isal 1.9.0

version 1.9.0

on Python PyPI

By far most of the work on this release has been done by
@ptweezy. Almost all bug fixes and improvements
are their work.

Major changes:

• Updated included ISA-L library to 2.32.1.
• Python 3.10 and PyPy 3.10 are no longer supported. Added support for PyPy
  3.12 and Python 3.15.
• Restore PyPy wheel builds.
• Fix builds on macOS with Apple Silicon, which used the portable C code of
  ISA-L rather than its optimized aarch64 assembly. On an Apple M4,
  decompression is 1.6 to 3 times faster, CRC-32 is 50 times faster, and
  Adler-32 is 7 times faster.

Optimizations and features:

• isal_zlib.compressobj supports strategy=Z_FIXED at compression
  level 0. Level 0 then uses fixed Huffman codes instead of starting each
  block with a Huffman table of about 110 bytes. This shrinks small flushed
  writes, such as websocket messages: JSON messages of about 440 bytes, each
  followed by flush(Z_SYNC_FLUSH), compress to an average of 75 instead
  of 188 bytes. Blocks of a few KiB or more can be larger: about 5% for JSON
  and 10% for Python source code at 64 KiB. Other strategies, and Z_FIXED
  at levels 1 to 3, still give a warning and use the default strategy.
• Speed up import isal.igzip and import isal.igzip_threaded by
  importing argparse and shutil only when the command line interface
  runs, and by dropping the multiprocessing import.
• Speed up reading lines from binary igzip files through iteration,
  readlines(), and readline(). On an Apple M4, reading FASTQ data is
  1.13 times faster, and reading lines of a few bytes is 1.5 times faster.
• Speed up each write() call to binary igzip files on CPython. On an
  AMD Ryzen 7 9800X3D with Python 3.14.8, writing FASTQ data or Python source
  code one line at a time is 1.1 times faster.
• Speed up many small writes to igzip.IGzipFile on Python 3.12 and later.
  IGzipFile collects these writes in the write buffer of
  gzip.GzipFile and compresses them in large chunks, which makes writing
  lines 1.4 times faster.

Bug fixes:

• Fix zlib streams compressed with a preset dictionary so they can be decoded
  by standard zlib, including builds linked against older ISA-L releases.
• Fix backward seeking in igzip files by discarding buffered compressed data
  when rewinding the underlying file.
• Fix a crash when igzip.decompress gets an object that is not
  bytes-like, such as a str. It raises a TypeError instead.
• Fix a crash after creating igzip_lib.IgzipDecompressor objects with a
  zdict.
• Fix error handling when more than one interpreter imports
  isal.isal_zlib or isal.igzip_lib, for example, the main interpreter
  and a legacy subinterpreter that Py_NewInterpreter() creates. Each
  import replaced the error classes that the C code raises, so
  except isal_zlib.error, except igzip_lib.error, and
  except gzip.BadGzipFile caught errors only in the interpreter that
  imported the modules last.
• Fix a hang in isal_zlib.Compress: after flush() raised an error for
  an unsupported mode, the next call on the same object never returned.
• Fix a crash when you call flush(Z_FINISH) twice on a level 0
  isal_zlib.Compress object.
• Fix unused_data of isal_zlib.Decompress and
  igzip_lib.IgzipDecompressor, which could contain bytes from outside the
  input buffer when max_length split the end of a stream. At the end of a
  stream, isal_zlib.Decompress.unconsumed_tail is empty, matching the
  CPython fix for gh-158169.
• Fix igzip_lib.IgzipDecompressor.needs_input, which could be True
  after decompress() returned max_length bytes while more output was
  still available. Readers that stop at the end of the file when
  needs_input is True, such as CPython's DecompressReader, raised
  EOFError for valid data.
• Fix isal_zlib.decompressobj with a wbits of 40 to 47, which detects
  zlib or gzip data. It failed on gzip data when the first decompress()
  call got fewer than 2 bytes, for example, when the data arrived byte by
  byte.
• Fix isal_zlib.decompressobj with a wbits of 40 to 47 after a
  flush() call that came before any data. It then failed on both zlib and
  gzip data.
• igzip_threaded.open raises ValueError for a block_size of about
  3.6 GiB or more and TypeError for a compresslevel that is not an
  integer. Both raised SystemError before.
• Fix igzip_threaded writers at compression level 0, which raised
  OverflowError for data that does not compress, such as random bytes or
  already compressed files. At compression level 0, block_size must now be
  less than about 2 GiB.
• Extraordinarily large gzip headers with either a very long comment or
  filename will now cause an OverflowError rather than use more and more
  memory. The size is limited to the internal buffer size which is currently
  set at 512 KiB.
• Fix igzip_threaded readers and writers that waited for a polling timeout
  of up to 50 ms at the end of the stream or when closed. Writing and closing
  a small file takes less than a millisecond instead of 60 ms.
• Fix igzip_threaded writers that waited forever when writing to the
  underlying file failed, for example, because the disk was full. A later call
  to write(), flush(), or close() raises the error. After an error,
  close() no longer closes a file object that you passed to
  igzip_threaded.open.
• Fix igzip_threaded writers with more than one thread that waited
  forever when compressing a block failed, for example, because there was not
  enough memory. A later call to write(), flush(), or close()
  raises the error.
• Fix a crash on the free-threaded build when one thread reads
  unused_data or unconsumed_tail of an isal_zlib decompressor
  while another thread decompresses with it. Reading the attributes of
  isal_zlib decompressors and igzip_lib.IgzipDecompressor from other
  threads is now safe, and so is calling decompress() from several
  threads at the same time when wbits detects the format.
• isal_zlib.crc32_combine and the read() and seek() methods of the
  private isal_zlib._GzipReader class raise TypeError instead of
  SystemError for arguments of the wrong type or number.
  igzip.IGzipFile was not affected.
• Fix a crash on PyPy when you create an isal_zlib.Compress or
  isal_zlib.Decompress object by calling the type, for example,
  isal_zlib.Compress(), and then use it. Calling the type raises
  TypeError now, as it does on CPython. Use compressobj() and
  decompressobj() instead.
• Fix readinto() of the private isal_zlib._GzipReader class, which
  wrote into read-only objects such as bytes. It raises TypeError now.
  igzip.IGzipFile was not affected.
• Fix the mtime attribute of igzip.IGzipFile objects in read mode,
  which was None for gzip files with a modification time of 0. It is 0 now,
  as in the gzip module. gzip -n, python -m isal.igzip -n, and
  igzip_threaded writers make such files, and so does gzip.compress by
  default on Python 3.14.
• Fix a crash when there is not enough memory to create an
  igzip_lib.IgzipDecompressor. It raises MemoryError now.
• Fix a crash when there is not enough memory to raise BadGzipFile for
  data that is not in the gzip format, for example, in igzip.decompress.
  It raises MemoryError now.
• Fix a crash in igzip_lib.IgzipDecompressor.decompress() after a call on
  the same object raised MemoryError. The next call used the input of the
  failed call after the caller had released it. This matches the CPython fix
  for gh-148395.
• Fix a crash on macOS and PyPy when there is not enough memory to create the
  internal lock of an isal_zlib.Compress, isal_zlib.Decompress, or
  igzip_lib.IgzipDecompressor object. It raises MemoryError now.
• Fix the -0 and --fast options of python -m isal.igzip, which
  compressed at level 2 instead of level 0.
• Fix crashes and wrong data when one thread calls seek() of the
  private isal_zlib._GzipReader class while another thread reads from the
  same object. igzip.IGzipFile was not affected.