arpeggio-shredder 0.2.1

Reversible JSON ↔ Arrow shredding with schema stability

arpeggio-shredder

Reversible shredding (flatten) and unshredding (rebuild) of semi-structured JSONL into Apache Arrow RecordBatches.

Source Code: https://codeberg.org/alwyna/shredder

This package provides a thin Python binding over a C++ core that converts JSON documents into a columnar “atoms” representation suitable for Arrow / Parquet workflows, while preserving the ability to reconstruct the original documents exactly.

The scope is intentionally narrow: flattening with reversibility, not general JSON processing.

Key properties

• Reversible: JSON → Arrow atoms → JSON
• Columnar-first: output is optimized for Arrow-native pipelines
• Deterministic identity: optional object and transaction tagging
• Minimal Python surface: High-level shredder package or raw shredder_ext extension

Installation

pip install arpeggio-shredder

Usage

Shredder 0.2.0 supports two explicit "codec" strategies: json (UTF-8 text) and arrow (native Struct/List DOM).

Known limitations / Non-goals for 0.2.x

• No support for array-of-array: Shredder throws a runtime_error on nested arrays (e.g., [[1,2]]).
• Row expansion semantics (no sibling back-fill): Fields encountered after an array may only be present in the last expanded row. Reconstruction remains correct via internal metadata.
• Depth/complexity caveat: The iterative core is frozen for 0.2.0. We prioritize stability over redesigning the traversal engine for extreme nesting.
• Codec strategy is explicit: codec="json" requires strings; codec="arrow" requires Arrow DOM (struct/list). No auto-detection.

1. JSON Text Codec (Default)

Warning: The json codec expects UTF-8 JSON strings (e.g., pa.string()), not Python dicts or Arrow structs. No silent casting is performed.

The classic workflow: JSONL → Arrow atoms → JSONL.

import pyarrow as pa
import shredder

# 1. Load JSONL as Arrow string array
lines = ['{"id": 1, "msg": "hello"}', '{"id": 2, "msg": "world"}']
array = pa.array(lines, type=pa.string())

# 2. Shred into columnar "atoms"
# codec="json" is default, but can be explicit
atoms = shredder.shred(array, codec="json")

# 3. Unshred back to JSON
# Returns RecordBatch with single column 'doc'
reconstructed = shredder.unshred(atoms, codec="json")
json_docs = reconstructed.column(0)

for doc in json_docs:
    print(doc.as_py())

2. Arrow DOM Codec

New in 0.2.0: Treat Arrow nested structures as the source of truth. No JSON parsing involved.

import pyarrow as pa
import pyarrow.json as pajson
import shredder

# 1. Read JSONL into Arrow-native columns
# table = pajson.read_json("input.jsonl")
table = pa.table({"id": [1, 2], "val": [10.5, 20.0]})

# 2. Convert to a single struct column (Arrow DOM input)
doc_struct = pa.StructArray.from_arrays(table.columns, names=table.column_names)

# 3. Use the stateful Shredder class
s = shredder.Shredder(codec="arrow")

# 4. Shred/Unshred
atoms = s.shred(doc_struct)
reconstructed = s.unshred(atoms)

# Output 'doc' column is the reconstructed StructArray
doc_out = reconstructed.column(0)
print(doc_out.to_pylist())

How it works

1. Codecs:
   • JsonTextCodec: 1-col UTF-8 string → internal DOM → atoms.
   • ArrowDomCodec: 1-col Arrow struct/list/etc → internal DOM → atoms.
2. Flattening: Nested objects are flattened into path-based columns (e.g., root_user_name).
3. Row Expansion: Nested arrays are "shredded" by duplicating parent values for each array element, adding an __idx column to preserve order.
4. Identity: The engine automatically injects __obj_id and __txn_id (BLAKE3 hashes of content) to ensure perfect reconstruction.
5. Reversibility: Shredder preserves enough metadata to rebuild the exact original structure, regardless of the input codec.

Native extension

The package ships with a prebuilt native extension (.so) built against Apache Arrow and exposed via pybind11.

• No runtime compilation
• No system Arrow installation required
• Shared libraries are bundled into the wheel

Platform support

• OS: Linux (manylinux-compatible)
• Architecture: x86_64
• Python: CPython 3.12
• ABI: glibc (manylinux)

Other platforms are not currently supported.

Relationship to the C++ project

This package is the Python distribution layer for the Shredder C++ project hosted on Codeberg: https://codeberg.org/alwyna/shredder.

License

This package is dual-licensed:

• AGPL-3.0 for open-source use and networked deployments
• Commercial license for proprietary or closed-source use

Commercial licensing is available via https://arpeggio.one/shop.