Creating persistent workers

Persistent workers can make your build faster. If you have repeated actions in your build that have a high startup cost or would benefit from cross-action caching, you may want to implement your own persistent worker to perform these actions.

The worker implementation has two parts:

Making the worker

A worker upholds a few requirements:

  • It reads WorkRequests from its stdin.
  • It writes WorkResponses (and only WorkResponses) to its stdout.
  • It accepts the --persistent_worker flag.

If your program upholds these requirements, it can be used as a worker!

Work requests

A WorkRequest contains a list of arguments to the worker, a list of path-digest pairs representing the inputs the worker can access (this isn’t enforced, but you can use this info for caching), and a request id, which is 0 for singleplex workers.

{
  “args” : [“--some_argument”],
  “inputs” : [
    { “/path/to/my/file/1 : “fdk3e2ml23d”},
    { “/path/to/my/file/2 : 1fwqd4qdd” }
 ],
  “request_id” : 12
}

Work responses

A WorkResponse should contain the same request id, a zero or nonzero exit code, and an output string that contains any errors encountered in processing or executing the request. Workers may write additional output to stderr, but they must only write WorkResponses to stdout.

{
  “exit_code” : 1,
  “output” : “Action failed with the following message:\nCould not find input
    file “/path/to/my/file/1,
  “request_id” : 12
}

As per the norm for protobufs, the fields are optional. However, Bazel requires the WorkRequest and the corresponding WorkResponse, to have the same request id, so the request id must be specified if it is nonzero. This is a valid WorkResponse.

{
  “request_id” : 12,
}

A request_id of 0 indicates a “singleplex” request, i.e. this request cannot be processed in parallel with other requests. The server guarantees that a given worker receives requests with either only request_id 0 or only request_id greater than zero. Singleplex requests are sent in serial, i.e. the server doesn’t send another request until it has received a response (except for cancel requests, see below).

Notes

  • Each protocol buffer is preceded by its length in varint format (see MessageLite.writeDelimitedTo().
  • JSON requests and responses are not preceded by a size indicator.
  • JSON requests uphold the same structure as the protobuf, but use standard JSON.
  • Bazel stores requests as protobufs and converts them to JSON using protobuf’s JSON format

Cancellation

Workers can optionally allow work requests to be cancelled before they finish. This is particularly useful in connection with dynamic execution, where local execution can regularly be interrupted by a faster remote execution. To allow cancellation, add supports-worker-cancellation: 1 to the execution-requirements field (see below) and set the --experimental_worker_cancellation flag.

A cancel request is a WorkRequest with the cancel field set (and similarly a cancel response is a WorkResponse with the was_cancelled field set). The only other field that must be in a cancel request or cancel response is request_id, indicating which request to cancel. The request_id field will be 0 for singleplex workers or the non-0 request_id of a previously sent WorkRequest for multiplex workers. The server may send cancel requests for requests that the worker has already responded to, in which case the cancel request must be ignored.

Each non-cancel WorkRequest message must be answered exactly once, whether or not it was cancelled. Once the server has sent a cancel request, the worker may respond with a WorkResponse with the request_id set and the was_cancelled field set to true. Sending a regular WorkResponse is also accepted, but the output and exit_code fields will be ignored.

Once a response has been sent for a WorkRequest, the worker must not touch the files in its working directory. The server is free to clean up the files, including temporary files.

Making the rule that uses the worker

You’ll also need to create a rule that generates actions to be performed by the worker. Making a Starlark rule that uses a worker is just like creating any other rule.

In addition, the rule needs to contain a reference to the worker itself, and there are some requirements for the actions it produces.

Referring to the worker

The rule that uses the worker needs to contain a field that refers to the worker itself, so you’ll need to create an instance of a \*\_binary rule to define your worker. If your worker is called MyWorker.Java, this might be the associated rule:

java_binary(
    name = worker,
    srcs = [MyWorker.Java],
)

This creates the “worker” label, which refers to the worker binary. You’ll then define a rule that uses the worker. This rule should define an attribute that refers to the worker binary.

If the worker binary you built is in a package named “work”, which is at the top level of the build, this might be the attribute definition:

"worker": attr.label(
    default = Label("//work:worker"),
    executable = True,
    cfg = "host",
)

cfg = "host" indicates that the worker should be built to run on your host platform.

Work action requirements

The rule that uses the worker creates actions for the worker to perform. These actions have a couple of requirements.

  • The “arguments” field. This takes a list of strings, all but the last of which are arguments passed to the worker upon startup. The last element in the “arguments” list is a flag-file (@-preceded) argument. Workers read the arguments from the specified flagfile on a per-WorkRequest basis. Your rule can write non-startup arguments for the worker to this flagfile.

  • The “execution-requirements” field, which takes a dictionary containing “supports-workers” : “1”, “supports-multiplex-workers” : “1”, or both.

    The “arguments” and “execution-requirements” fields are required for all actions sent to workers. Additionally, actions that should be executed by JSON workers need to include “requires-worker-protocol” : “json” in the execution requirements field. “requires-worker-protocol” : “proto” is also a valid execution requirement, though it’s not required for proto workers, since they are the default.

    You can also set a “worker-key-mnemonic” in the execution requirements. This may be useful if you’re reusing the executable for multiple action types and want to distinguish actions by this worker.

  • Temporary files generated in the course of the action should be saved to the worker’s directory. This enables sandboxing.

Assuming a rule definition with “worker” attribute described above, in addition to a “srcs” attribute representing the inputs, an “output” attribute representing the outputs, and an “args” attribute representing the worker startup args, the call to ctx.actions.run might be:

ctx.actions.run(
  inputs=ctx.files.srcs,
  outputs=[ctx.attr.output],
  executable=ctx.attr.worker,
  mnemonic="someMnemonic",
  execution_requirements={
    supports-workers : 1,
    requires-worker-protocol : json},
  arguments=ctx.attr.args + [@flagfile]
 )

Examples

The Bazel code base uses Java compiler workers, in addition to an example JSON worker that is used in our integration tests.

You can use their scaffolding to make any Java-based tool into a worker by passing in the correct callback.

For an example of a rule that uses a worker, take a look at Bazel’s worker integration test.

External contributors have implemented workers in a variety of languages; you can find many more examples on GitHub!