Skip to content

Commit 9d43df2

Browse files
Dev-iLclaude
andcommitted
Update the stable ABI report and changelog for the final candidate
Rederive every figure in the benchmark report from the re-timed data: the control spread, the operation table, the 19 cases beyond the control maximum, the 16 MiB + 137 summary, the absence of faster cases, and the repair check. Disclose the _ZlibDecompressor copies above a 16 MiB limit and for output shorter than its limit, discount the repeated Python input in the strategy study in both directions, and restate the end-of-stream timing evidence with its one outlier. The changelog adds the decompressobj() input beyond 4 GiB and the read-only readinto() buffer fixes. build.json loses slashes in its prose strings and records that no CI has run for this branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1 parent 85fb4a0 commit 9d43df2

3 files changed

Lines changed: 62 additions & 52 deletions

File tree

‎CHANGELOG.rst‎

Lines changed: 14 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -17,12 +17,16 @@ version 1.1.1-dev
1717
``zlib_ng.compress()``, ``zlib_ng.decompress()``, ``compressobj()`` output,
1818
and ``decompressobj().decompress()`` and ``_ZlibDecompressor.decompress()``
1919
without ``max_length``. In the project's benchmark, decompressing 16 MiB took
20-
1.05 to 1.20 times as long and the worst case, 1 MiB of zeros, 2.86 times as
20+
up to 19% longer, and the worst case, 1 MiB of zeros, took 2.74 times as
2121
long; compression times did not change measurably. ``gzip_ng`` file reads
22-
and ``_ZlibDecompressor.decompress(data, max_length)`` avoid the copy.
23-
``decompressobj().decompress(data, max_length)`` with a ``max_length`` above
24-
16 KiB also copies, and briefly holds up to twice ``max_length``. PyPy and
25-
free-threaded CPython are not affected.
22+
avoid the copy, and so does ``_ZlibDecompressor.decompress(data,
23+
max_length)`` when the output fills a ``max_length`` of at most 16 MiB. A
24+
bounded call that returns less than its limit copies the shorter result
25+
while still holding the full allocation. With a larger limit, the copy is
26+
also made by ``_ZlibDecompressor.decompress(data, max_length)`` above 16 MiB
27+
and by ``decompressobj().decompress(data, max_length)`` above 16 KiB, which
28+
briefly holds up to twice ``max_length``. PyPy and free-threaded CPython are
29+
not affected.
2630
+ Compression or decompression whose output ends exactly at the size of its
2731
current output buffer no longer grows that buffer once more before
2832
returning.
@@ -35,6 +39,11 @@ version 1.1.1-dev
3539
+ ``_ZlibDecompressor.decompress()`` no longer loses input when given more than
3640
4 GiB of data and a ``max_length`` that stops it early; the bytes are kept
3741
for the next call or in ``unused_data``.
42+
+ ``decompressobj().decompress()`` no longer loses unread input beyond the
43+
first 4 GiB of one call when the output limit or the end of the stream is
44+
reached; it is kept in ``unconsumed_tail`` or ``unused_data``.
45+
+ The internal ``_GzipReader.readinto()`` raises ``TypeError`` for a read-only
46+
buffer instead of writing into it, which crashed on a read-only ``mmap``.
3847
+ Reading with ``gzip_ng.open()`` from a file object whose ``read`` attribute
3948
lookup raises an error other than ``AttributeError`` now raises that error
4049
instead of an unrelated ``TypeError``.

0 commit comments

Comments
 (0)