NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #774 most downloaded on PyPI
Modern high-performance serialization utilities for Python
Latest stable
3.0.1
Last release 18 days ago
03 Nov 2026
Release timing varies
gaps range from 9 days to 1.2 years
Most releases are documented
notes for 31 of 40 stable releases
1 version withdrawn
withdrawn after publishing
8 years old
48 releases · first in 2018
Drag across the ruler to compare, or tap a release to open it.
Fix reading gzipped JSONL files with blank lines ( #125 ). read_gzip_jsonl reads the file in binary mode, so blank lines arrived as b"" and weren't sk
read_gzip_jsonl reads the file in binary mode, so blank lines arrived as b"" and weren't skipped. Instead it raised ValueError: Invalid JSON on line N: b''. This included files written by write_gzip_jsonl(..., append=True), which adds a blank separator line, so a srsly write-then-append round trip couldn't be read back. The bug was present since read_gzip_jsonl was added in v2.4.6.Invalid JSON on line N errors from read_jsonl and read_gzip_jsonl now matches the line in the file. Previously, blank lines and lines skipped as invalid weren't counted.read_gzip_jsonl now show the offending line as text instead of b'...'.json_dumps(data, indent=None), which raised TypeError unless sort_keys=True was also set. indent=None now gives compact output, the same as indent=0.Your coding agent can read these notes before it upgrades. Set up the MCP server →
Align with thinc's workflow: trigger only on release-v*/prerelease-v* tags instead of branch pushes/PRs, pin the reusable workflow to a specific SHA,
Align with thinc's workflow: trigger only on release-v*/prerelease-v*
tags instead of branch pushes/PRs, pin the reusable workflow to a
specific SHA, remove the now-unnecessary create-release input, and add
the retag_release job that re-points the draft release to a clean v* tag.
Backport of the v3.0.1 fixes to the v2.x series.
Backport of the v3.0.1 fixes to the v2.x series.
read_gzip_jsonl reads the file in binary mode, so blank lines arrived as b"" and weren't skipped. Instead it raised ValueError: Invalid JSON on line N: b''. This included files written by write_gzip_jsonl(..., append=True), which adds a blank separator line, so a srsly write-then-append round trip couldn't be read back. The bug was present since read_gzip_jsonl was added in v2.4.6.Invalid JSON on line N errors from read_jsonl and read_gzip_jsonl now matches the line in the file. Previously, blank lines and lines skipped as invalid weren't counted.read_gzip_jsonl now show the offending line as text instead of b'...'.json_dumps(data, indent=None), which raised TypeError unless sort_keys=True was also set. indent=None now gives compact output, the same as indent=0.Add Python 3.15 support and real free-threading support, set version …
Add Python 3.15 support and real free-threading support, set version …
…to v2.5.5
- Raise python_requires cap to <3.16; add 3.14/3.15 classifiers
- Declare the vendored ujson and msgpack extensions free-threading
compatible. Previously importing srsly re-enabled the GIL on
free-threaded builds.
- Fix use-after-free in the ujson encoder when a list is mutated during
encoding (by another thread, or a __json__/toDict callback): iterate
over a private copy. Take a strong reference to dict values in the
sort_keys path.
- Require Cython>=3.1 for the freethreading_compatible directive
- Test on 3.14t, 3.15 and 3.15t; update checkout/setup-python to v6
- Build with eMudhra-hardened gha-cibuildwheel (cibuildwheel v4.3.3)
Security fix for the vendored msgpack, plus fixes for crashes in the vendored ujson.
Security fix for the vendored msgpack, plus fixes for crashes in the vendored ujson.
srsly.msgpack.Unpacker after a failed unpack could crash the process. Backports the upstream msgpack-python fix. The high-level helpers such as srsly.msgpack_loads create a new unpacker for each call, so they were not exposed. (#124, #127)srsly.ujson.dumps(..., sort_keys=True) on a dict. (#128)srsly.ujson.dumps(..., sort_keys=True) is given keys that can't be sorted, such as {1: 1, "a": 2}. It now raises TypeError, like the standard library json module. (#128)srsly.ujson when encoding a large object fails partway through. (#128)srsly.json_dumps was not affected by the sort_keys crashes, because it uses the standard library json module when sort_keys=True.
Python 3.14 support via updated cloudpickle
Python 3.14 support via updated cloudpickle
Nothing published for this version
Now that Github have runners for Linux ARM, we can build wheels for that again 🎉
Now that Github have runners for Linux ARM, we can build wheels for that again 🎉
Update vendored msgpack-python to v1.1.0
msgpack-python to v1.1.0Updates for Python 3.12 support (#102).
Restrict to Cython 0.29.x due to incompatibilities with Cython 3.
Fix typing for JSONInput and JSONInputBin (#79).
JSONInput and JSONInputBin (#79).read_gzip_jsonl and write_gzip_jsonl (#84).Update to cloudpickle v2.2.0 for Python 3.11 support.
Nothing published for this version
…handling of invalid surrogate pair characters (CVE-2022-31116)
Port https://github.com/ultrajson/ultrajson/pull/519 to fix buffer overflows (CVE-2021-4595)
Nothing published for this version
Fix json_loads behaviour for string "-"
write_jsonljson_loads behaviour for string "-"Support custom msgpack serializers and deserializers via function registries (mostly internals).
msgpack serializers and deserializers via function registries (mostly internals).Nothing published for this version
Nothing published for this version
Nothing published for this version
Expose data loaders for spaCy v3.
Remove unsafe YAML implementations, raising if called (see #28)
Include vendorized `ruamel.yaml`.
ruamel.yaml.Nothing published for this version
Nothing published for this version
Nothing published for this version
Support cupy.ndarray in msgpack_numpy.
cupy.ndarray in msgpack_numpy.msgpack raise correct errors for unserializable objects if numpy is not installed.Port over esnme/ultrajson#323 by @timdawborn.
Nothing published for this version
Nothing published for this version
…handling of invalid surrogate pair characters (CVE-2022-31116)
Drop support for Python 3.5 and earlier.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Support cupy.ndarray in msgpack_numpy.
cupy.ndarray in msgpack_numpy.msgpack raise correct errors for unserializable objects if numpy is not installed.Nothing published for this version
Fix issue #10: Update cloudpickle to 1.2.2 to support Python 3.8.
append_new_line option to write_jsonl to control whether newline is appended.Thanks to @nyejon for the pull request!
Add read_gzip_json and write_gzip_json for working with compressed JSON.
read_gzip_json and write_gzip_json for working with compressed JSON.Thanks to @polm for the pull request!
Fix issue #6: Don't write to JSONL as "a" (append) by default.
"a" (append) by default.Thanks to @sadovnychyi for the pull request and to @Adriandorr for the fix.
Thanks to @sadovnychyi for the pull request and to @Adriandorr for the fix.
Fix length limits on msgpack serialization.
msgpack serialization.Move ujson to srsly.ujson for API consistency.
ujson to srsly.ujson for API consistency.pickle_loads and pickle_dumps.msgpack_loads, msgpack_dumps, write_msgpack, read_msgpack.Expose json_dumps and json_loads helpers.
json_dumps and json_loads helpers.sort_keys via json.Fix tests across Python versions and platforms.
Nothing published for this version
One column per quarter.