sccache - Shared Compilation Cache

Sccache is a ccache-like tool. It is used as a compiler wrapper and avoids compilation when possible, storing a cache in a remote storage using the Amazon Simple Cloud Storage Service (S3) API, the Google Cloud Storage (GCS) API, or Redis.

Sccache now includes experimental Rust support.

It works as a client-server. The client spawns a server if one is not running already, and sends the wrapped command line as a request to the server, which then does the work and returns stdout/stderr for the job. The client-server model allows the server to be more efficient in its handling of the remote storage.

Sccache can also be used with local storage instead of remote.

Table of Contents (ToC)

Build Requirements
Build
Installation
Usage
Storage Options
Debugging
Interaction with GNU make jobserver
Known Caveats

Build Requirements

Sccache is a Rust program. Building it requires cargo (and thus rustc). sccache currently requires Rust 1.27.

We recommend you install Rust via Rustup. The generated binaries can be built so that they are very portable. By default sccache supports a local disk cache. To build sccache with support for S3 and/or Redis cache backends, add --features=all or select a specific feature by passing s3, gcs, and/or redis. Refer the Cargo Documentation for details.

Build

$ cargo build [--features=all|redis|s3|gcs] [--release]

Building portable binaries

When building with the gcs feature, sccache will depend on OpenSSL, which can be an annoyance if you want to distribute portable binaries. It is possible to statically link against OpenSSL using the steps below before building with cargo.

Linux

You will need to download and build OpenSSL with -fPIC in order to statically link against it.

./config -fPIC --prefix=/usr/local --openssldir=/usr/local/ssl
make
make install
export OPENSSL_LIB_DIR=/usr/local/lib
export OPENSSL_INCLUDE_DIR=/usr/local/include
export OPENSSL_STATIC=yes

Build with cargo and use ldd to check that the resulting binary does not depend on OpenSSL anymore.

macOS

Just setting the below environment variable will enable static linking.

export OPENSSL_STATIC=yes

Build with cargo and use otool -L to check that the resulting binary does not depend on OpenSSL anymore.

Windows

On Windows it is fairly straight forward to just ship the required libcrpyto and libssl DLLs with sccache.exe, but the binary might also depend on a few MSVC CRT DLLs that are not available on older Windows versions.

It is possible to statically link against the CRT using a .cargo/config file with the following contents.

[target.x86_64-pc-windows-msvc]
rustflags = ["-Ctarget-feature=+crt-static"]

Build with cargo and use dumpbin /dependents to check that the resulting binary does not depend on MSVC CRT DLLs anymore.

In order to statically link against both the CRT and OpenSSL, you will need to either build OpenSSL static libraries (with a statically linked CRT) yourself or get a pre-built distribution that provides these.

Then you can set environment variables which get picked up by the openssl-sys crate.

See the following example for using pre-built libraries from Shining Light Productions, assuming an installation in C:\OpenSSL-Win64:

set OPENSSL_LIB_DIR=C:\OpenSSL-Win64\lib\VC\static
set OPENSSL_INCLUDE_DIR=C:\OpenSSL-Win64\include
set OPENSSL_LIBS=libcrypto64MT:libssl64MT

Installation

$ cargo install

Usage

Running sccache is like running ccache: wrap your compilation commands with it, like so:

$ sccache gcc -o foo.o -c foo.c

or use it with rust, like so:

$ RUSTC_WRAPPER=[path to sccache] cargo build

Sccache (tries to) support gcc, clang and MSVC. If you don't specify otherwise, sccache will use a local disk cache.

You can run sccache --start-server to start the background server process without performing any compilation.

You can run sccache --stop-server to terminate the server. It will terminate after 10 minutes of inactivity.

Running sccache --show-stats will print a summary of cache statistics.

Some notes about using sccache with Jenkins are here.

Storage Options

Sccache defaults to using local disk storage. You can set the SCCACHE_DIR environment variable to change the disk cache location. By default it will use a sensible location for the current platform: ~/.cache/sccache on Linux, %LOCALAPPDATA%\Mozilla\sccache on Windows, and ~/Library/Caches/Mozilla.sccache on MacOS.

If you want to use S3 storage for the sccache cache, you need to set the SCCACHE_BUCKET environment variable to the name of the S3 bucket to use. You can use AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY to set the S3 credentials and if you need to override the default endpoint you can set SCCACHE_ENDPOINT. To connect to a minio storage for example you can set SCCACHE_ENDPOINT=<ip>:<port>.

Set SCCACHE_REDIS to a Redis url in format redis://[:<passwd>@]<hostname>[:port][/<db>] to store the cache in a Redis instance. Redis can be configured as a LRU (least recently used) cache with a fixed maximum cache size. Set maxmemory and maxmemory-policy according to the Redis documentation. The allkeys-lru policy which discards the least recently accessed or modified key fits well for the sccache use case.

Set SCCACHE_MEMCACHED to a Memcached url in format tcp://<hostname>:<port> ... to store the cache in a Memcached instance.

To use Google Cloud Storage, you need to set the SCCACHE_GCS_BUCKET environment variable to the name of the GCS bucket. If you're using authentication, set SCCACHE_GCS_KEY_PATH to the location of your JSON service account credentials. By default, SCCACHE on GCS will be read-only. To change this, set SCCACHE_GCS_RW_MODE to either READ_ONLY or READ_WRITE.

To use Azure Blob Storage, you'll need your Azure connection string and an existing Blob Storage container name. Set the SCCACHE_AZURE_CONNECTION_STRING environment variable to your connection string, and SCCACHE_AZURE_BLOB_CONTAINER to the name of the container to use. Note that sccache will not create the container for you - you'll need to do that yourself.

Important: The environment variables are only taken into account when the server starts, so only on the first run.

Debugging

You can run the server manually in foreground mode by running SCCACHE_START_SERVER=1 SCCACHE_NO_DAEMON=1 sccache, and send logging to stderr by setting the RUST_LOG environment variable, the format of which is described in more detail in the env_logger documentation.

Alternately, you can set the SCCACHE_ERROR_LOG environment variable to a path and set RUST_LOG to get the server process to redirect its logging there (including the output of unhandled panics, since the server sets RUST_BACKTRACE=1 internally).

Interaction with GNU `make` jobserver

Sccache provides support for a GNU make jobserver. When the server is started from a process that provides a jobserver, sccache will use that jobserver and provide it to any processes it spawns. (If you are running sccache from a GNU make recipe, you will need to prefix the command with + to get this behavior.) If the sccache server is started without a jobserver present it will create its own with the number of slots equal to the number of available CPU cores.

This is most useful when using sccache for Rust compilation, as rustc supports using a jobserver for parallel codegen, so this ensures that rustc will not overwhelm the system with codegen tasks. Cargo implements its own jobserver (see the information on NUM_JOBS in the cargo documentation) for rustc to use, so using sccache for Rust compilation in cargo via RUSTC_WRAPPER should do the right thing automatically.

Known caveats

(and possible future improvements)

Sccache doesn't try to be smart about the command line arguments it uses when computing a key for a given compilation result (like skipping preprocessor-specific arguments)
It doesn't support all kinds of compiler flags, and is certainly broken with a few of them. Really only the flags used during Firefox builds have been tested.
It doesn't support ccache's direct mode.
It doesn't support an option like CCACHE_BASEDIR.

SingingTree/sccache