yarutsk 0.2.1

A YAML round-trip library that preserves comments and insertion order

yarutsk

A Python YAML library that round-trips documents while preserving comments, insertion order, scalar styles, tags, anchors and aliases, blank lines, and explicit document markers.

What it does

Most YAML libraries silently drop comments on load. yarutsk keeps them attached to their keys — both inline (key: value # like this) and block-level (# above a key) — so a load → modify → dump cycle leaves the rest of the file intact.

import io
import yarutsk

doc = yarutsk.load(io.StringIO("""
# database config
host: localhost  # primary
port: 5432
"""))

doc["port"] = 5433

out = io.StringIO()
yarutsk.dump(doc, out)
print(out.getvalue())
# # database config
# host: localhost  # primary
# port: 5433

YamlMapping is a subclass of dict and YamlSequence is a subclass of list, so they work everywhere a dict or list is expected:

import json

doc = yarutsk.loads("name: Alice\nscores: [10, 20, 30]")

isinstance(doc, dict)           # True
isinstance(doc["scores"], list) # True
json.dumps(doc)                 # '{"name": "Alice", "scores": [10, 20, 30]}'

Round-trip fidelity

yarutsk reproduces the source text exactly for everything it understands. A loads followed by dumps gives back the original string byte-for-byte in the common case:

src = """\
defaults: &base
  timeout: 30
  retries: 3

service:
  name: api
  config: *base
"""
assert yarutsk.dumps(yarutsk.loads(src)) == src

Specifically preserved:

• Scalar styles — plain, 'single-quoted', "double-quoted", literal block |, folded block >
• Non-canonical scalars — yes/no/on/off, ~, Null, True/False, 0xFF, 0o77 — reproduced as written, not re-canonicalised to true/false/null/255
• YAML tags — !!str, !!python/tuple, and any custom tag are emitted back verbatim
• Anchors and aliases — &name on the anchor node and *name for references are preserved; the Python layer returns the resolved value transparently
• Blank lines between mapping entries and sequence items
• Explicit document markers — --- and ...

Installation

Built with Maturin. From the repo root:

pip install maturin
maturin develop

API

Loading and dumping

# Load from stream (StringIO / BytesIO)
doc  = yarutsk.load(stream)            # first document
docs = yarutsk.load_all(stream)        # all documents as a list

# Load from string
doc  = yarutsk.loads(text)
docs = yarutsk.loads_all(text)

# Dump to stream
yarutsk.dump(doc, stream)
yarutsk.dump_all(docs, stream)

# Dump to string
text = yarutsk.dumps(doc)
text = yarutsk.dumps_all(docs)

load / loads return a YamlMapping, YamlSequence, or YamlScalar (for a top-level scalar document), or None for empty input. Nested container nodes are YamlMapping or YamlSequence; scalar leaves inside mappings and sequences are returned as native Python primitives (int, float, bool, str, bytes, datetime.datetime, datetime.date, or None).

Type conversions

Implicit coercion

Plain YAML values (no tag) are converted to Python types automatically:

Value pattern	Python type	Examples
Decimal integer	int	42, -7
Hex / octal integer	int	0xFF → 255, 0o17 → 15
Float	float	3.14, 1.5e2, .inf, -.inf, .nan
true / false (any case)	bool	True, FALSE
yes / no / on / off (any case)	bool	YAML 1.1 booleans
null, Null, NULL, ~, empty value	None	—
Anything else	str	hello, "quoted"

Non-canonical forms are reproduced as written on dump — yes stays yes, 0xFF stays 0xFF, ~ stays ~.

Explicit tags

A !!tag overrides implicit coercion and controls which Python type is returned:

Tag	Python type	Notes
!!str	str	Forces string even if the value looks like an int, bool, or null
!!int	int	Parses decimal, hex (0xFF), and octal (0o17)
!!float	float	Promotes integer literals (!!float 1 → 1.0)
!!bool	bool	—
!!null	None	Forces null regardless of content (!!null "" → None)
!!binary	bytes	Base64-decoded on load; base64-encoded on dump
!!timestamp	datetime.datetime or datetime.date	Date-only values return date; datetime values return datetime

Tags are preserved through the round-trip: load → dump reproduces the original tag and source text exactly.

import datetime

# !!binary
doc = yarutsk.loads("data: !!binary aGVsbG8=\n")
doc["data"]                            # b'hello'
yarutsk.dumps(doc)                     # 'data: !!binary aGVsbG8=\n'

# !!timestamp
doc = yarutsk.loads("ts: !!timestamp 2024-01-15T10:30:00\n")
doc["ts"]                              # datetime.datetime(2024, 1, 15, 10, 30)

doc = yarutsk.loads("ts: !!timestamp 2024-01-15\n")
doc["ts"]                              # datetime.date(2024, 1, 15)

# !!float promotes integers
doc = yarutsk.loads("x: !!float 1\n")
doc["x"]                               # 1.0  (float, not int)

# !!str forces a string
doc = yarutsk.loads("x: !!str 42\n")
doc["x"]                               # '42'

# Dumping Python bytes / datetime automatically produces the right tag
mapping = yarutsk.loads("x: placeholder\n")
mapping["x"] = b"hello"
yarutsk.dumps(mapping)                 # 'x: !!binary aGVsbG8=\n'

mapping["x"] = datetime.datetime(2024, 1, 15, 10, 30)
yarutsk.dumps(mapping)                 # 'x: !!timestamp 2024-01-15T10:30:00\n'

Schema — custom types

Schema lets you register loaders (tag → Python object, fired on load) and dumpers (Python type → tag + data, fired on dump). Pass it as a keyword argument to any load or dump function.

Mapping types

The loader receives a YamlMapping (dict-like); the dumper returns a (tag, dict) tuple:

import yarutsk

class Point:
    def __init__(self, x, y): self.x, self.y = x, y

schema = yarutsk.Schema()
schema.add_loader("!point", lambda d: Point(d["x"], d["y"]))
schema.add_dumper(Point, lambda p: ("!point", {"x": p.x, "y": p.y}))

doc = yarutsk.loads("origin: !point\n  x: 0\n  y: 0\n", schema=schema)
doc["origin"]                          # Point(0, 0)

doc["pos"] = Point(3, 4)               # assigning custom objects works too
yarutsk.dumps(doc, schema=schema)
# origin: !point
#   x: 0
#   y: 0
# pos: !point
#   x: 3
#   y: 4

Scalar types

The loader receives the raw scalar string; the dumper returns a (tag, str) tuple:

class Color:
    def __init__(self, r, g, b): self.r, self.g, self.b = r, g, b

schema = yarutsk.Schema()
schema.add_loader("!color", lambda s: Color(*[int(x) for x in s.split(",")]))
schema.add_dumper(Color, lambda c: ("!color", f"{c.r},{c.g},{c.b}"))

doc = yarutsk.loads("bg: !color 255,0,128\n", schema=schema)
doc["bg"]                              # Color(255, 0, 128)
yarutsk.dumps(doc, schema=schema)      # 'bg: !color 255,0,128\n'

Overriding built-in tags

Registering a loader for !!int, !!float, !!bool, !!null, or !!str bypasses the built-in coercion. The callable receives the raw YAML string rather than the already-converted Python value:

schema = yarutsk.Schema()
schema.add_loader("!!int", lambda raw: int(raw, 0))  # parses 0xFF, 0o77, etc.

doc = yarutsk.loads("x: !!int 0xFF\n", schema=schema)
doc["x"]                               # 255

Multiple dumpers for the same type are checked in registration order; the first isinstance match wins.

YamlScalar

Top-level scalar documents are wrapped in a YamlScalar node:

doc = yarutsk.loads("42")
doc.value                              # 42 (Python int)
doc.to_dict()                          # same as .value

# Scalar style
doc = yarutsk.loads("---\n'hello'\n")
doc.style                              # 'single'
doc.style = "double"                   # 'plain'|'single'|'double'|'literal'|'folded'

# YAML tag
doc = yarutsk.loads("!!str 42")
doc.tag                                # '!!str'
doc.tag = None                         # clear tag

# Explicit document markers
doc = yarutsk.loads("---\n42\n...")
doc.explicit_start                     # True
doc.explicit_end                       # True
doc.explicit_start = False
doc.explicit_end   = False

YamlMapping

YamlMapping is a subclass of dict with insertion-ordered keys. All standard dict operations work directly:

# Standard dict interface (inherited)
doc["key"]                             # get (KeyError if missing)
doc["key"] = value                     # set (preserves position if key exists)
del doc["key"]                         # delete
"key" in doc                           # membership test
len(doc)                               # number of entries
for key in doc: ...                    # iterate over keys in order
doc.keys()                             # KeysView in insertion order
doc.values()                           # ValuesView in insertion order
doc.items()                            # ItemsView of (key, value) pairs
doc.get("key")                         # returns None if missing
doc.get("key", default)                # returns default if missing
doc.pop("key")                         # remove & return (KeyError if missing)
doc.pop("key", default)                # remove & return, or default
doc.setdefault("key", default)         # get or insert default
doc.update(other)                      # merge from dict or YamlMapping
doc == {"a": 1}                        # equality comparison

# Works with any dict-expecting library
isinstance(doc, dict)                  # True
json.dumps(doc)                        # works

# Conversion
doc.to_dict()                          # deep conversion to plain Python dict

# Comments (1-arg = get, 2-arg = set; pass None to clear)
doc.comment_inline("key")             # -> str | None
doc.comment_before("key")             # -> str | None
doc.comment_inline("key", text)
doc.comment_before("key", text)

# YAML tag
doc.tag                                # -> str | None  (e.g. '!!python/object:Foo')
doc.tag = "!!map"

# Explicit document markers
doc.explicit_start                     # bool
doc.explicit_end                       # bool
doc.explicit_start = True
doc.explicit_end   = True

# Node access — returns YamlScalar/YamlMapping/YamlSequence preserving style/tag/anchor
node = doc.node("key")                # KeyError if absent

# Scalar style shortcut (equivalent to: doc.node("key").style = "single")
doc.scalar_style("key", "single")     # 'plain'|'single'|'double'|'literal'|'folded'

# Sorting
doc.sort_keys()                        # alphabetical, in-place
doc.sort_keys(reverse=True)            # reverse alphabetical
doc.sort_keys(key=lambda k: len(k))    # custom key function on key strings
doc.sort_keys(recursive=True)          # also sort all nested mappings

YamlSequence

YamlSequence is a subclass of list. All standard list operations work directly:

# Standard list interface (inherited)
doc[0]                                 # get by index (negative indices supported)
doc[0] = value                         # set by index
del doc[0]                             # delete by index
value in doc                           # membership test
len(doc)                               # number of items
for item in doc: ...                   # iterate over items
doc.append(value)                      # add to end
doc.insert(idx, value)                 # insert before index
doc.pop()                              # remove & return last item
doc.pop(idx)                           # remove & return item at index
doc.remove(value)                      # remove first occurrence (ValueError if missing)
doc.extend(iterable)                   # append items from list or YamlSequence
doc.index(value)                       # index of first occurrence
doc.count(value)                       # number of occurrences
doc.reverse()                          # reverse in-place
doc == [1, 2, 3]                       # equality comparison

# Works with any list-expecting library
isinstance(doc, list)                  # True
json.dumps(doc)                        # works

# Conversion
doc.to_dict()                          # deep conversion to plain Python list

# Comments (1-arg = get, 2-arg = set; pass None to clear)
doc.comment_inline(idx)               # -> str | None
doc.comment_before(idx)               # -> str | None
doc.comment_inline(idx, text)
doc.comment_before(idx, text)

# YAML tag
doc.tag                                # -> str | None  (e.g. '!!python/tuple')
doc.tag = None

# Explicit document markers
doc.explicit_start                     # bool
doc.explicit_end                       # bool
doc.explicit_start = True
doc.explicit_end   = True

# Sorting (preserves comment metadata)
doc.sort()                             # natural order, in-place
doc.sort(reverse=True)
doc.sort(key=lambda v: len(v))         # custom key function on item values

Sorting preserves all comments — each entry or item carries its inline and before-key comments with it when reordered.

Benchmarks

Compare load, dump, and round-trip performance against PyYAML and ruamel.yaml across small, medium, and large inputs:

uv sync --group benchmark
uv run maturin develop --release
uv run pytest benchmarks/ -v --benchmark-sort=name

Running tests

You need Rust 1.85+ and Python 3.12+ with uv. Python 3.12 is the minimum — YamlSequence subclasses list, which requires PyO3's extends = PyList support introduced in Python 3.12.

# 1. Clone with the yaml-test-suite submodule
git clone --recurse-submodules https://github.com/theyugin/yarutsk
cd yarutsk

# 2. Create a virtual environment and install dev dependencies
uv sync --group dev

# 3. Build the extension in dev (debug) mode
uv run maturin develop

# 4. Run the suites
uv run pytest tests/ --ignore=tests/test_yaml_suite.py -v  # core library tests
uv run pytest tests/test_yaml_suite.py -q                   # yaml-test-suite compliance

test_yaml_suite.py requires the yaml-test-suite submodule. Tests that fail due to known YAML normalisation differences are marked xfail and do not count as failures.

Internals

The scanner and parser are vendored from yaml-rust2 (MIT licensed) with one targeted modification: the comment-skipping loop in the scanner now emits Comment tokens instead of discarding them. Everything else — block/flow parsing, scalar type coercion, multi-document support — comes from yaml-rust2 unchanged. The builder layer wires those tokens to the data model, and a hand-written block-style emitter serialises it back out.

YamlMapping and YamlSequence are PyO3 pyclasses that extend Python's built-in dict and list types. A Rust inner field stores the full YAML data model (including comments); the parent dict/list is kept in sync on every mutation so that all standard Python operations work transparently.

Disclaimer

This library was created with Claude Code (Anthropic). The design, implementation, tests, and this README were written by Claude under human direction.