "record"):
    n += 1
    total_score += int(record.findtext("score", "0"))
    record.clear()

print(f"{n} records, average score {total_score / n:.1f}")
# peak memory: roughly constant, regardless of how big huge_export.xml is

Streaming — Constant-Memory Parsing for Big XML¶

iterparse / iterfind¶

StreamElement¶

Streaming straight to dict or JSON¶

Sources accepted everywhere¶

Memory model¶

`iterparse` / `iterfind`¶

`StreamElement`¶

Streaming straight to `dict` or JSON¶