Remote Caching

Report an issue View source Nightly · 7.4 . 7.3 · 7.2 · 7.1 · 7.0 · 6.5

This page covers remote caching, setting up a server to host the cache, and running builds using the remote cache.

A remote cache is used by a team of developers and/or a continuous integration (CI) system to share build outputs. If your build is reproducible, the outputs from one machine can be safely reused on another machine, which can make builds significantly faster.

Overview

Bazel breaks a build into discrete steps, which are called actions. Each action has inputs, output names, a command line, and environment variables. Required inputs and expected outputs are declared explicitly for each action.

You can set up a server to be a remote cache for build outputs, which are these action outputs. These outputs consist of a list of output file names and the hashes of their contents. With a remote cache, you can reuse build outputs from another user's build rather than building each new output locally.

To use remote caching:

  • Set up a server as the cache's backend
  • Configure the Bazel build to use the remote cache
  • Use Bazel version 0.10.0 or later

The remote cache stores two types of data:

  • The action cache, which is a map of action hashes to action result metadata.
  • A content-addressable store (CAS) of output files.

Note that the remote cache additionally stores the stdout and stderr for every action. Inspecting the stdout/stderr of Bazel thus is not a good signal for estimating cache hits.

How a build uses remote caching

Once a server is set up as the remote cache, you use the cache in multiple ways:

  • Read and write to the remote cache
  • Read and/or write to the remote cache except for specific targets
  • Only read from the remote cache
  • Not use the remote cache at all

When you run a Bazel build that can read and write to the remote cache, the build follows these steps:

  1. Bazel creates the graph of targets that need to be built, and then creates a list of required actions. Each of these actions has declared inputs and output filenames.
  2. Bazel checks your local machine for existing build outputs and reuses any that it finds.
  3. Bazel checks the cache for existing build outputs. If the output is found, Bazel retrieves the output. This is a cache hit.
  4. For required actions where the outputs were not found, Bazel executes the actions locally and creates the required build outputs.
  5. New build outputs are uploaded to the remote cache.

Setting up a server as the cache's backend

You need to set up a server to act as the cache's backend. A HTTP/1.1 server can treat Bazel's data as opaque bytes and so many existing servers can be used as a remote caching backend. Bazel's HTTP Caching Protocol is what supports remote caching.

You are responsible for choosing, setting up, and maintaining the backend server that will store the cached outputs. When choosing a server, consider:

  • Networking speed. For example, if your team is in the same office, you may want to run your own local server.
  • Security. The remote cache will have your binaries and so needs to be secure.
  • Ease of management. For example, Google Cloud Storage is a fully managed service.

There are many backends that can be used for a remote cache. Some options include:

nginx

nginx is an open source web server. With its [WebDAV module], it can be used as a remote cache for Bazel. On Debian and Ubuntu you can install the nginx-extras package. On macOS nginx is available via Homebrew:

brew tap denji/nginx
brew install nginx-full --with-webdav

Below is an example configuration for nginx. Note that you will need to change /path/to/cache/dir to a valid directory where nginx has permission to write and read. You may need to change client_max_body_size option to a larger value if you have larger output files. The server will require other configuration such as authentication.

Example configuration for server section in nginx.conf:

location /cache/ {
  # The path to the directory where nginx should store the cache contents.
  root /path/to/cache/dir;
  # Allow PUT
  dav_methods PUT;
  # Allow nginx to create the /ac and /cas subdirectories.
  create_full_put_path on;
  # The maximum size of a single file.
  client_max_body_size 1G;
  allow all;
}

bazel-remote

bazel-remote is an open source remote build cache that you can use on your infrastructure. It has been successfully used in production at several companies since early 2018. Note that the Bazel project does not provide technical support for bazel-remote.

This cache stores contents on disk and also provides garbage collection to enforce an upper storage limit and clean unused artifacts. The cache is available as a [docker image] and its code is available on GitHub. Both the REST and gRPC remote cache APIs are supported.

Refer to the GitHub page for instructions on how to use it.

Google Cloud Storage

[Google Cloud Storage] is a fully managed object store which provides an HTTP API that is compatible with Bazel's remote caching protocol. It requires that you have a Google Cloud account with billing enabled.

To use Cloud Storage as the cache:

  1. Create a storage bucket. Ensure that you select a bucket location that's closest to you, as network bandwidth is important for the remote cache.

  2. Create a service account for Bazel to authenticate to Cloud Storage. See Creating a service account.

  3. Generate a secret JSON key and then pass it to Bazel for authentication. Store the key securely, as anyone with the key can read and write arbitrary data to/from your GCS bucket.

  4. Connect to Cloud Storage by adding the following flags to your Bazel command:

    • Pass the following URL to Bazel by using the flag: --remote_cache=https://storage.googleapis.com/bucket-name where bucket-name is the name of your storage bucket.
    • Pass the authentication key using the flag: --google_credentials=/path/to/your/secret-key.json, or --google_default_credentials to use Application Authentication.

  5. You can configure Cloud Storage to automatically delete old files. To do so, see Managing Object Lifecycles.

Other servers

You can set up any HTTP/1.1 server that supports PUT and GET as the cache's backend. Users have reported success with caching backends such as Hazelcast, Apache httpd, and AWS S3.

Authentication

As of version 0.11.0 support for HTTP Basic Authentication was added to Bazel. You can pass a username and password to Bazel via the remote cache URL. The syntax is https://username:password@hostname.com:port/path. Note that HTTP Basic Authentication transmits username and password in plaintext over the network and it's thus critical to always use it with HTTPS.

HTTP caching protocol

Bazel supports remote caching via HTTP/1.1. The protocol is conceptually simple: Binary data (BLOB) is uploaded via PUT requests and downloaded via GET requests. Action result metadata is stored under the path /ac/ and output files are stored under the path /cas/.

For example, consider a remote cache running under http://localhost:8080/cache. A Bazel request to download action result metadata for an action with the SHA256 hash 01ba4719... will look as follows:

GET /cache/ac/01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b HTTP/1.1
Host: localhost:8080
Accept: */*
Connection: Keep-Alive

A Bazel request to upload an output file with the SHA256 hash 15e2b0d3... to the CAS will look as follows:

PUT /cache/cas/15e2b0d3c33891ebb0f1ef609ec419420c20e320ce94c65fbc8c3312448eb225 HTTP/1.1
Host: localhost:8080
Accept: */*
Content-Length: 9
Connection: Keep-Alive

0x310x320x330x340x350x360x370x380x39

Run Bazel using the remote cache

Once a server is set up as the remote cache, to use the remote cache you need to add flags to your Bazel command. See list of configurations and their flags below.

You may also need configure authentication, which is specific to your chosen server.

You may want to add these flags in a .bazelrc file so that you don't need to specify them every time you run Bazel. Depending on your project and team dynamics, you can add flags to a .bazelrc file that is:

  • On your local machine
  • In your project's workspace, shared with the team
  • On the CI system

Read from and write to the remote cache

Take care in who has the ability to write to the remote cache. You may want only your CI system to be able to write to the remote cache.

Use the following flag to read from and write to the remote cache:

build --remote_cache=http://your.host:port

Besides HTTP, the following protocols are also supported: HTTPS, grpc, grpcs.

Use the following flag in addition to the one above to only read from the remote cache:

build --remote_upload_local_results=false

Exclude specific targets from using the remote cache

To exclude specific targets from using the remote cache, tag the target with no-remote-cache. For example:

java_library(
    name = "target",
    tags = ["no-remote-cache"],
)

Delete content from the remote cache

Deleting content from the remote cache is part of managing your server. How you delete content from the remote cache depends on the server you have set up as the cache. When deleting outputs, either delete the entire cache, or delete old outputs.

The cached outputs are stored as a set of names and hashes. When deleting content, there's no way to distinguish which output belongs to a specific build.

You may want to delete content from the cache to:

  • Create a clean cache after a cache was poisoned
  • Reduce the amount of storage used by deleting old outputs

Unix sockets

The remote HTTP cache supports connecting over unix domain sockets. The behavior is similar to curl's --unix-socket flag. Use the following to configure unix domain socket:

   build --remote_cache=http://your.host:port
   build --remote_cache_proxy=unix:/path/to/socket

This feature is unsupported on Windows.

Disk cache

Bazel can use a directory on the file system as a remote cache. This is useful for sharing build artifacts when switching branches and/or working on multiple workspaces of the same project, such as multiple checkouts. Enable the disk cache as follows:

build --disk_cache=path/to/build/cache

You can pass a user-specific path to the --disk_cache flag using the ~ alias (Bazel will substitute the current user's home directory). This comes in handy when enabling the disk cache for all developers of a project via the project's checked in .bazelrc file.

Garbage collection

Starting with Bazel 7.4, you can use --experimental_disk_cache_gc_max_size and --experimental_disk_cache_gc_max_age to set a maximum size for the disk cache or for the age of individual cache entries. Bazel will automatically garbage collect the disk cache while idling between builds; the idle timer can be set with --experimental_disk_cache_gc_idle_delay (defaulting to 5 minutes).

As an alternative to automatic garbage collection, we also provide a tool to run a garbage collection on demand.

Known issues

Input file modification during a build

When an input file is modified during a build, Bazel might upload invalid results to the remote cache. You can enable a change detection with the --experimental_guard_against_concurrent_changes flag. There are no known issues and it will be enabled by default in a future release. See [issue #3360] for updates. Generally, avoid modifying source files during a build.

Environment variables leaking into an action

An action definition contains environment variables. This can be a problem for sharing remote cache hits across machines. For example, environments with different $PATH variables won't share cache hits. Only environment variables explicitly whitelisted via --action_env are included in an action definition. Bazel's Debian/Ubuntu package used to install /etc/bazel.bazelrc with a whitelist of environment variables including $PATH. If you are getting fewer cache hits than expected, check that your environment doesn't have an old /etc/bazel.bazelrc file.

Bazel does not track tools outside a workspace

Bazel currently does not track tools outside a workspace. This can be a problem if, for example, an action uses a compiler from /usr/bin/. Then, two users with different compilers installed will wrongly share cache hits because the outputs are different but they have the same action hash. See issue #4558 for updates.

Incremental in-memory state is lost when running builds inside docker containers Bazel uses server/client architecture even when running in single docker container. On the server side, Bazel maintains an in-memory state which speeds up builds. When running builds inside docker containers such as in CI, the in-memory state is lost and Bazel must rebuild it before using the remote cache.