libmizu 0.2.0
Lock-free shared-memory IPC at thread speed in C11
Loading...
Searching...
No Matches
Changelog

All notable changes to libmizu are documented here. The format follows Keep a Changelog. Versions carry an ABI contract: any wire-format change bumps MIZU_ABI_VERSION in include/mizu.h, and the soname tracks MIZU_VERSION_MAJOR.

[Unreleased]

Added

Changed

  • mizu_arrow_col_utf8 (binding-author tier) also concatenates a null-free large_utf8 column, narrowing its offsets, where it declined before.
  • A batch submit wakes at most one idle worker per task: a short batch no longer wakes every idle worker in the pool, and no worker is woken twice for the same batch.

[0.2.0] - 2026-11-12

Added

  • mizu_pool_collect_terminal (binding-author tier): collect every terminal task in a set in one call, never parking.
  • mizu_arrow_open_vector (binding-author tier): open a bare Arrow vector without applying the frame's automatic-row.names row limit.
  • mizu_arrow_col_validity and mizu_arrow_col_utf8 (binding-author tier): concatenate a chunked Arrow column's validity bitmap, and a null-free utf8 column's offsets and bytes, into a binding's own buffers.

Changed

  • Much faster conversions of large nullable Arrow columns.
  • Faster conversions of large int32 and int64 Arrow columns.
  • Faster first reads of large zero-copy views on Linux.
  • Repeated sends of a payload larger than 64 MB no longer create a fresh shared-memory region per send.

Fixed

  • mizu_pool_stop() could wait an extra 50 ms after the last worker exited cleanly.
  • An event-driven pool consumer could miss the wake for a task result published immediately after submission and wait indefinitely.

[0.1.0] - 2026-11-04

Added

First public release.

  • Lock-free SPSC channels and work-stealing task pools over POSIX shared memory (Linux, macOS) and Win32 file mappings, behind a language-agnostic, FFI-safe C ABI: opaque handles, mizu_status terminal states, vectored collects, batch submit, and pool introspection. 64-bit platforms; Linux needs kernel >= 5.3.
  • Crash-atomic wire formats: every cross-process claim is copy-then-CAS, so a peer dying mid-operation cannot wedge a queue, and a per-waiter parker protocol (futex / __ulock / WaitOnAddress) guarantees no lost wakeups.
  • Process-liveness verdicts from a kernel-released exclusive lock per entity, with per-platform death listeners and an orphan reaper reclaiming dead creators' regions.
  • Payload tiers from immediate values to spill regions and zero-copy layout regions (SHM_VEC) under a cross-process refcount, with churn-gated fallback to the copy tiers.
  • A binding seam where the core moves opaque frames and never sees a language object: bindings register stage / read / exec / check / park / sweep / drop callbacks per handle, and staging is transactional. A built-in bytes binding ships as the reference consumer.
  • A three-tier API: the stable mizu.h, the binding-author mizu_ext.h (dual-form header inlines plus exported symbols, pinned per minor release), and private internals.
  • The cross-language wire contract, normative in DESIGN.md and pinned by a byte-exact golden corpus: per-handle identity words with append-only capability bits, the 'I' interchange stream, and the MIZH / MIZS / MIZL zero-copy layout formats.
  • Cross-language pools: a neutral task stream with a bounded error envelope, task arguments by reference, and the morsel-claim map protocol (mizu_morsel_*) for self-scheduling runners.
  • An Arrow C Data Interface → MIZL frame writer with dictionary, temporal and uint64-verbatim forms, fixed-width conversion kernels, an inline small-table form, and per-column writers a binding's own emitters can share.
  • Named, edge-triggered notify endpoints for event-driven bindings, signaled at the publish checkpoint; close and peer death also signal.
  • L'Ecuyer-CMRG (MRG32k3a) stream jumping for deterministic per-element map seeds.
  • Distribution as a four-file amalgamation (mizu.c + mizu.h + mizu_ext.h + mizu_ext.c), make install with pkg-config, and a doxygen API reference.