This article covers sandboxing in Bazel, installing
sandboxfs, and debugging
your sandboxing environment.
Sandboxing is a permission restricting strategy that isolates processes from each other or from resources in a system. For Bazel, this means restricting file system access.
Bazel's file system sandbox runs processes in a working directory that only contains known inputs, such that compilers and other tools don't see source files they should not access, unless they know the absolute paths to them.
Sandboxing doesn't hide the host environment in any way. Processes can freely access all files on the file system. However, on platforms that support user namespaces, processes can't modify any files outside their working directory. This ensures that the build graph doesn't have hidden dependencies that could affect the reproducibility of the build.
More specifically, Bazel constructs an
execroot/ directory for each action,
which acts as the action's work directory at execution time.
contains all input files to the action and serves as the container for any
generated outputs. Bazel then uses an operating-system-provided technique,
containers on Linux and
sandbox-exec on macOS, to constrain the action within
Reasons for sandboxing
Without action sandboxing, Bazel doesn't know if a tool uses undeclared input files (files that are not explicitly listed in the dependencies of an action). When one of the undeclared input files changes, Bazel still believes that the build is up-to-date and won’t rebuild the action. This can result in an incorrect incremental build.
Incorrect reuse of cache entries creates problems during remote caching. A bad cache entry in a shared cache affects every developer on the project, and wiping the entire remote cache is not a feasible solution.
Sandboxing mimics the behavior of remote execution — if a build works well with sandboxing, it will likely also work with remote execution. By making remote execution upload all necessary files (including local tools), you can significantly reduce maintenance costs for compile clusters compared to having to install the tools on every machine in the cluster every time you want to try out a new compiler or make a change to an existing tool.
What sandbox strategy to use
You can choose which kind of sandboxing to use, if any, with the
strategy flags. Using the
strategy makes Bazel pick one of the sandbox implementations listed below,
preferring an OS-specific sandbox to the less hermetic generic one.
Persistent workers run in a generic sandbox if you pass
standalone) strategy does not do any kind of sandboxing.
It simply executes the action's command line with the working directory set to
the execroot of your workspace.
processwrapper-sandbox is a sandboxing strategy that does not require any
"advanced" features - it should work on any POSIX system out of the box. It
builds a sandbox directory consisting of symlinks that point to the original
source files, executes the action's command line with the working directory set
to this directory instead of the execroot, then moves the known output artifacts
out of the sandbox into the execroot and deletes the sandbox. This prevents the
action from accidentally using any input files that are not declared and from
littering the execroot with unknown output files.
linux-sandbox goes one step further and builds on top of the
processwrapper-sandbox. Similar to what Docker does under the hood, it uses
Linux Namespaces (User, Mount, PID, Network and IPC namespaces) to isolate the
action from the host. That is, it makes the entire filesystem read-only except
for the sandbox directory, so the action cannot accidentally modify anything on
the host filesystem. This prevents situations like a buggy test accidentally rm
-rf'ing your $HOME directory. Optionally, you can also prevent the action from
accessing the network.
linux-sandbox uses PID namespaces to prevent the action
from seeing any other processes and to reliably kill all processes (even daemons
spawned by the action) at the end.
darwin-sandbox is similar, but for macOS. It uses Apple's
to achieve roughly the same as the Linux sandbox.
linux-sandbox and the
darwin-sandbox do not work in a "nested"
scenario due to restrictions in the mechanisms provided by the operating
systems. Because Docker also uses Linux namespaces for its container magic, you
cannot easily run
linux-sandbox inside a Docker container, unless you use
docker run --privileged. On macOS, you cannot run
sandbox-exec inside a
process that's already being sandboxed. Thus, in these cases, Bazel
automatically falls back to using
If you would rather get a build error — such as to not accidentally build with a
less strict execution strategy — explicitly modify the list of execution
strategies that Bazel tries to use (for example,
sandboxfs is a FUSE file system that exposes an arbitrary view of the
underlying file system without time penalties. Bazel uses
execroot/ instantaneously for each action, avoiding the cost of
issuing thousands of system calls. Note that further I/O within
be slower due to FUSE overhead.
Use the following steps to install
sandboxfs and perform a Bazel build with
Download and install
sandboxfs so that the
sandboxfs binary ends up in your
- (macOS-only) Install OSXFUSE.
sudo sysctl -w vfs.generic.osxfuse.tunables.allow_other=1
You will need to do this after installation and after every reboot to ensure core macOS system services work through sandboxfs.
Run a Bazel build with
bazel build target --experimental_use_sandboxfs
If you see
local instead of
linux-sandbox as an
annotation for the actions that are executed, this may mean that sandboxing is
--genrule_strategy=sandboxed --spawn_strategy=sandboxed to
Follow the strategies below to debug issues with sandboxing.
On some platforms, such as Google Kubernetes
Engine cluster nodes
or Debian, user namespaces are deactivated by default due to security
concerns. If the
/proc/sys/kernel/unprivileged_userns_clone file exists and
contains a 0, you can activate user namespaces by running:
sudo sysctl kernel.unprivileged_userns_clone=1
Rule execution failures
The sandbox may fail to execute rules because of the system setup.
If you see a message like
namespace-sandbox.c:633: execvp(argv, argv): No
such file or directory, try to deactivate the sandbox with
--strategy=Genrule=local for genrules, and
for other rules.
Detailed debugging for build failures
If your build failed, use
--sandbox_debug to make
Bazel show the exact command it ran when your build failed, including the part
that sets up the sandbox.
Example error message:
ERROR: path/to/your/project/BUILD:1:1: compilation of rule '//path/to/your/project:all' failed: Sandboxed execution failed, which may be legitimate (such as a compiler error), or due to missing dependencies. To enter the sandbox environment for easier debugging, run the following command in parentheses. On command failure, a bash shell running inside the sandbox will then automatically be spawned namespace-sandbox failed: error executing command (cd /some/path && \ exec env - \ LANG=en_US \ PATH=/some/path/bin:/bin:/usr/bin \ PYTHONPATH=/usr/local/some/path \ /some/path/namespace-sandbox @/sandbox/root/path/this-sandbox-name.params -- /some/path/to/your/some-compiler --some-params some-target)
You can now inspect the generated sandbox directory and see which files Bazel created and run the command again to see how it behaves.
Note that Bazel does not delete the sandbox directory when you use
--sandbox_debug. Unless you are actively debugging, you should disable
--sandbox_debug because it fills up your disk over time.