v1.4.1 — ParallelZip is now @CacheableTask

Highlights

• ParallelZip is now @CacheableTask. Gradle's own Zip/AbstractArchiveTask are @DisableCachingByDefault upstream — Zip's own stated reason is "Not worth caching," a judgment call from when archive packaging was assumed cheap. ParallelZip previously inherited that same default rather than choosing it deliberately.
• With org.gradle.caching=true (or --build-cache), an unchanged rebuild now restores the archive straight from the build cache, skipping both the copy-walk and compression entirely — the single biggest available speedup for repeat builds, bigger than anything the compression pipeline itself can offer.
• This is safe specifically because the task already produces byte-for-byte reproducible output for identical inputs, and already separates @Input properties (store, level, skipAlreadyCompressed — affect the archive bytes) from @Internal ones (threads — doesn't).
• Verified with a real functional test: runs the task, wipes all local build state (not just the output file), reruns, and asserts the second run's outcome is genuinely FROM_CACHE with a still-valid archive — not just an annotation check.

Published to the Gradle Plugin Portal.

v1.4.0 — skip re-deflating already-compressed content, mmap fixes, OOM guard

Highlights

• Magic-byte already-compressed detection — entries are recognized by file signature (jars/zips, gzip, bzip2, xz, 7-zip, zstd, rar, lz4, png, jpeg, gif, webp, mp4/mov, webm/mkv, ogg, woff/woff2) and STORED immediately instead of attempting DEFLATE. Geometric-mean speedup across nine real production Zip tasks jumps from ~5.4× (1.3.0) to ~26.5×.
• New skipAlreadyCompressed task property (default true) — the speed above trades real archive size for it (6–17% larger across the in-build benchmark suite). Set skipAlreadyCompressed = false to always attempt DEFLATE instead, for 1.3.x-like size when that trade isn't worth it for a given archive.
• Fixed a Windows-only mmap file-lock bug — an unreleased MappedByteBuffer on a large entry's source file kept it locked against deletion/rewrite until GC, which could break a later task reusing the same file in the same Gradle daemon.
• OOM guard for small Gradle daemon heaps — the mmap-path buffer budget now scales down with heap size instead of a fixed 128 MiB ceiling, and a one-time warning fires when the daemon's max heap is under 2048 MiB.

Benchmarks

Full refresh across all 9 in-build projects and all 11 fixed-corpus projects — see docs/BENCHMARKS.md.

Published to the Gradle Plugin Portal.

v1.3.0 — native libdeflate on all six platforms, fused CRC32, verified benchmarks

Highlights

• Native libdeflate accelerator on all six bundled platforms — linux/windows/macOS × x64/arm64. A single published jar carries every platform's native; unsupported platforms fall back to the pure-Java Deflater automatically.
• Fused CRC32 into compression and mmap for large entries in the native path, cutting a redundant pass over entry data.
• Verified in-build benchmark framework — swaps each sample project's real Zip task for ParallelZip and takes isolated, repeated measurements.

CI / build

• Fixed the macos-x64 leg after the macos-13 runner retirement (Dec 2025).
• windows-arm64 leg now uses Microsoft's OpenJDK 17 (Temurin ships no arm64 JDK 17).
• Fixed the native build on non-MSVC (MinGW) Windows toolchains.

Published to the Gradle Plugin Portal.

v1.2.0 — native acceleration on every major architecture

Native acceleration now covers every major platform/arch

Previously the native libdeflate accelerator only ran on linux-x64 and windows-x64 —
everywhere else (Apple Silicon, ARM Windows, Linux ARM, Intel Macs) silently fell back to
the pure-Java Deflater. This release adds native builds for the four remaining
architectures, so all six are now covered:

• linux-x64, linux-arm64
• windows-x64, windows-arm64
• macos-arm64, macos-x64

CI now builds, tests, and bundles a native library per architecture, and asserts each one
is present and actually exercised before the release jar is assembled — a regression on
any single platform fails the build instead of silently falling back.

Compatibility

• No breaking changes — plugin ID, Maven coordinates, and usage are unchanged from 1.1.0.
• Any other platform/arch (e.g. linux-riscv64) continues to fall back to the pure-Java
  path automatically, exactly as before.

See the README for the
full supported-platform list.

v1.1.0 — faster codec path, new plugin coordinates

Breaking: plugin coordinates moved

The Maven group and the Gradle plugin ID moved from io.github.kukis13 to
com.ljarocki:

plugins {
    id 'com.ljarocki.parallel-zip' version '1.1.0'   // was io.github.kukis13.parallel-zip
}

tasks.register('dist', com.ljarocki.parallelzip.ParallelZip) {   // was io.github.kukis13.parallelzip.ParallelZip
    from 'build/staging'
    archiveFileName = 'dist.zip'
    store = false   // true = STORE everything (fastest, ~larger)
    threads = 12    // default: available processors
}

Existing users need to update the id and any direct ParallelZip/ParallelZipPlugin
class references. Nothing else about usage changes.

What's faster

• New incompressibility sniff: probes large entries before compressing and STOREs
  them instead if they clearly won't shrink (already-compressed jars/binaries/media) —
  the single biggest win in this release, especially on jar-heavy archives.
• Per-thread native libdeflate handle reuse + batched JNI calls, replacing
  alloc-per-call, cutting overhead on archives with many small entries.
• Small entries batched into fewer compression tasks, and unfiltered entries read
  lazily on a worker thread instead of eagerly on Gradle's single-threaded copy walk
  (the bigger of the two).

Net effect across the 11 projects in BENCHMARKS.md: parallel-zip
DEFLATE (libdeflate) is now 1.65×–9.67× faster than Gradle's Zip task (up from
1.64×–4.78× before this release), with no regressions on any corpus. DEFLATE now trades
up to ~0.8% archive size for that speed on a few corpora — still the safe default.

Compatibility

• Verified Gradle 9 support (fixed Gradle 9 deprecations) alongside Gradle 8.
• Declared Configuration Cache compatibility.
• Dropped the foojay-resolver plugin: the build now fails loudly if JDK 17 isn't
  already installed, instead of silently auto-downloading one.
• Fixed the native accelerator build on non-MSVC (MinGW) Windows toolchains.

Benchmarks

Full refreshed 11-project table and methodology in
BENCHMARKS.md.

v1.0.0 — first release

A multithreaded, reproducible, drop-in replacement for Gradle's Zip task.

What it does

• Drop-in for Zip — extends AbstractArchiveTask, so the full CopySpec DSL
  (from/into/include/exclude/rename/filter/duplicatesStrategy) works
  unchanged. Swapping type: Zip for type: ParallelZip is a one-word change.
• Parallel compression across all cores, or STORE (skip DEFLATE entirely) for
  archives that are already jar/binary-heavy.
• Byte-for-byte reproducible archives — fixed write order means output is stable
  across thread counts and runs, sidestepping the reproducibility blocker that kept
  gradle/gradle#2774 from shipping.
• ZIP64 support for archives beyond 4 GiB or 65,535 entries.
• Streaming for arbitrarily large individual entries (tested past 2 GiB), with
  memory bounded to a fraction of the heap.
• Native libdeflate acceleration on linux-x64/windows-x64 — measured 2.0–3.7×
  faster than the JDK's Deflater on real archives, with a safe pure-Java fallback on
  every other platform/arch or if the native build ever fails to load.
• Zero runtime dependencies.

Proof, not just claims

Benchmarked against the official binary distributions of 11 popular open-source Java
projects (ZooKeeper, Cassandra, Kafka, Groovy, Solr, HBase, Spark, Gradle, Flink,
SonarQube, Hadoop) — full numbers and methodology in
BENCHMARKS.md.

Usage

plugins {
    id 'io.github.kukis13.parallel-zip' version '1.0.0'
}

tasks.register('dist', io.github.kukis13.parallelzip.ParallelZip) {
    from 'build/staging'
    archiveFileName = 'dist.zip'
    store = false   // true = STORE everything (fastest, ~larger)
    threads = 12    // default: available processors
}

See the README for full docs.

Also published to the Gradle Plugin Portal (pending first-time approval).