Skip to main content

RBE Platforms

BuildBuddy default

BuildBuddy's default platform is Ubuntu 16.04 with Java 8 installed. Building on our basic command can specify this platform with the --host_platform flag:

--host_platform=@buildbuddy_toolchain//:platform

Using a custom Docker image

You can configure BuildBuddy RBE to use a custom docker image, by adding the following rule to a BUILD file:

BUILD
platform(
name = "docker_image_platform",
constraint_values = [
"@platforms//cpu:x86_64",
"@platforms//os:linux",
"@bazel_tools//tools/cpp:clang",
],
exec_properties = {
"OSFamily": "Linux",
"dockerNetwork": "off",
"container-image": "docker://gcr.io/YOUR:IMAGE",
},
)

Make sure to replace gcr.io/YOUR:IMAGE with your docker image url.

You can then pass this configuration to BuildBuddy RBE with the following flag:

--host_platform=//:docker_image_platform

This assumes you've placed this rule in your root BUILD file. If you place it elsewhere, make sure to update the path accordingly.

ENTRYPOINT and CMD

Remote build actions will be run in your container via CMD, so note that any CMD instructions in your Dockerfile will be ignored. ENTRYPOINT, on the other hand, is not ignored, so make sure that the container image's ENTRYPOINT is either unset, or is a wrapper that is compatible with your build actions' commands.

For more information, see Understand how CMD and ENTRYPOINT interact.

Passing credentials for Docker images

You can use images from private container registries by adding the following flags to your bazel command (replace USERNAME and ACCESS_TOKEN with the appropriate credentials for the container registry):

--remote_exec_header=x-buildbuddy-platform.container-registry-username=USERNAME
--remote_exec_header=x-buildbuddy-platform.container-registry-password=ACCESS_TOKEN

For the value of ACCESS_TOKEN, we recommend generating a short-lived token using the command-line tool for your cloud provider.

To generate a short-lived token for GCR (Google Container Registry), the username must be _dcgcloud_token and the token can be generated with gcloud auth print-access-token:

--remote_exec_header=x-buildbuddy-platform.container-registry-username=_dcgcloud_token
--remote_exec_header=x-buildbuddy-platform.container-registry-password="$(gcloud auth print-access-token)"

For Amazon ECR (Elastic Container Registry), the username must be AWS and a short-lived token can be generated with aws ecr get-login-password --region REGION (replace REGION with the region matching the ECR image URL):

--remote_exec_header=x-buildbuddy-platform.container-registry-username=AWS
--remote_exec_header=x-buildbuddy-platform.container-registry-password="$(aws ecr get-login-password --region REGION)"

Some cloud providers may also allow the use of long-lived tokens, which can also be used in remote headers. For example, GCR allows setting a username of _json_key and then using a service account's JSON-format private key as the password. Note that remote headers cannot have newlines; the command tr '\n' ' ' is used in this example to remove them:

--remote_exec_header=x-buildbuddy-platform.container-registry-username=_json_key
--remote_exec_header=x-buildbuddy-platform.container-registry-password="$(cat service-account-keyfile.json | tr '\n' ' ')"

Specifying a custom executor pool

You can configure BuildBuddy RBE to use a custom executor pool, by adding the following rule to a BUILD file:

BUILD
platform(
name = "gpu_platform",
constraint_values = [
"@platforms//cpu:x86_64",
"@platforms//os:linux",
"@bazel_tools//tools/cpp:clang",
],
exec_properties = {
"OSFamily": "Linux",
"dockerNetwork": "off",
"Pool": "my-gpu-pool",
},
)

Make sure to replace my-gpu-pool with your pool name.

You can then pass this configuration to BuildBuddy RBE with the following flag:

--host_platform=//:gpu_platform

This assumes you've placed this rule in your root BUILD file. If you place it elsewhere, make sure to update the path accordingly.

For instructions on how to deploy custom executor pools, we the RBE Executor Pools docs.

Target level execution properties

If you want different targets to run in different RBE environments, you can specify exec_properties at the target level. For example if you want to run one set of tests in a high-memory pool, or another set of targets on executors with GPUs.

BUILD
go_test(
name = "memory_hogging_test",
srcs = ["memory_hogging_test.go"],
embed = [":go_default_library"],
exec_properties = {
"Pool": "high-memory-pool",
},
)

Execution properties

BuildBuddy RBE supports various exec_properties that modify remote action execution.

These properties can be used in different ways:

  • Set exec_properties in the execution platform definition.
  • Set exec_properties in each BUILD target.
  • Set --remote_default_exec_properties=KEY=VALUE in .bazelrc or at the Bazel command line. Note that these properties are not apply if you are already using a platform.
  • Set --remote_header=x-buildbuddy-platform.KEY=VALUE. This is a BuildBuddy-specific feature and is not generally recommended except for certain properties, described in [Setting properties via remote headers][#setting-properties-via-remote-headers]

Execution groups allow more control over which execution properties can be used for each group of actions in each BUILD target. Execution groups are typically implemented by individual rules, but notably, Bazel includes a built-in execution group called "test" that allows applying execution properties only to tests.

Setting properties via remote headers

BuildBuddy supports setting execution properties via remote headers. This can be done by setting --remote_header=x-buildbuddy-platform.KEY=VALUE at the Bazel command line.

This feature is useful as a more secure option for passing secret property values, such as container-registry-password or container-registry-username, since other methods of setting exec properties result in the properties being stored in the remote cache, while header values are not stored.

Properties which are set via remote headers take precedence over any exec properties set via any other method.

The following is the complete list of properties which are officially supported for use in remote_header. Other properties may work but are not officially supported and can break at any time.

  • container-registry-username
  • container-registry-password

Action scheduling properties

These execution properties affect how BuildBuddy's scheduler selects an executor for action execution:

  • Pool: selects which executor pool to use.
  • OSFamily: selects which operating system the executor must be running. Available options are linux (default), darwin, and windows (darwin and windows are currently only available for self-hosted executors).
  • Arch: selects which CPU architecture the executor must be running on. Available options are amd64 (default) and arm64.
  • use-self-hosted-executors: use self-hosted executors instead of BuildBuddy's managed executor pool. Available options are true and false. The default value is configurable from organization settings.

Action isolation and hermeticity properties

When executing actions, each BuildBuddy executor can spin up multiple action runners. Each runner executes one action at a time. Each runner has a workspace which represents the working directory of the action and contains the action's input tree. Each runner also has an isolation strategy which decides which technology is used to isolate the action from others running on the same machine. Isolation strategies may also be loosely referred to as containers or sandboxes.

The following properties allow customizing the behavior of the runner:

  • workload-isolation-type: selects which isolation technology is the runner should use. When using BuildBuddy Cloud executors, podman (the default) and firecracker are supported. For self-hosted executors, the available options are docker, podman, firecracker, sandbox, and none. The executor must have relevant flags enabled.
  • recycle-runner: whether to retain the runner after action execution and reuse it to execute subsequent actions. The runner's container is paused between actions, and the workspace is cleaned between actions by default. This option may be useful to improve performance in some situations, but is not generally recommended for most actions as it reduces action hermeticity. Available options are true and false.
  • preserve-workspace: only applicable when "recycle-runner": "true" is set. Whether to re-use the Workspace directory from the previous action. Available options are true and false.
  • clean-workspace-inputs: a comma-separated list of glob values that decides which files in the action's input tree to clean up before the action starts. Has no effect unless preserve-workspace is true. Glob patterns should follow the specification in gobwas/glob library.
  • nonroot-workspace: If set to true, the workspace directory will be writable by non-root users (permission 0o777). Otherwise, it will be read-only to non-root users (permission 0o755).

Runner resource allocation

BuildBuddy's scheduler intelligently allocates resources to actions, so it's generally not needed to manually configure resources for actions. However, some exec_properties are provided as manual overrides:

  • EstimatedCPU: the CPU time allocated for the action. Example values:
    • 2: 2 CPU cores
    • 0.5: 500 MilliCPU
    • 4000m: 4000 MilliCPU
  • EstimatedMemory: the memory allocated to the action. Example values:
    • 1M: 1 MB
    • 2GB: 2 GB
    • 4.5GB: 4.5 GB
  • EstimatedComputeUnits: a convenience unit that specifies both CPU and memory. One compute unit is defined as 1 CPU and 2.5GB of memory. Accepts numerical values, e.g. 1 or 9.
  • EstimatedFreeDiskBytes: the amount of disk space allocated to the action. Example values:
    • 1M: 1 MB
    • 2GB: 2 GB
    • 4.5GB: 4.5 GB

Remote persistent worker properties

Similar to the local execution environment, the remote execution environment may also retain a long-running process acting as a persistent worker to help reduce the total cost of cold-starts for build actions with high startup cost.

To use remote persistent workers, the action must have "recycle-runner": "true" in exec_properties.

The Bazel's flag "--experimental_remote_mark_tool_inputs" should help set these automatically so you don't have to set them manually. However, we provide these exec_properties for rules authors to experiment with.

  • persistentWorkerKey: unique key for the persistent worker. This should be automatically set by Bazel.
  • persistentWorkerProtocol: the serialization protocol used by the persistent worker. Available options are proto (default) and json.

Runner container support

For docker, podman, and firecracker isolation, the executor supports running actions in user-provided container images.

Some of these have a docker prefix. This is just a historical artifact; all properties with this prefix also apply to podman. Our goal is to support all of these for firecracker as well, but we are not there yet.

The following execution properties provide more customization.

  • container-image: the container image to use, in the format docker://<container-image>:<tag>. The default value is docker://gcr.io/flame-public/executor-docker-default:enterprise-v1.6.0. For Firecracker, the image will be converted to a VM root image automatically.
  • container-registry-username and container-registry-password: credentials to be used to pull private container images. These are not needed if the image is public. We recommend setting these via remote headers, to avoid storing them in the cache.
  • dockerUser: determines which user the action should be run with inside the container. The default is the user set on the image. If setting to a non-root user, you may also need to set nonroot-workspace to true.

The following properties apply to podman and docker isolation, and are currently unsupported by firecracker. (The docker prefix is just a historical artifact.)

  • dockerInit: determines whether --init should be used when starting a container. Available options are true and false. Defaults to false.
  • dockerRunAsRoot: when set to true, forces the container to run as root, even the image specification specifies a non-root USER. Available options are true and false. Defaults to false.
  • dockerNetwork: determines which network mode should be used. For sandbox isolation, this determines whether the network is enabled or not. Available options are off and bridge. The default is bridge, but we strongly recommend setting this to off for faster runner startup time. The latest version of the BuildBuddy toolchain does this for you automatically.

Runner secrets

Please consult RBE secrets for more information on the related properties.

Docker daemon support

For firecracker isolation, we support starting a Docker daemon (dockerd) which allows actions to run Docker containers. Check out our RBE with Firecracker MicroVMs doc for examples.

The following exec_properties are supported:

  • init-dockerd: whether to start the dockerd process inside the VM before execution. Available options are true and false. Defaults to false.
  • enable-dockerd-tcp: whether dockerd should listen on TCP port 2375 in addition to the default Unix domain socket. Available options are true and false. Defaults to false.