zccache 1.1.20

A high-performance local compiler cache daemon

zccache

A blazing fast compiler cache for C/C++ and Rust

Inspired by sccache, but optimized for local-first use with aggressive file metadata caching and filesystem watching.

Quick Install

curl -LsSf https://github.com/zackees/zccache/releases/latest/download/install.sh | sh

powershell -ExecutionPolicy Bypass -c "irm https://github.com/zackees/zccache/releases/latest/download/install.ps1 | iex"

Verify:

zccache --version

Performance

50 files per benchmark, median of 5 trials. Run it yourself: ./perf

Cache Hit (warm cache)

Benchmark	Bare Compiler	sccache	zccache	vs sccache	vs bare
C++ single-file	11.705s	1.576s	0.050s	32x	236x
C++ multi-file	11.553s	11.530s	0.017s	695x	696x
C++ response-file (single)	12.540s	1.558s	0.047s	33x	267x
C++ response-file (multi)	12.049s	12.434s	0.019s	669x	648x
Rust build	6.592s	8.604s	0.045s	193x	148x
Rust check	3.716s	5.922s	0.049s	121x	76x

Cache Miss (cold compile)

Benchmark	Bare Compiler	sccache	zccache	vs sccache	vs bare
C++ single-file	12.641s	20.632s	13.430s	1.5x	0.9x
C++ multi-file	11.358s	11.759s	12.867s	0.9x	0.9x
C++ response-file (single)	12.063s	20.607s	14.087s	1.5x	0.9x
C++ response-file (multi)	13.030s	25.303s	13.975s	1.8x	0.9x
Rust build	7.119s	10.023s	8.507s	1.2x	0.8x
Rust check	4.289s	7.056s	5.060s	1.4x	0.8x

Benchmark details

• Single-file = 50 sequential clang++ -c unit.cpp invocations
• Multi-file = one clang++ -c *.cpp invocation (sccache cannot cache these — its "warm" time is a full recompile)
• Response-file = args via nested .rsp files: 200 -D defines + 50 -I paths + 30 warning flags (~283 expanded args)
• Rust build = --emit=dep-info,metadata,link (cargo build)
• Rust check = --emit=dep-info,metadata (cargo check)
• Cold = first compile (empty cache). Warm = median of 5 subsequent runs.
• sccache gets cache hits but each hit still costs ~170ms subprocess overhead. zccache serves hits in ~1ms via in-process IPC.

---

Why is zccache so much faster on warm hits?

The difference comes from architecture, not better caching:

	sccache	zccache
IPC model	Subprocess per invocation (fork + exec + connect)	Persistent daemon, single IPC message per compile
Cache lookup	Client hashes inputs, sends to server, server checks disk	Daemon has inputs in memory (file watcher + metadata cache)
On hit	Server reads artifact from disk, sends back via IPC	Daemon hardlinks cached file to output path (1 syscall)
Multi-file	Compiles every file (no multi-file cache support)	Parallel per-file cache lookups, only misses go to the compiler
Per-hit cost	~170ms (process spawn + hash + disk I/O + IPC)	~1ms (in-memory lookup + hardlink)

Architecture enhancements that make the difference:

• Filesystem watcher — a background notify watcher tracks file changes in real time, so the daemon already knows whether inputs are dirty before you even invoke a compile. No redundant stat/hash work on hit.
• In-memory metadata cache — file sizes, mtimes, and content hashes live in a lock-free DashMap. Cache key computation is a memory lookup, not disk I/O.
• Single-roundtrip IPC — each compile is one length-prefixed bincode message over a Unix socket (or named pipe on Windows). No subprocess spawning, no repeated handshakes.
• Hardlink delivery — cache hits are served by hardlinking the cached artifact to the output path — a single syscall instead of reading + writing the file contents.
• Multi-file fast path — when a build system passes N source files in one invocation, zccache checks all N against the cache in parallel, serves hits immediately, and batches only the misses into a single compiler process.

Broader tool coverage — zccache supports modes that other compiler caches don't:

Mode	Description
Multi-file compilation	clang++ -c a.cpp b.cpp c.cpp — per-file caching with parallel lookups
Response files	Nested .rsp files with hundreds of flags — fully expanded and cached
clang-tidy	Static analysis results cached and replayed
include-what-you-use	IWYU output cached per translation unit
Emscripten (emcc/em++)	WebAssembly compilation cached end-to-end
wasm-ld	WebAssembly linking cached
rustfmt	Formatting results cached
clippy	Lint results cached
Rust check & build	cargo check and cargo build with extern crate content hashing

Install

curl -LsSf https://github.com/zackees/zccache/releases/latest/download/install.sh | sh

powershell -ExecutionPolicy Bypass -c "irm https://github.com/zackees/zccache/releases/latest/download/install.ps1 | iex"

This installs the standalone native Rust binaries (zccache, zccache-daemon, and zccache-fp) directly from GitHub Releases.

Default install locations:

• Linux/macOS user install: ~/.local/bin
• Linux/macOS global install: /usr/local/bin
• Windows user install: %USERPROFILE%\.local\bin
• Windows global install: %ProgramFiles%\zccache\bin

Global install examples:

curl -LsSf https://github.com/zackees/zccache/releases/latest/download/install.sh | sudo sh -s -- --global

powershell -ExecutionPolicy Bypass -c "$env:ZCCACHE_INSTALL_MODE='global'; irm https://github.com/zackees/zccache/releases/latest/download/install.ps1 | iex"

Each GitHub release also publishes standalone per-platform archives:

• Linux: zccache-vX.Y.Z-x86_64-unknown-linux-musl.tar.gz, zccache-vX.Y.Z-aarch64-unknown-linux-musl.tar.gz
• macOS: zccache-vX.Y.Z-x86_64-apple-darwin.tar.gz, zccache-vX.Y.Z-aarch64-apple-darwin.tar.gz
• Windows: zccache-vX.Y.Z-x86_64-pc-windows-msvc.zip, zccache-vX.Y.Z-aarch64-pc-windows-msvc.zip

PyPI remains available if you prefer pip install zccache; those wheels also install the native binaries directly onto your PATH. Pre-built wheels are available for:

Platform	Architecture
Linux	x86_64, aarch64
macOS	x86_64, Apple Silicon
Windows	x86_64

Verify the install:

zccache --version

Rust crates are also published on crates.io. The main installable/runtime crates are:

• zccache-cli
• zccache-daemon
• zccache-core
• zccache-hash
• zccache-protocol
• zccache-fscache
• zccache-artifact

Use it as a drop-in replacement for sccache — just substitute zccache:

Integration Summary

RUSTC_WRAPPER=zccache cargo build
export CC="zccache clang"
export CXX="zccache clang++"

• Rust: set RUSTC_WRAPPER=zccache or add rustc-wrapper = "zccache" to .cargo/config.toml.
• Bash: export RUSTC_WRAPPER, CC, and CXX once in your shell or CI environment.
• Python: pass RUSTC_WRAPPER, CC, and CXX through subprocess env when invoking cargo or clang.
• First commands to check: zccache --version, zccache start, zccache status.

Rust zccache integration

Use zccache as Cargo's compiler wrapper:

# one-off invocation
RUSTC_WRAPPER=zccache cargo build
RUSTC_WRAPPER=zccache cargo check

# optional: start the daemon explicitly
zccache start

Add to .cargo/config.toml for automatic use:

[build]
rustc-wrapper = "zccache"

Recommended project-local config:

[build]
rustc-wrapper = "zccache"

[env]
ZCCACHE_DIR = { value = "/tmp/.zccache", force = false }

Supports --emit=metadata (cargo check), --emit=dep-info,metadata,link (cargo build), extern crate content hashing, and cacheable crate types such as lib, rlib, and staticlib. Proc-macro and binary crates are passed through without caching, matching the usual sccache behavior.

Useful Rust workflow commands:

# inspect status
zccache status

# clear local cache
zccache clear

# validate wrapper is active
RUSTC_WRAPPER=zccache cargo clean
RUSTC_WRAPPER=zccache cargo check
zccache status

Bash integration

For shell-driven builds, export the wrapper once in your session or CI step:

export RUSTC_WRAPPER=zccache
export CC="zccache clang"
export CXX="zccache clang++"

zccache start
cargo build
ninja

If you want this active in interactive shells, add it to ~/.bashrc:

export RUSTC_WRAPPER=zccache
export PATH="$HOME/.local/bin:$PATH"

For per-build stats in Bash:

eval "$(zccache session-start --stats)"
cargo build
zccache session-end "$ZCCACHE_SESSION_ID"

Python integration

Python projects can use zccache when invoking Rust or C/C++ toolchains through subprocess, build backends, or extension-module builds.

import os
import subprocess

env = os.environ.copy()
env["RUSTC_WRAPPER"] = "zccache"
env["CC"] = "zccache clang"
env["CXX"] = "zccache clang++"

subprocess.run(["cargo", "build", "--release"], check=True, env=env)

This is useful for:

• setuptools-rust
• maturin
• scikit-build-core
• custom Python build/test harnesses that shell out to cargo, clang, or clang++

Example with maturin:

RUSTC_WRAPPER=zccache maturin build

Example with Python driving cargo check:

subprocess.run(["cargo", "check"], check=True, env=env)

C/C++ build system integration (ninja, meson, cmake, make)

zccache is a drop-in compiler wrapper. Point your build system's compiler at zccache <real-compiler> and it handles the rest:

# meson native file
[binaries]
c = ['zccache', '/usr/bin/clang']
cpp = ['zccache', '/usr/bin/clang++']

# CMake
set(CMAKE_C_COMPILER_LAUNCHER zccache)
set(CMAKE_CXX_COMPILER_LAUNCHER zccache)

The first build (cold cache) runs at near-bare speed. Subsequent rebuilds (ninja -t clean && ninja, or touching source files) serve cached artifacts via hardlinks in under a second.

Single-roundtrip IPC: In drop-in mode, zccache sends a single CompileEphemeral message that combines session creation, compilation, and session teardown — eliminating 2 of 3 IPC roundtrips per invocation.

Session stats: Track hit rates per-build with --stats:

eval $(zccache session-start --stats --log build.log)
export ZCCACHE_SESSION_ID=...
# ... build runs ...
zccache session-stats $ZCCACHE_SESSION_ID   # query mid-build
zccache session-end $ZCCACHE_SESSION_ID     # final stats

Persistent cache: Artifacts are stored in ~/.zccache/artifacts/ and survive daemon restarts. No need to re-warm the cache after a reboot.

Compile journal (build replay): Every compile and link command is recorded to ~/.zccache/logs/compile_journal.jsonl as a JSONL file with enough detail to replay the entire build:

{"ts":"2026-03-17T10:30:00.123Z","outcome":"hit","compiler":"/usr/bin/clang++","args":["-c","foo.cpp","-o","foo.o"],"cwd":"/project/build","env":[["CC","clang"]],"exit_code":0,"session_id":"uuid","latency_ns":1234567}

Fields: ts (ISO 8601 UTC), outcome (hit/miss/error/link_hit/link_miss), compiler (full path), args (full argument list), cwd, env (omitted when inheriting daemon env), exit_code, session_id (null for ephemeral), latency_ns (wall-clock nanoseconds). One JSON object per line — pipe through jq to filter, or replay builds by extracting compiler + args + cwd.

Per-session compile journal: Pass --journal <path> to session-start to write a dedicated JSONL log containing only the commands from that session. The path must end in .jsonl:

result=$(zccache session-start --journal build.jsonl)
session_id=$(echo "$result" | jq -r .session_id)
export ZCCACHE_SESSION_ID=$session_id

# ... build runs ...

# Inspect this session's commands only (no noise from other sessions)
jq . build.jsonl

zccache session-end $session_id

The session journal uses the same JSONL schema as the global journal. Entries are written to both the global and session journals simultaneously. The session file handle is released when session-end is called.

Multi-file compilation (fast path)

When a build system passes multiple source files to a single compiler invocation (e.g. gcc -c a.cpp b.cpp c.cpp -o ...), zccache treats this as a fast path:

1. Each source file is checked against the cache in parallel.
2. Cache hits are served immediately — their .o files are written from the cache.
3. Remaining cache misses are batched into a single compiler process, preserving the compiler's own process-reuse and memory-sharing benefits.
4. The outputs of the batched compilation are cached individually for future hits.

This hybrid approach means the first build populates the cache per-file, and subsequent builds serve as many files as possible from cache while still letting the compiler handle misses efficiently in bulk.

Recommendation: Configure your build system to pass multiple source files per compiler invocation whenever possible. This gives zccache the best opportunity to parallelize cache lookups and minimize compiler launches.

Concurrency

The daemon uses lock-free concurrent data structures (DashMap) for artifact and metadata lookups, so parallel compilation requests from multiple build workers never serialize on a global lock.

Status

Early development — architecture and scaffolding phase.

Goals

• Extremely fast on local machines (daemon keeps caches warm)
• Portable across Linux, macOS, and Windows
• Correct under heavy parallel compilation (no stale cache hits)
• Simple deployment (single binary)

Tool Compatibility

zccache works as a drop-in wrapper for these compilers and tools:

• Clang Toolchain: clang, clang-tidy, IWYU
• Emscripten / WebAssembly: emcc, wasm-ld
• Rust Toolchain: rustc, rustfmt, clippy

Architecture

See docs/ARCHITECTURE.md for the full system design.

Key components

Crate	Purpose
zccache-cli	Command-line interface (zccache binary)
zccache-daemon	Daemon process (IPC server, orchestration)
zccache-core	Shared types, errors, config, path utilities
zccache-protocol	IPC message types and serialization
zccache-ipc	Transport layer (Unix sockets / named pipes)
zccache-hash	blake3 hashing and cache key computation
zccache-fscache	In-memory file metadata cache
zccache-artifact	Disk-backed artifact store with redb index
zccache-watcher	File watcher subsystem: daemon notify pipeline plus Rust-backed Python watcher bindings
zccache-compiler	Compiler detection and argument parsing
zccache-test-support	Test utilities and fixtures

Building

cargo build --workspace

Testing

cargo test --workspace

Documentation

• Architecture
• Design Decisions
• Roadmap

Watcher APIs

zccache exposes watcher-related APIs in three different places, depending on how you want to consume change detection:

• CLI: zccache fp ... for daemon-backed fingerprint checks in scripts and CI
• Python: zccache.watcher for cross-platform library-style file watching
• Rust: zccache-watcher for the daemon-facing watcher pipeline primitives

CLI API

The CLI watcher entrypoint is the fingerprint API. It answers "should I rerun?" by consulting the daemon's in-memory watch state and cached file fingerprints.

zccache fp --cache-file .cache/headers.json check \
  --root . \
  --include '**/*.cpp' \
  --include '**/*.h' \
  --exclude build \
  --exclude .git

Exit codes:

• 0: files changed, run the expensive step
• 1: no changes detected, skip the step

After a successful or failed run, update the daemon's watch state:

zccache fp --cache-file .cache/headers.json mark-success
zccache fp --cache-file .cache/headers.json mark-failure
zccache fp --cache-file .cache/headers.json invalidate

The fingerprint API is the best fit for shell scripts, CI jobs, and build steps that only need a yes/no change answer rather than a stream of file events.

Python API

The repository also exposes zccache.watcher for robust cross-platform file watching from Python. The public API is polling- and callback-friendly, while the backend runs the filesystem scan loop in Rust and only crosses into Python when delivering events.

from zccache.watcher import watch_files

watcher = watch_files(
    ".",
    include_folders=["src", "include"],
    include_globs=["src/**/*.cpp", "include/**/*.h"],
    excluded_patterns=["build", "dist/**", ".git"],
    debounce_seconds=0.2,
    poll_interval=0.1,
)

event = watcher.poll(timeout=1.0)
if event is not None:
    print(event.paths)

watcher.stop()

For explicit lifecycle control, use the class API:

from zccache.watcher import FileWatcher

watcher = FileWatcher(".", include_globs=["**/*.cpp"], autostart=False)
watcher.start()
event = watcher.poll(timeout=1.0)
watcher.stop()
watcher.resume()
watcher.stop()

Python watcher features:

• include_folders to narrow the scan roots
• include_globs to include only matching files
• excluded_patterns to skip directories or files
• debounce_seconds to coalesce bursts of edits
• optional notification_predicate applied at Python delivery time
• callback API plus polling API
• explicit start(), stop(), resume(), and context-manager support

Compatibility wrappers used by fastled-wasm are also available:

• FileWatcherProcess
• DebouncedFileWatcherProcess
• watch_files
• FileWatcher

See crates/zccache-watcher/README.md for the full Python watcher surface.

Rust API

For Rust consumers, the public watcher crate is zccache-watcher. It now exposes both the daemon-facing watcher pipeline and a library-style polling watcher API:

• PollingWatcherConfig

• PollingWatcher

• PollWatchBatch

• PollWatchObserver

• IgnoreFilter for directory-name-based filtering

• NotifyWatcher for notify-backed OS watch registration

• SettleBuffer and SettledEvent for burst coalescing

• OverflowRecovery for overflow-driven rescan scheduling

• WatchEvent and WatcherConfig for event/config plumbing

Example:

use std::time::Duration;
use zccache_watcher::{PollingWatcher, PollingWatcherConfig};

let mut config = PollingWatcherConfig::new(".");
config.include_globs = vec!["**/*.cpp".to_string()];
config.poll_interval = Duration::from_millis(50);
config.debounce = Duration::from_millis(50);

let watcher = PollingWatcher::new(config)?;
watcher.start()?;
let batch = watcher.poll_timeout(Duration::from_secs(1))?;
watcher.stop()?;

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.