This page describes multiplex workers, how to write multiplex-compatible rules, and workarounds for certain limitations.
Multiplex workers allow Bazel to handle multiple requests with a single worker process. For multi-threaded workers, Bazel can use fewer resources to achieve the same, or better performance. For example, instead of having one worker process per worker, Bazel can have four multiplexed workers talking to the same worker process, which can then handle requests in parallel. For languages like Java and Scala, this saves JVM warm-up time and JIT compilation time, and in general it allows using one shared cache between all workers of the same type.
Overview
There are two layers between the Bazel server and the worker process. For certain
mnemonics that can run processes in parallel, Bazel gets a WorkerProxy from
the worker pool. The WorkerProxy forwards requests to the worker process
sequentially along with a request_id, the worker process processes the request
and sends responses to the WorkerMultiplexer. When the WorkerMultiplexer
receives a response, it parses the request_id and then forwards the responses
back to the correct WorkerProxy. Just as with non-multiplexed workers, all
communication is done over standard in/out, but the tool cannot just use
stderr for user-visible output (see below).
Each worker has a key. Bazel uses the key's hash code (composed of environment
variables, the execution root, and the mnemonic) to determine which
WorkerMultiplexer to use. WorkerProxys communicate with the same
WorkerMultiplexer if they have the same hash code. Therefore, assuming
environment variables and the execution root are the same in a single Bazel
invocation, each unique mnemonic can only have one WorkerMultiplexer and one
worker process. The total number of workers, including regular workers and
WorkerProxys, is still limited by --worker_max_instances.
Writing multiplex-compatible rules
The rule's worker process should be multi-threaded to take advantage of
multiplex workers. Protobuf allows a ruleset to parse a single request even
though there might be multiple requests piling up in the stream. Whenever the
worker process parses a request from the stream, it should handle the request in
a new thread. Because different thread could complete and write to the stream at
the same time, the worker process needs to make sure the responses are written
atomically (messages don't overlap). Responses must contain the
request_id of the request they're handling.
Handling multiplex output
Multiplex workers need to be more careful about handling their output than
singleplex workers. Anything sent to stderr will go into a single log file
shared among all WorkerProxys of the same type,
randomly interleaved between concurrent requests. While redirecting stdout
into stderr is a good idea, do not collect that output into the output
field of WorkResponse, as that could show the user mangled pieces of output.
If your tool only sends user-oriented output to stdout or stderr, you will
need to change that behaviour before you can enable multiplex workers.
Enabling multiplex workers
Multiplex workers are not enabled by default. A ruleset can turn on multiplex
workers by using the supports-multiplex-workers tag in the
execution_requirements of an action (just like the supports-workers tag
enables regular workers). As is the case when using regular workers, a worker
strategy needs to be specified, either at the ruleset level (for example,
--strategy=[some_mnemonic]=worker) or generally at the strategy level (for
example, --dynamic_local_strategy=worker,standalone.) No additional flags are
necessary, and supports-multiplex-workers takes precedence over
supports-workers, if both are set. You can turn off multiplex workers
globally by passing --noexperimental_worker_multiplex.
A ruleset is encouraged to use multiplex workers if possible, to reduce memory pressure and improve performance. However, multiplex workers are not currently compatible with dynamic execution unless they implement multiplex sandboxing. Attempting to run non-sandboxed multiplex workers with dynamic execution will silently use sandboxed singleplex workers instead.
Multiplex sandboxing
Multiplex workers can be sandboxed by adding explicit support for it in the worker implementations. While singleplex worker sandboxing can be done by running each worker process in its own sandbox, multiplex workers share the process working directory between multiple parallel requests. To allow sandboxing of multiplex workers, the worker must support reading from and writing to a subdirectory specified in each request, instead of directly in its working directory.
To support multiplex sandboxing, the worker must use the sandbox_dir field
from the WorkRequest and use that as a prefix for all file reads and writes.
While the arguments and inputs fields remain unchanged from an unsandboxed
request, the actual inputs are relative to the sandbox_dir. The worker must
translate file paths found in arguments and inputs to read from this
modified path, and must also write all outputs relative to the sandbox_dir.
This includes paths such as '.', as well as paths found in files specified
in the arguments (such as "argfile" arguments).
Once a worker supports multiplex sandboxing, the ruleset can declare this
support by adding supports-multiplex-sandboxing to the
execution_requirements of an action. Bazel will then use multiplex sandboxing
if the --experimental_worker_multiplex_sandboxing flag is passed, or if
the worker is used with dynamic execution.
The worker files of a sandboxed multiplex worker are still relative to the
working directory of the worker process. Thus, if a file is
used both for running the worker and as an input, it must be specified both as
an input in the flagfile argument as well as in tools, executable, or
runfiles.