Worker Manager

The worker manager and its functionality was introduced in version 22.9.

The details of this section are intended for more advanced usages and not necessary to get started.

The purpose of the manager is to create consistency and flexibility between development and production environments. Whether you intend to run a single worker, or multiple workers, whether with, or without auto-reload: the experience will be the same.

In general it looks like this:

When you run Sanic, the main process instantiates a WorkerManager. That manager is in charge of running one or more WorkerProcess. There generally are two kinds of processes:

• server processes, and
• non-server processes.

For the sake of ease, the User Guide generally will use the term "worker" or "worker process" to mean a server process, and "Manager" to mean the single worker manager running in your main process.

How Sanic Server starts processes#

Sanic will start processes using the spawn start method. This means that for every process/worker, the global scope of your application will be run on its own thread. The practical impact of this that if you do not run Sanic with the CLI, you will need to nest the execution code inside a block to make sure it only runs on __main__.

if __name__ == "__main__":
    app.run()

If you do not, you are likely to see an error message like this:

sanic.exceptions.ServerError: Sanic server could not start: [Errno 98] Address already in use.

This may have happened if you are running Sanic in the global scope and not inside of a `if __name__ == "__main__"` block.

See more information: https://sanic.dev/en/guide/deployment/manager.html#how-sanic-server-starts-processes

The likely fix for this problem is nesting your Sanic run call inside of the __name__ == "__main__" block. If you continue to receive this message after nesting, or if you see this while using the CLI, then it means the port you are trying to use is not available on your machine and you must select another port.

Starting a worker#

All worker processes must send an acknowledgement when starting. This happens under the hood, and you as a developer do not need to do anything. However, the Manager will exit with a status code 1 if one or more workers do not send that ack message, or a worker process throws an exception while trying to start. If no exceptions are encountered, the Manager will wait for up to thirty (30) seconds for the acknowledgement.

from sanic.worker.manager import WorkerManager

WorkerManager.THRESHOLD = 100  # Value is in 0.1s

See worker ack for more information.

from sanic import Sanic

Sanic.start_method = "fork"

Worker ack#

When all of your workers are running in a subprocess a potential problem is created: deadlock. This can occur when the child processes cease to function, but the main process is unaware that this happened. Therefore, Sanic servers will automatically send an ack message (short for acknowledge) to the main process after startup.

In version 22.9, the ack timeout was short and limited to 5s. In version 22.12, the timeout was lengthened to 30s. If your application is shutting down after thirty seconds then it might be necessary to manually increase this threshhold.

from sanic.worker.manager import WorkerManager

WorkerManager.THRESHOLD = 600

Zero downtime restarts#

By default, when restarting workers, Sanic will teardown the existing process first before starting a new one.

If you are intending to use the restart functionality in production then you may be interested in having zero-downtime reloading. This can be accomplished by forcing the reloader to change the order to start a new process, wait for it to ack, and then teardown the old process.

app.m.restart(zero_downtime=True)

Added in v22.12

Using shared context between worker processes#

Python provides a few methods for exchanging objects, synchronizing, and sharing state between processes. This usually involves objects from the multiprocessing and ctypes modules.

If you are familiar with these objects and how to work with them, you will be happy to know that Sanic provides an API for sharing these objects between your worker processes. If you are not familiar, you are encouraged to read through the Python documentation linked above and try some of the examples before proceeding with implementing shared context.

Similar to how application context allows an applicaiton to share state across the lifetime of the application with app.ctx, shared context provides the same for the special objects mentioned above. This context is available as app.shared_ctx and should ONLY be used to share objects intended for this purpose.

The shared_ctx will:

• NOT share regular objects like int, dict, or list
• NOT share state between Sanic instances running on different machines
• NOT share state to non-worker processes
• only share state between server workers managed by the same Manager

Attaching an inappropriate object to shared_ctx will likely result in a warning, and not an error. You should be careful to not accidentally add an unsafe object to shared_ctx as it may not work as expected. If you are directed here because of one of those warnings, you might have accidentally used an unsafe object in shared_ctx.

from multiprocessing import Queue

@app.main_process_start
async def main_process_start(app):
    app.shared_ctx.queue = Queue()

Trying to attach to the shared_ctx object outside of this listener may result in a RuntimeError.

from multiprocessing import Queue

@app.get("")
async def handler(request):
    request.app.shared_ctx.queue.put(1)
    ...

Access to the multiplexer#

The application instance has access to an object that provides access to interacting with the Manager and other worker processes. The object is attached as the app.multiplexer property, but it is more easily accessed by its alias: app.m.

@app.on_request
async def print_state(request: Request):
    print(request.app.m.name)
    print(request.app.m.pid)
    print(request.app.m.state)

Sanic-Server-0-0
99999
{'server': True, 'state': 'ACKED', 'pid': 99999, 'start_at': datetime.datetime(2022, 10, 1, 0, 0, 0, 0, tzinfo=datetime.timezone.utc), 'starts': 2, 'restart_at': datetime.datetime(2022, 10, 1, 0, 0, 12, 861332, tzinfo=datetime.timezone.utc)}

# shutdown the entire application and all processes
app.m.name.terminate()

# restart the current worker only
app.m.name.restart()

# restart specific workers only (comma delimited)
app.m.name.restart("Sanic-Server-4-0,Sanic-Server-7-0")

# restart ALL workers
app.m.name.restart(all_workers=True)  # Available v22.12+

Worker state#

@app.on_request
async def print_state(request: Request):
    print(request.app.m.workers)

{
    'Sanic-Main': {'pid': 99997},
    'Sanic-Server-0-0': {
        'server': True,
        'state': 'ACKED',
        'pid': 9999,
        'start_at': datetime.datetime(2022, 10, 1, 0, 0, 0, 0, tzinfo=datetime.timezone.utc),
        'starts': 2,
        'restart_at': datetime.datetime(2022, 10, 1, 0, 0, 12, 861332, tzinfo=datetime.timezone.utc)
    },
    'Sanic-Reloader-0': {
        'server': False,
        'state': 'STARTED',
        'pid': 99998,
        'start_at': datetime.datetime(2022, 10, 1, 0, 0, 0, 0, tzinfo=datetime.timezone.utc),
        'starts': 1
    }
}

The possible states are:

• NONE - The worker has been created, but there is no process yet
• IDLE - The process has been created, but is not running yet
• STARTING - The process is starting
• STARTED - The process has started
• ACKED - The process has started and sent an acknowledgement (usually only for server processes)
• JOINED - The process has exited and joined the main process
• TERMINATED - The process has exited and terminated
• RESTARTING - The process is restarting
• FAILED - The process encountered an exception and is no longer running
• COMPLETED - The process has completed its work and exited successfully

Built-in non-server processes#

As mentioned, the Manager also has the ability to run non-server processes. Sanic comes with two built-in types of non-server processes, and allows for creating custom processes.

The two built-in processes are

• the auto-reloader, optionally enabled to watch the file system for changes and trigger a restart
• inspector, optionally enabled to provide external access to the state of the running instance

Inspector#

Sanic has the ability to expose the state and the functionality of the multiplexer to the CLI. Currently, this requires the CLI command to be run on the same machine as the running Sanic instance. By default the inspector is disabled.

app.config.INSPECTOR = True

You will now have access to execute any of these CLI commands:

sanic inspect reload                      Trigger a reload of the server workers
sanic inspect shutdown                    Shutdown the application and all processes
sanic inspect scale N                     Scale the number of workers to N
sanic inspect <custom>                    Run a custom command

app.config.INSPECTOR_HOST =  "localhost"
app.config.INSPECTOR_PORT =  6457

Learn more to find out what is possible with the Inspector.

Running custom processes#

To run a managed custom process on Sanic, you must create a callable. If that process is meant to be long-running, then it should handle a shutdown call by a SIGINT or SIGTERM signal.

from time import sleep

def my_process(foo):
    try:
        while True:
            sleep(1)
    except KeyboardInterrupt:
        print("done")

@app.main_process_ready
async def ready(app: Sanic, _):
#   app.manager.manage(<name>, <callable>, <kwargs>)
    app.manager.manage("MyProcess", my_process, {"foo": "bar"})

Transient v. durable processes#

@app.main_process_ready
async def ready(app: Sanic, _):
    app.manager.manage(
        "MyProcess",
        my_process,
        {"foo": "bar"},
        transient=True,
    )

Tracked v. untracked processes#

@app.main_process_ready
async def ready(app: Sanic, _):
    app.manager.manage(
        "OneAndDone",
        do_once,
        {},
        tracked=False,
    )

Added in v23.12

Restartable custom processes#

@app.main_process_ready
async def ready(app: Sanic, _):
    app.manager.manage(
        "MyProcess",
        my_process,
        {"foo": "bar"},
        restartable=True,
    )

@app.get("/restart")
async def restart_handler(request: Request):
    request.app.m.restart("Sanic-MyProcess-0")
    return json({"foo": request.app.m.name})

Added in v23.12

On the fly process management#

@app.post("/start")
async def start_handler(request: Request):
    request.app.m.manage(
        "MyProcess",
        my_process,
        {"foo": "bar"},
        workers=2,
    )
    return json({"foo": request.app.m.name})

Added in v23.12

Single process mode#

sanic path.to.server:app --single-process

if __name__ == "__main__":
    app.run(single_process=True)

if __name__ == "__main__":
    app.prepare(single_process=True)
    Sanic.serve_single()

Sanic and multiprocessing#

Sanic makes heavy use of the multiprocessing module to manage the worker processes. You should generally avoid lower level usage of this module (like setting the start method) as it may interfere with the functionality of Sanic.

Start methods in Python#

Before explaining what Sanic tries to do, it is important to understand what the start_method is and why it is important. Python generally allows for three different methods of starting a process:

• fork
• spawn
• forkserver

The fork and forkserver methods are only available on Unix systems, and spawn is the only method available on Windows. On Unix systems where you have a choice, fork is generally the default system method.

You are encouraged to read the Python documentation to learn more about the differences between these methods. However, the important thing to know is that fork basically copies the entire memory of the parent process into the child process, whereas spawn will create a new process and then load the application into that process. This is the reason why you need to nest your Sanic run call inside of the __name__ == "__main__" block if you are not using the CLI.

Sanic and start methods#

By default, Sanic will try and use spawn as the start method. This is because it is the only method available on Windows, and it is the safest method on Unix systems.

from sanic import Sanic

Sanic.start_method = "fork"

Overcoming a RuntimeError#

You might have received a RuntimeError that looks like this:

RuntimeError: Start method 'spawn' was requested, but 'fork' was already set.

If so, that means somewhere in your application you are trying to set the start method that conflicts with what Sanic is trying to do. You have a few options to resolve this:

from sanic import Sanic

Sanic.START_METHOD_SET = True

from sanic import Sanic

Sanic.start_method = "fork"

import multiprocessing

multiprocessing.set_start_method("spawn")

In any of these options, you should run this code as early as possible in your application. Depending upon exactly what your specific scenario is, you may need to combine some of the options.

# This is BAD; avoid the global scope
from multiprocessing import Queue

q = Queue()

# This is GOOD; the queue is made in a listener and shared to all the processes on the shared_ctx
from multiprocessing import Queue

@app.main_process_start
async def main_process_start(app):
    app.shared_ctx.q = Queue()