Getting Started


This section is incomplete.

Liability Waiver

Before proceeding, it is critical you understand what you’re involving yourself and possibly your team and its successors with:

  • Constructing the most fundamental class, Broker, causes a new thread to be spawned, exposing a huge class of difficult to analyse behaviours that Python software generally does not suffer from.

    While every effort is made to hide this complexity, you should expect threading-related encounters during development, and crucially, years after your program reached production. See Troubleshooting for more information.

  • While high-level abstractions are provided, they are only a convenience, you must still understand how Mitogen works before depending on it. Mitogen interacts with many aspects of the operating system, threading, SSH, sudo, sockets, TTYs, shell, Python runtime, and timing and ordering uncertainty introduced through interaction with the network, GIL and OS scheduling.

    Knowledge of this domain is typically attained through painful years of failed attempts hacking system-level programs, and learning through continual suffering how to debug the atrocities left behind. If you feel you lack resources or willpower to diagnose problems independently, Mitogen is not appropriate, prefer a higher level solution instead.

First Principles

Before starting, take a moment to reflect on writing a program that will operate across machines and privilege domains:

  • As with multithreaded programming, writing a program that spans multiple hosts is exposed to many asynchrony issues. Unlike multithreaded programming, the margin for unexpected failures is much higher, even between only two peers, as communication may be fail at any moment, since that communication depends on reliability of an external network.

  • Since a multi-host program always spans trust and privilege domains, trust must be taken into consideration in your design from the outset. Mitogen attempts to protect the consuming application by default where possible, however it is paramount that trust considerations are always in mind when exposing any privileged functionality to a potentially untrusted network of peers.

    A parent must always assume data received from a child is suspect, and must not base privileged control decisions on that data. As a small example, a parent should not form a command to execute in a subprocess using strings received from a child.

  • As the program spans multiple hosts, its design will benefit from a strict separation of program and data. This entails avoiding some common Python idioms that rely on its ability to manipulate functions and closures as if they were data, such as passing a lambda closed over some program state as a callback parameter.

    In the general case this is both difficult and unsafe to support in a distributed program, and so (for now at least) it should be assumed this functionality is unlikely to appear in future.

Broker And Router


Execution starts when your program constructs a Broker and associated Router. The broker is responsible for multiplexing IO to children from a private thread, while in children, it is additionally responsible for ensuring robust destruction if communication with the master is lost.

Router is responsible for receiving messages and dispatching them to a callback from the broker thread (registered by add_handler()), or forwarding them to a Stream. See Message Routing for an in-depth description. Router also doubles as the entry point to Mitogen’s public API:

>>> import mitogen.master

>>> broker = mitogen.master.Broker()
>>> router = mitogen.master.Router(broker)

>>> try:
...     # Your code here.
...     pass
... finally:
...     broker.shutdown()

As Python will not stop if threads still exist after the main thread exits, Broker.shutdown() must be called reliably at exit. Helpers are provided by mitogen.utils to ensure Broker is reliably destroyed:

def do_mitogen_stuff(router):
    # Your code here.


If your program cannot live beneath mitogen.utils.run_with_router() on the stack, you must arrange for Broker.shutdown() to be called anywhere the main thread may exit.

Enable Logging

Mitogen makes heavy use of the logging package, both for child stdio redirection, and soft errors and warnings that may be generated.

You should always configure the logging package in any program that integrates Mitogen. If your program does not otherwise use the logging package, a basic configuration can be performed by calling mitogen.utils.log_to_file():

>>> import mitogen.utils

# Errors, warnings, and child stdio will be written to stderr.
>>> mitogen.utils.log_to_file()

Additionally, if your program has logging.DEBUG as the default logging level, you may wish to update its configuration to restrict the mitogen logger to logging.INFO, otherwise vast amounts of output will be generated by default.

Logging Environment Variables


Overrides the logging package log level set by any call to mitogen.utils.log_to_file(). Defaults to INFO.

If set to IO, equivalent to DEBUG but additionally enabled IO logging for any call to mitogen.utils.log_to_file(). IO logging produces verbose records of any IO interaction, which is useful for debugging hangs and deadlocks.

Logging Records

Messages received from a child context via mitogen.master.LogForwarder receive extra attributes:

  • mitogen_context: mitogen.parent.Context referring to the message source.
  • mitogen_name: original logger name in the source context.
  • mitogen_msg: original message in the source context.

Creating A Context

Contexts are simply external Python programs over which your program has control, and can execute code within. They can be created as subprocesses on the local machine, in another user account via sudo, on a remote machine via ssh, or any recursive combination of the above.

Now a Router exists, our first contexts can be created. To demonstrate basic functionality, we will start with some local() contexts created as subprocesses:

>>> local = router.local()
>>> local_with_name = router.local(remote_name='i-have-a-name')

Examination of the system process list with the pstree utility reveals the resulting process hierarchy:

| |   \-+= 27660 dmw python
| |     |--- 27661 dmw mitogen:dmw@Eldil.local:27660
| |     \--- 27663 dmw mitogen:i-have-a-name

Both contexts are visible as subprocesses of the interactive Python interpreter, with their argv[0] including a description of their identity. To aid systems administrators in identifying errant software running on their machines, the default remote_name includes the location of the program that started the context, however as shown, this can be overridden.


Presently contexts are constructed in a blocking manner on the thread that invoked the context factory. In a future release, the factory will instead return immediately, and construction will happen asynchronously on the broker thread.

Calling A Function

Now that some contexts exist, it is time to execute code in them. Any regular function, static method, or class method reachable directly from module scope may be used, including built-in functions such as time.time().

The method is used to execute a function and block the caller until the return value is available or an exception is raised:

>>> import time
>>> import os

>>> # Returns the current time.
>>> print('Time in remote context:',

>>> try:
...     # Raises OSError.
..., '/nonexistent')
... except mitogen.core.CallError, e:
...     print('Call failed:', str(e))

It is a simple wrapper around the more flexible Context.call_async(), which immediately returns a Receiver wired up to receive the return value instead. A receiver may simply be discarded, kept around indefinitely without ever reading its result, or used to wait on the results from several calls. Here get() is called to block the thread until the result arrives:

>>> call = local.call_async(time.time)
>>> msg = call.get()
>>> print(msg.unpickle())

Running User Functions

So far we have used the interactive interpreter to call some standard library functions, but if since source code typed at the interpreter cannot be recovered, Mitogen is unable to execute functions defined in this way.

We must therefore continue by writing our code as a script:

import mitogen.utils

def my_first_function():
    print('Hello from remote context!')
    return 123

def main(router):
    local = router.local()

if __name__ == '__main__':

Let’s try running it:

$ python
19:11:32 I mitogen.ctx.local.32466: stdout: Hello from remote context!

Waiting On Multiple Calls

Using Context.call_async() it is possible to start multiple function calls then sleep waiting for responses as they are available. This makes it trivial to run tasks in parallel across processes (including remote processes) without the need for writing asynchronous code:

hostnames = ['host1', 'host2', 'host3', 'host4']
contexts = [router.ssh(hostname=hn) for hn in hostnames]
calls = [ for context in contexts]

for msg in
    print('Reply from %s: %s' % (recv.context, data))

Running Code That May Hang

When executing code that may hang due to, for example, talking to network peers that may become unavailable, it is desirable to be able to recover control in the case a remote call has hung.

By specifying the timeout parameter to Receiver.get() on the receiver returned by Context.call_async, it becomes possible to wait for a function to complete, but time out if its result does not become available.

When a context has become hung like this, it is still possible to gracefully terminate it using the Context.shutdown() method. This method sends a shutdown message to the target process, where its IO multiplexer thread can still process it independently of the hung function running on on the target’s main thread.

Recovering Mitogen Object References In Children

def func1(a, b, econtext):

def func2(a, b, router):


Let’s try something a little more complex:

RPC Serialization Rules

The following built-in types may be used as parameters or return values in remote procedure calls:

User-defined types may not be used, except for:

Subclasses of built-in types must be undecorated using mitogen.utils.cast().

Test Your Design

tc qdisc add dev eth0 root netem delay 250ms



This section is incomplete.

A typical example is a hang due to your application’s main thread exitting perhaps due to an unhandled exception, without first arranging for any Broker to be shut down gracefully.

Another example would be your main thread hanging indefinitely because a bug in Mitogen fails to notice an event (such as RPC completion) your thread is waiting for will never complete. Solving this kind of hang is a work in progress.