Change Log

Version 3.5.0

12.10.2026

🐍 Python 3.15 and free-threaded Python 3.14t / 3.15t wheels

⚡ Up to 2x faster DAG-CBOR encode, up to 10% faster decode

Encode	v3.4.1	v3.5.0	Speedup
`citm_catalog`	1.04 ms	0.52 ms	2.01x (+101%)
`github`	0.05 ms	0.03 ms	1.78x (+78%)
`twitter`	0.47 ms	0.27 ms	1.74x (+74%)
`canada`	0.83 ms	0.64 ms	1.29x (+29%)

Decode	v3.4.1	v3.5.0	Speedup
`canada`	4.22 ms	3.82 ms	1.10x (+10%)
`torture_cids`	63.5 ms	59.0 ms	1.08x (+8%)
`github`	0.07 ms	0.06 ms	1.06x (+6%)
`atproto_repo.car`	65.0 ms	61.8 ms	1.05x (+5%)
`twitter`	0.77 ms	0.75 ms	1.02x (+2%)
`citm_catalog`	1.71 ms	1.71 ms	1.00x (+0%)

❗ Behavior changes: -0.0 is now rejected on decode and encoded as 0.0, as the DAG-CBOR spec requires; encoding deeply nested data raises RecursionError instead of crashing; malformed CAR header types raise ValueError.

Add Python 3.15 support; bump PyO3 0.29 and deps, adopt Ruff 0.17 default rules by @MarshalX in #129
Add free-threaded wheels for Python 3.14t and 3.15t by @MarshalX in #130
Ship py.typed so type checkers pick up the bundled stubs by @MarshalX
Optimize DAG-CBOR encode by reusing a thread-local buffer and validating CIDs without Multihash construction by @MarshalX in #116
Optimize DAG-CBOR decode by reading each CBOR header once instead of re-decoding it per value by @MarshalX in #118
Optimize DAG-CBOR decode by reading multi-byte header arguments at a fixed width by @MarshalX in #121
Optimize DAG-CBOR encode by sorting map entries on a pooled stack and borrowing container items by @MarshalX in #122
Optimize DAG-CBOR encode/decode by dispatching leaf values inline and recursing only for containers by @MarshalX in #123
Forbid negative zero in DAG-CBOR: reject -0.0 on decode and encode it as 0.0 by @MarshalX in #126
Raise RecursionError instead of overflowing the stack when encoding deeply nested DAG-CBOR by @MarshalX in #128
Document raised exceptions and normalize CAR header type errors to ValueError by @MarshalX in #127
Restructure single-file lib.rs into per-domain modules by @MarshalX in #114
Replace _PyDict_SetItem_KnownHash with public PyDict_SetItem by @MarshalX in #124
Fix decode error handling: PyList leak, unbounded preallocation, and swallowed errors in decode_dag_cbor_multi by @MarshalX in #117
Fix list and bytearray races under free-threading by @MarshalX in #130
Fix Windows build after PyO3 0.29 by @MarshalX in #131
Fix encode_multibase docs: 'u' is base64url, not base58btc by @MarshalX

Version 3.4.1

27.05.2026

This is a focused decode-side optimization. The numbers below come from pytest-benchmark on Apple M1 / CPython 3.14 (PGO-built wheels), comparing v3.4.0 → v3.4.1 head-to-head.

⚡ Decode (DAG-CBOR)

Fixture	v3.4.0	v3.4.1	Speedup
`canada`	4.75 ms	4.43 ms	1.07x (+7%)
`github`	69 µs	66 µs	1.05x (+5%)
`twitter`	790 µs	766 µs	1.03x (+3%)
`citm_catalog`	1.79 ms	1.77 ms	1.01x (+1%)

The decoder used to fetch the Python recursion limit on every recursive step. We now read it once per top-level decode call and pass it through, closing the last decode gap with cbrrr (we're within 0.7% of it on canada and 30–40% ahead on the others).

🔧 Optimizations

Optimize DAG-CBOR decode by reading Py_GetRecursionLimit once per call by @MarshalX in #112

Deserialization v3.4.0 → v3.4.1

Serialization v3.4.0 → v3.4.1

Version 3.4.0

26.05.2026

❗ Up to ~2x faster DAG-CBOR encode and decode!

This release is all about throughput. The numbers below come from pytest-benchmark on an Apple M1 / CPython 3.14, comparing v3.3.2 → v3.4.0 head-to-head.

⚡ Decode (DAG-CBOR)

Fixture	v3.3.2	v3.4.0	Speedup
`github`	138 µs	70 µs	1.96x (+96%)
`twitter`	1.44 ms	0.80 ms	1.81x (+81%)
`citm_catalog`	2.87 ms	1.81 ms	1.59x (+59%)
`canada`	5.49 ms	4.77 ms	1.15x (+15%)

⚡ Encode (DAG-CBOR)

Fixture	v3.3.2	v3.4.0	Speedup
`canada`	2.31 ms	1.17 ms	1.98x (+98%)
`citm_catalog`	2.25 ms	1.26 ms	1.79x (+79%)
`twitter`	902 µs	635 µs	1.42x (+42%)
`github`	75.4 µs	69.7 µs	1.08x (+8%)

🔧 Optimizations

Optimize DAG-CBOR encode/decode and CAR block parsing by @MarshalX in #104
Optimize DAG-CBOR string decode with an ASCII fast path by @MarshalX in #105
Optimize DAG-CBOR map decode with a key-intern cache and KnownHash insert by @MarshalX in #106
Optimize int encode with a direct PyLong digit read by @MarshalX in #107

🧰 Toolchain & Build

Bump to Rust 1.95 and PyO3 0.28 by @MarshalX in #102
Drop the macos-13 PGO runner; cross-compile x86_64 macOS wheels from arm64 by @MarshalX in #108
Bump CI actions (Node 24) and pytest-codspeed v5 by @MarshalX in #103

📊 Benchmarks & Docs

Add a head-to-head benchmark against other Python DAG-CBOR libs by @MarshalX in #109
Update README and copyright years by @MarshalX in #110

Deserialization v3.3.2 → v3.4.0

Serialization v3.3.2 → v3.4.0

🏎️ Old-style performance comparison. Before (v3.3.2) VS After (v3.4.0):

Version 3.3.2

05.12.2025

Fix DASL testing CI by @MarshalX in #99
Fix leading zero handling of CIDs in DAG-CBOR by @MarshalX in #98

Version 3.3.1

03.12.2025

Add DASL testing harnesses to CI workflow by @MarshalX in #97
Fix error handling in array and map length checks by @MarshalX in #95
Fix error handling in CID and ULEB128 processing functions by @MarshalX in #96

Version 3.3.0

01.12.2025

❗ Almost x2 speed for CAR decoding (Bluesky firehose), ~25-35% faster DAG-CBOR decoding, and ~17% faster CID decoding!

I do have more ideas on what to try to speed it up even further! For example, applying BOLT on top of the existing PGO.

⚡ Optimizations

Huge performance gain for DAG-CBOR decoding and CID by replacing deprecated Rust crate libipld with cbor4ii by @vmx in #80 and even contributed upstream!
Optimize DAG-CBOR encoding for arrays and maps by @MarshalX in #88
Optimize CAR decoding by reducing allocations by @MarshalX in #90
Optimize executable size (panic = "abort") by @MarshalX in #86

Other

Add pytest-benchmark GitHub Actions workflow by @MarshalX in #82
Add lint and format check GHA workflow by @MarshalX in #83
Add CONTRIBUTING.md by @MarshalX in #85
Update CodSpeed by @MarshalX in #81
Fix formatting in profiling by @MarshalX in #84

🏎️ Performance comparison. Before (v3.2.0) VS After (v3.3.0):

You can find more benchmark results here: #92

Version 3.2.0

22.10.2025

❗ Possible breaking change for you: dropped support for PyPy 3.9 and PyPy 3.10 (both no longer supported upstream) in #77. CPython is still supported from 3.8 to 3.14.

Add Python 3.14 support by @moreati in #74
Update Rust to 1.90 by @MarshalX in #75
Update PyO3 to 0.27.1 by @MarshalX in #75
Fix repository URL specification in pyproject.toml file by @moreati in #73
Fix clippy warnings by @MarshalX in #76
Fix CodSpeed by @MarshalX in #78

Version 3.1.1

24.06.2025

Bump rust to 1.87.0 by @MarshalX in #67
Specify Rust version using env variable in workflows by @MarshalX in #67
Fix bignum encoding by @MarshalX in #66

Version 3.1.0

20.06.2025

Add type stubs by @zzstoatzz in #56
Add PyPy 3.11 support by @MarshalX in #63
Update PyO3 to 0.25.1 with performance optimizations by @MarshalX in #63
Update README by @MarshalX in #64
Fix panic on encoding map with integer keys by @MarshalX in #60
Fix panic on non UTF-8 string by @MarshalX in #59
Fix smallest and largest CBOR integer roundtrip by @MarshalX in #58
Fix decoding of infinities and NaNs according to specs by @MarshalX in #62
Fix handling of concatenated objects / CBOR sequence by @MarshalX in #65

Version 3.0.1

18.02.2025

Update PyO3 to 0.23.4 by @MarshalX in #54
Specify the dynamic version in pyproject.toml by @fxcoudert in #53
Fix stack overflow when decoding deeply nested structures like arrays and maps by @MarshalX in #51

Possible breaking change for you: dropped support for PyPy 3.8

Version 3.0.0

24.10.2024

❗ Breaking changes

Drop Python 3.7 support by @MarshalX in #50

🆕 Features

Add Python 3.13 support by @MarshalX in #50

Other

Update PyO3 to 0.22.5 by @MarshalX in #50
Update decode_car signature, bytes CIDs instead of str in README by @snarfed in #49

Version 2.0.0

07.09.2024

❗ Breaking changes

CID in CAR block keys and CID DAG-CBOR are now in raw byte form instead of stringified one!

Encode and decode CID as bytes (major 2) instead of string (major 3) by @MarshalX in #36

⚡ Bunch of new optimizations

Optimize duplicate check by @MarshalX in #34
Optimize type checking by @MarshalX in #37
Optimize keys duplication check by @MarshalX in #39
Optimize array creation by @MarshalX in #40
Optimize string encoding by @MarshalX in #41

🆕 Features

Add CID encoding (encode_cid(bytes) -> str) by @MarshalX in #38
Add DAG-CBOR profiling by @MarshalX in #35

Other

Fix upload to PyPI by @MarshalX in #33
Update CodSpeed to v3 by @MarshalX in #47
Update PyO3 to 0.22.2 by @MarshalX in #48
Lock rust toolchain on v1.76.0 because higher versions segfaults during PGO gather by @MarshalX in #46

🏎️ Performance comparison. Before (v1.2.3) VS After (v2.0.0):

Raw bench results

Before (v1.2.3):

Hello World Decode:
===================
libipld  : 144 ns

Hello World Encode:
===================
libipld  : 150 ns

Realistic Decode Tests:
=======================
canada.json.dagcbor            libipld  : 5.11 ms (197.31 MB/s)
citm_catalog.json.dagcbor      libipld  : 3.43 ms (95.31 MB/s)
github.json.dagcbor            libipld  : 0.17 ms (270.36 MB/s)
twitter.json.dagcbor           libipld  : 2.15 ms (178.96 MB/s)

Realistic Decode Car Tests:
===========================
atproto_repo.car               libipld  : 119.97 ms (154.22 MB/s)

Realistic Encode Tests:
=======================
canada.json.dagcbor            libipld  : 1.19 ms (848.65 MB/s)
citm_catalog.json.dagcbor      libipld  : 1.66 ms (196.42 MB/s)
github.json.dagcbor            libipld  : 0.11 ms (429.15 MB/s)
twitter.json.dagcbor           libipld  : 0.87 ms (443.77 MB/s)

After (v2.0.0):

Hello World Decode:
===================
libipld  : 124 ns

Hello World Encode:
===================
libipld  : 116 ns

Realistic Decode Tests:
=======================
canada.json.dagcbor            libipld  : 4.65 ms (216.57 MB/s)
citm_catalog.json.dagcbor      libipld  : 2.96 ms (110.37 MB/s)
github.json.dagcbor            libipld  : 0.15 ms (311.00 MB/s)
twitter.json.dagcbor           libipld  : 1.75 ms (219.74 MB/s)

Realistic Decode Car Tests:
===========================
atproto_repo.car               libipld  : 92.44 ms (200.13 MB/s)

Realistic Encode Tests:
=======================
canada.json.dagcbor            libipld  : 1.00 ms (1011.47 MB/s)
citm_catalog.json.dagcbor      libipld  : 1.60 ms (203.57 MB/s)
github.json.dagcbor            libipld  : 0.06 ms (733.96 MB/s)
twitter.json.dagcbor           libipld  : 0.56 ms (686.31 MB/s)

Test stand: M1 Pro, Python 3.12.

I highly recommend using Python 3.12. The lower version will be slower.

P.S. Breaking change not only gives a performance boost but also aligns implementations with other languages.

Version 1.2.3

11.03.2024

⚡ Thanks to PGO and new Bound API in pyo3 v0.21 I achieved huge performance boost ones again! The pyo3 team did an awesome work! Many thanks to all contributors! Also, I implemented own CAR reader that gives +11%

Migrate to pyo3 v0.21 Bound API by @MarshalX in #23
Add PGO-optimized wheels by @MarshalX in #30, #21
Add own implementation of CAR reader by @MarshalX in #28
Add IPLD Codec Fixtures by @MarshalX in #19
Add CAR and CID benchmarks by @MarshalX in #20
Improve work with numbers by @MarshalX in #25
Use refs in args and Bound in return type where possible by @MarshalX in #31

🏎️ Performance comparison. Before (v1.2.2) VS After (v1.2.3):

Raw bench results

Before (v1.2.2):

Hello World Decode:
===================
libipld  : 275 ns

Hello World Encode:
===================
libipld  : 237 ns

Realistic Decode Tests:
=======================
canada.json.dagcbor            libipld  : 7.51 ms (134.15 MB/s)
citm_catalog.json.dagcbor      libipld  : 4.97 ms (65.71 MB/s)
github.json.dagcbor            libipld  : 0.23 ms (201.43 MB/s)
twitter.json.dagcbor           libipld  : 2.70 ms (142.51 MB/s)

Realistic Decode Car Tests:
===========================
atproto_repo.car               libipld  : 170.09 ms (108.77 MB/s)

Realistic Encode Tests:
=======================
canada.json.dagcbor            libipld  : 1.89 ms (533.01 MB/s)
citm_catalog.json.dagcbor      libipld  : 2.86 ms (114.02 MB/s)
github.json.dagcbor            libipld  : 0.14 ms (335.64 MB/s)
twitter.json.dagcbor           libipld  : 1.07 ms (358.30 MB/s)

After (v1.2.3):

Hello World Decode:
===================
libipld  : 130 ns

Hello World Encode:
===================
libipld  : 143 ns

Realistic Decode Tests:
=======================
canada.json.dagcbor            libipld  : 5.24 ms (192.11 MB/s)
citm_catalog.json.dagcbor      libipld  : 3.49 ms (93.51 MB/s)
github.json.dagcbor            libipld  : 0.18 ms (257.51 MB/s)
twitter.json.dagcbor           libipld  : 2.11 ms (182.31 MB/s)

Realistic Decode Car Tests:
===========================
atproto_repo.car               libipld  : 123.72 ms (149.55 MB/s)

Realistic Encode Tests:
=======================
canada.json.dagcbor            libipld  : 1.22 ms (823.65 MB/s)
citm_catalog.json.dagcbor      libipld  : 1.69 ms (192.80 MB/s)
github.json.dagcbor            libipld  : 0.11 ms (410.94 MB/s)
twitter.json.dagcbor           libipld  : 0.80 ms (481.72 MB/s)

Test stand: M1 Pro, Python 3.12.

I highly recommend using Python 3.12. The lower version will be slower.

Version 1.2.2

02.03.2024

Add unit tests and benchmarks by @MarshalX in #11
Add CodSpeed by @MarshalX in #14
Fix encode_multibase input data type by @MarshalX in #12
Fix DAG CBOR string encoding by @MarshalX in #13
Fix duplicated map keys check in decode_dag_cbor by @MarshalX in #16
Fix non-string map keys check in decode_dag_cbor by @MarshalX in #17
Fix DAG-CBOR map keys order for encoding and decoding by @MarshalX in #18

🏎️ Performance comparison

Before (v1.2.1):

canada.json.dagcbor            libipld  : 1.98 ms (508.99 MB/s)
citm_catalog.json.dagcbor      libipld  : 3.96 ms (82.46 MB/s)
twitter.json.dagcbor           libipld  : 1.92 ms (200.10 MB/s)

After (v1.2.2):

canada.json.dagcbor            libipld  : 2.04 ms (493.43 MB/s)
citm_catalog.json.dagcbor      libipld  : 2.76 ms (118.31 MB/s)
twitter.json.dagcbor           libipld  : 1.03 ms (371.93 MB/s)

CodSpeed Performance Report: #18#issuecomment-1974061926

Version 1.2.1

27.02.2024

Improve performance of DAG-CBOR encoding by @MarshalX in #10

Encoding results compared to v1.2.0:

canada.json.dagcbor x20.6
citm_catalog.json.dagcbor x4.7
twitter.json.dagcbor x4.1

Version 1.2.0

26.02.2024

Add encode_dag_cbor by @MarshalX in #5
Add profiling by @MarshalX in #6
Improve performance of DAG-CBOR decoding (remove a lot of unnecessary allocs and drops) by @MarshalX in #8

Decoding results compared to v1.1.1:

citm_catalog.json.dagcbor x2
canada.json.dagcbor x1.64
twitter.json.dagcbor x1.29

Decoding results compared to v1.1.0:

citm_catalog.json.dagcbor x4.9
canada.json.dagcbor x4.6
twitter.json.dagcbor x4.37

Version 1.1.1

24.02.2024

Add error handling by @MarshalX in #4
Improve performance (at least x2 boost for CAR and DAG-CBOR decoding) by @MarshalX in #4
Fix CID decoding in DAG-CBOR by @MarshalX in #4

Version 1.1.0

10.01.2024

Add multibase encode and decode by @MarshalX in #3

Version 1.0.1

27.10.2023

Add Python 3.12 by @MarshalX in #2

Version 1.0.0

13.09.2023

The first stable release

Features

Decode CID (decode_cid(str) -> dict)
Decode DAG-CBOR (decode_dag_cbor(bytes) -> dict, decode_dag_cbor_multi(bytes) -> list[dict])
Decode CAR (decode_car(bytes) -> tuple[dict, dict[str, dict]]). Returns a header and blocks mapped by CID.

by @MarshalX

Navigation Menu

Repository navigation

CHANGES.md

CHANGES.md

Change Log

Version 3.5.0

Version 3.4.1

⚡ Decode (DAG-CBOR)

🔧 Optimizations

Deserialization v3.4.0 → v3.4.1

Serialization v3.4.0 → v3.4.1

Version 3.4.0

⚡ Decode (DAG-CBOR)

⚡ Encode (DAG-CBOR)

🔧 Optimizations

🧰 Toolchain & Build

📊 Benchmarks & Docs

Deserialization v3.3.2 → v3.4.0

Serialization v3.3.2 → v3.4.0

Version 3.3.2

Version 3.3.1

Version 3.3.0

Version 3.2.0

Version 3.1.1

Version 3.1.0

Version 3.0.1

Version 3.0.0

Version 2.0.0

Version 1.2.3

Version 1.2.2

Version 1.2.1

Version 1.2.0

Version 1.1.1

Version 1.1.0

Version 1.0.1

Version 1.0.0

Features

FilesExpand file tree

CHANGES.md

Latest commit

History

CHANGES.md

File metadata and controls

Change Log

Version 3.5.0

Version 3.4.1

⚡ Decode (DAG-CBOR)

🔧 Optimizations

Deserialization v3.4.0 → v3.4.1

Serialization v3.4.0 → v3.4.1

Version 3.4.0

⚡ Decode (DAG-CBOR)

⚡ Encode (DAG-CBOR)

🔧 Optimizations

🧰 Toolchain & Build

📊 Benchmarks & Docs

Deserialization v3.3.2 → v3.4.0

Serialization v3.3.2 → v3.4.0

Version 3.3.2

Version 3.3.1

Version 3.3.0

Version 3.2.0

Version 3.1.1

Version 3.1.0

Version 3.0.1

Version 3.0.0

Version 2.0.0

Version 1.2.3

Version 1.2.2

Version 1.2.1

Version 1.2.0

Version 1.1.1

Version 1.1.0

Version 1.0.1

Version 1.0.0

Features