Pants Daemon

The Pants Daemon (pantsd) is a system introduced to enable Pants to keep information about the build warm in memory between runs. It consists of a process running in the background (currently, one for each buildroot), which listens to filesystem events and keeps a build graph warm. It then passes that graph to subsequent runs that request it.

This document outlines all the moving pieces of pantsd, and explains how it works through an end-to-end run.


ProcessManager is a class designed to keep track of processes. Besides changing their state (paused, terminating...), it allows a process to fork in different ways. Classes that extend ProcessManager can (and often do) inject code to be run before and after forking by overriding the functions pre_fork, post_fork_child, post_fork_parent. Some examples of usage of these functions are DaemonPantsRunner invoking a LocalPantsRunner in its post_fork_child, to fulfill a request to PailgunServer.

The PantsDaemon Class

The PantsDaemon class is responsible for managing the lifetime of the pantsd process and the services associated with it. A PantsDaemon is created with the inner class Factory, and it has two modes of initialization:

  • Stub initialization, which parses options, launches watchman, and not much else. This mode is used to determine whether it needs to fully initialize the daemon or not.
  • Full initialization, which spawns the daemon process, initializes all the services, the legacy engine and the native code.

Initialization is encapsulated in the PantsDaemon.Factory.create() method. PantsDaemon is a ProcessManager, which means one can know if it's alive, or if it needs to restart.

PantsDaemons can be launch()ed, which will terminate the process it was running, and call daemon_spawn() to fork a new process. This new process will run the code in PantsDaemon.post_fork_child(), which in short means it will run os.spawnve to execute the function, which will call PantsDaemon.run_sync(). run_sync() does a lot of things, but the vital things are calling _setup_services() to spin up services, and _run_services() to start an infinte loop polling them.


The Nailgun Protocol is a protocol designed to allow clients to make command-line requests to a Nailgun Server. It supports an interface similar to ExecuteProcessRequest, except for it having no concern about making operations hermetic, and the Nailgun Protocol supports streaming access to stdin/stdout. Pailgun is an extension of the Nailgun Protocol, which is clients use to ask the Pants Daemon to spawn pants invocations.

The protocol is subject to change slightly in #6579.

In Pantsd, PailgunServer is the class responsible for reading Pailgun requests and handling them, by spawning a PailgunHandler.


A Pants Daemon process has several services that it polls in order. Every service runs in a separate thread, and can be paused, resumed or terminated. Services can communicate with each other.

Examples of services are FSEventService, which takes care of listening for watchman events, SchedulerService, which takes care of listening responding to those events to keep a warm Graph, and PailgunService, which listens to SchedulerService and manages the lifetime of a PailgunServer responsible for spawning pants runs when requested by clients, it takes a DaemonPantsRunner as one of its arguments, to use as a template to spawn pants runs when requested by the clients.


A PailgunService is a PantsService which spins up and polls a PailgunServer. A PailgunServer is a TCPServer with ThreadingMixIn, which listens to Pailgun requests in a socket and spins up instances of PailgunHandlers to handle them. It overrides ThreadingMixIn.process_request_thread() to spin up one thread and one handler per request. A PailgunServer holds a reference to the class DaemonPantsRunner, which can be used to run pants from the server. A PailgunHandler is a class that parses the requests sent to the server, and uses DaemonPantsRunner to invoke pants with the environment and arguments specified by the request. A DaemonPantsRunner is a ProcessManager, which implements a run() method that forks the current process (running the pantsd server), creating an instance of LocalPantsRunner in the child process. These child processes are named pantsd-runner [] processes.

An end-to-end run with Pantsd

To understand what the process is for spinning up and closing down pantsd, here is what happens when we run pantsd for the first time:

If we run the command ./pants --enable-pantsd list src/scala::, the following happens:

  1. PantsRunner::run() is called, which will prompt parsing of bootstrap options. --enable-pantsd and --shutdown-pantsd-after-run are both bootstrap options.
  2. In that function, we determine whether we need to run in pantsd mode or local mode. If we choose local mode, an instance of LocalPantsRunner is created and the run will continue as if pantsd didn't exist.

  3. Since --enable-pantsd was toggled on, we will create an instance of RemotePantsRunner, and run it. RemotePantsRunner will:

  4. Maybe launch pantsd, by calling PantsDaemon.Factory.maybe_launch(). This method will:
    • Create a stub instance of PantsDaemon.
    • Check if any of the fingerprinted options have changed by calling ProcessManager.needs_restart(). If they have (or there is no pantsd running at the moment), it will fully initialize an instance of PantsDaemon by calling PantsDaemon.Factory.create(), and launch it with PantsDaemon.launch(). In either case, it will return a PantsDaemon.Handle to a process running PantsDaemon.run_sync(), which will poll all the services.
  5. With a handle to the pantsd process, RemotePantsRunner will now call _run_pants_with_retry(), which will try to _connect_and_execute() to the port supplied by the handle, probably more than once. To do that, it will create a NailgunClient instance with will use the pailgun protocol described above to tell the pantsd process to invoke pants, with a call along the lines of: result = client.execute('./pants', *self._args, **modified_env)
  6. After that request is finished, it will record the result and the local process will exit.

  7. The PailgunService will receive that request, and it will handle it as follows:

  8. PailgunService is endlessly polling for requests via calling PailgunServer.handle_request().
  9. When that method receives a request, it will call SocketServer._handle_request_noblock(), which will call ThreadingMixIn.process_request() (it would usually call SocketServer.process_request, but PailgunServer also extends ThreadingMixIn, which intentionally overrides this function).
  10. ThreadingMixIn.process_request() will spin up a new Thread and call PailgunServer.process_request_thread() (it would usually call ThreadingMixIn.process_request_thread(), but we override it).
  11. (note: we are now in a separate thread) PailgunServer.process_request_thread() will create an instance of PailgunHandler, and call PailgunHandler.handle_request().
  12. PailgunHandler.handle_request() will create an instance of DaemonPantsRunner, and it will run() it. Creating an instance of DaemonPantsRunner with DaemonPantsRunner.create() means that it will call the SchedulerService to get a warm graph.
  13. DaemonPantsRunner will call daemonize(), which will:

    • Stop current services in the daemon (DaemonPantsRunner.pre_fork()).
    • Fork the process.
    • Resume all the services in the parent process (DaemonPantsRunner.post_fork_child()).
    • In the child process, rename the process to pantsd-runner [], create an instance of LocalPantsRunner with the graph (and options, and such) it got from the SchedulerService, and run as if the daemon didn't exist.
  14. The PailgunServer will wait until LocalPantsRunner is finished in handle_request.

Generated by publish_docs from dist/markdown/html/src/docs/architecture_pantsd.html 2019-04-13T09:49:40.247391