papa 1.0.6

Simple socket and process kernel

Summary

papa is a process kernel. It contains both a client library and a server component for creating sockets and launching processes from a stable parent process.

Dependencies

Papa has no external dependencies, and it never will! It has been tested under the following Python versions:

• 2.6

• 2.7

• 3.2

• 3.3

• 3.4

Installation

$> pip install papa

Purpose

Sometimes you want to be able to start a process and have it survive on its own, but you still want to be able to capture the output. You could daemonize it and pipe the output to files, but that is a pain and lacks flexibility when it comes to handling the output.

Process managers such as circus and supervisor are very good for starting and stopping processes, and for ensuring that they are automatically restarted when they die. However, if you need to restart the process manager, all of their managed processes must be brought down as well. In this day of zero downtime, that is no longer okay.

Papa is a process kernel. It has extremely limited functionality and it has zero external dependencies. If I’ve done my job right, you should never need to upgrade the papa package. There will probably be a few bug fixes before it is really “done”, but the design goal was to create something that did NOT do everything, but only did the bare minimum required. The big process managers can add the remaining features.

Papa has 3 types of things it manages:

• Sockets

• Values

• Processes

Here is what papa does:

• Create sockets and close sockets

• Set, get and clear named values

• Start processes and capture their stdout/stderr

• Allow you to retrieve the stdout/stderr of the processes started by papa

• Pass socket file descriptors and port numbers to processes as they start

Here is what it does NOT do:

• Stop processes

• Send signals to processes

• Restart processes

• Communicate with processes in any way other than to capture their output

Sockets

By managing sockets, papa can manage interprocess communication. Just create a socket in papa and then pass the file descriptor to your process to use it. See the Circus docs for a very good description of why this is so useful.

Papa can create Unix, INET and INET6 sockets. By default it will create an INET TCP socket on an OS-assigned port.

You can pass either the file descriptor (fileno) or the port of a socket to a process by including a pattern like this in the process arguments:

• $(socket.my_awesome_socket_name.fileno)

• $(socket.my_awesome_socket_name.port)

Values

Papa has a very simple name/value pair storage. This works much like environment variables. The values must be text, so if you want to store a complex structure, you will need to encode and decode with something like the JSON module.

The primary purpose of this facility is to store state information for your process that will survive between restarts. For instance, a process manager can store the current state that all of its managed processes are supposed to be in. Then if the process manager is restarted, it can restore its internal state, then go about checking to see if anything on the machine has changed. Are all processes that should be running actually running?

Processes

Processes can be started with or without output management. You can specify a maximum size for output to be cached. Each started process has a management thread in the Papa kernel watching its state and capturing output if necessary.

A Note on Naming (Namespacing)

Sockets, values and processes all have unique names. A name can only represent one item per class. So you could have an “aack” socket, an “aack” value and an “aack” process, but you cannot have two “aack” processes.

All of the monitoring commands support a final asterix as a wildcard. So you can get a list of sockets whose names match “uwsgi*” and you would get any socket that starts with “uwsgi”.

One good naming scheme is to prefix all names with the name of your own application. So, for instance, the Circus process manager can prefix all names with “circus.” and the Supervisor process manager can prefix all names with “supervisor.”. If you write your own simple process manager, just prefix it with “tweeter.” or “facebooklet.” or whatever your project is called.

If you need to have multiple copies of something, put a number after a dot for each of those as well. For instance, if you are starting 3 waitress instances in circus, call them circus.waitress.1, circus.waitress.2, and circus.waitress.3. That way you can query for all processes named circus.* to see all processes managed by circus, or query for circus.waitress.* to see all waitress processes managed by circus.

Starting the kernel

There are two ways to start the kernel. You can run it as a process, or you can just try to access it from the client library and allow it to autostart. The client library uses a lock to ensure that multiple threads do not start the server at the same time but there is currently no protection against multiple processes doing so.

By default, the papa kernel process will communicate over port 20202. You can change this by specifying a different port number or a path. By specifying a path, a Unix socket will be used instead.

If you are going to be creating papa client instances in many places in your code, you may want to just call papa.set_default_port or papa.set_default_path once when your application is starting and then just instantiate the Papa object with no parameters.

Telnet interface

Papa has been designed so that you can communicate with the process kernel entirely without code. Just start the Papa server, then do this:

telnet localhost 20202

You should get a welcome message and a prompt. Type “help” to get help. Type “help process” to get help on the process command.

The most useful commands from a monitoring standpoint are:

• list sockets

• list processes

• list values

All of these can by used with no arguments, or can be followed by a list of names, including wildcards. For instance, to see all of the values in the circus and supervisor namespaces, do this:

list values circus.* supervisor.*

You can abbreviate every command as short as you like. So “l p” means “list processes” and “h l p” means “help list processes”

Creating a Connection

You can create either long-lived or short-lived connections to the Papa kernel. If you want to have a long-lived connection, just create a Papa object to connect and close it when done, like this:

class MyObject(object):
    def __init__(self):
        self.papa = Papa()

    def start_stuff(self):
        self.papa.make_socket('uwsgi')
        self.papa.make_process('uwsgi', 'env/bin/uwsgi', args=('--ini', 'uwsgi.ini', '--socket', 'fd://$(socket.uwsgi.fileno)'), working_dir='/Users/aackbar/awesome', env=os.environ)
        self.papa.make_process('http_receiver', sys.executable, args=('http.py', '$(socket.uwsgi.port)'), working_dir='/Users/aackbar/awesome', env=os.environ)

    def close(self):
        self.papa.close()

If you want to just fire off a few commands and leave, it is better to use the with mechanism like this:

from papa import Papa

with Papa() as p:
    print(p.list_sockets())
    print(p.make_socket('uwsgi', port=8080))
    print(p.list_sockets())
    print(p.make_process('uwsgi', 'env/bin/uwsgi', args=('--ini', 'uwsgi.ini', '--socket', 'fd://$(socket.uwsgi.fileno)'), working_dir='/Users/aackbar/awesome', env=os.environ))
    print(p.make_process('http_receiver', sys.executable, args=('http.py', '$(socket.uwsgi.port)'), working_dir='/Users/aackbar/awesome', env=os.environ))
    print(p.list_processes())

This will make a new connection, do a bunch of work, then close the connection.

Socket Commands

There are 3 socket commands.

p.list_sockets(*args)

The sockets command takes a list of socket names to get info about. All of these are valid:

• p.list_sockets()

• p.list_sockets('circus.*')

• p.list_sockets('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

A dict is returned with socket names as keys and socket details as values.

p.make_socket(name, host=None, port=None, family=None, socket_type=None, backlog=None, path=None, umask=None, interface=None, reuseport=None)

All parameters are optional except for the name. To create a standard TCP socket on port 8080, you can do this:

p.make_socket('circus.uwsgi', port=8080)

To make a Unix socket, do this:

p.make_socket('circus.uwsgi', path='/tmp/uwsgi.sock')

A path for a Unix socket must be an absolute path or make_socket will raise a papa.Error exception.

You can also leave out the path and port to create a standard TCP socket with an OS-assigned port. This is really handy when you do not care what port is used.

If you call make_socket with the name of a socket that already exists, papa will return the original socket if all parameters match, or raise a papa.Error exception if some parameters differ.

See the make_sockets method of the Papa object for other parameters.

p.remove_sockets(*args)

The remove_sockets command also takes a list of socket names. All of these are valid:

• p.remove_sockets('circus.*')

• p.remove_sockets('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

Removing a socket will prevent any future processes from using it, but any processes that were already started using the file descriptor of the socket will continue to use the copy they inherited.

Value Commands

There are 4 value commands.

p.list_values(*args)

The list_values command takes a list of values to retrieve. All of these are valid:

• p.list_values()

• p.list_values('circus.*')

• p.list_list_values('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

A dict will be returned with all matching names and values.

p.set(name, value=None)

To set a value, do this:

p.set('circus.uswgi', value)

You can clear a single value by setting it to None.

p.get(name)

To retrieve a value, do this:

value = p.get('circus.uwsgi')

If no value is stored by that name, None will be returned.

p.remove_values(*args)

To remove a value or values, do something like this:

• p.remove_values('circus.*')

• p.remove_values('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

You cannot remove all variables so passing no names or passing * will raise a papa.Error exception.

Process Commands

There are 4 process commands:

p.list_processes(*args)

The list_processes command takes a list of process names to get info about. All of these are valid:

• p.list_processes()

• p.list_processes('circus.*')

• p.list_processes('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

A dict is returned with process names as keys and process details as values.

p.make_process(name, executable, args=None, env=None, working_dir=None, uid=None, gid=None, rlimits=None, stdout=None, stderr=None, bufsize=None, watch_immediately=None)

Every process must have a unique name and an executable. All other parameters are optional. The make_process method returns a dict that contains the pid of the process.

The args parameter should be a tuple of command-line arguments. If you have only one argument, papa conveniently supports passing that as a string.

You will probably want to pass working_dir. If you do not, the working directory will be that of the papa kernel process.

By default, stdout and stderr are captured so that you can retrieve them with the watch command. By default, the bufsize for the output is 1MB.

Valid values for stdout and stderr are papa.DEVNULL and papa.PIPE (the default). You can also pass papa.STDOUT to stderr to merge the streams.

If you pass bufsize=0, not output will be recorded. Otherwise, bufsize can be the number of bytes, or a number followed by ‘k’, ‘m’ or ‘g’. If you want a 2 MB buffer, you can pass bufsize='2m', for instance. If you do not retrieve the output quicky enough and the buffer overflows, older data is removed to make room.

If you specify uid, it can be either the numeric id of the user or the username string. Likewise, gid can be either the numeric group id or the group name string.

If you want to specify rlimits, pass a dict with rlimit names and numeric values. Valid rlimit names can be found in the resources module. Leave off the RLIMIT_ prefix. On my system, valid names are as, core, cpu, data, fsize, memlock, nofile, nproc, rss, and stack.

rlimit={'cpu': 2, 'nofile': 1024}

The env parameter also takes a dict with names and values. A useful trick is to do env=os.environ to copy your environment to the new process.

If you want to run a Python application and you wish to use the same Python executable as your client application, a useful trick is to pass sys.executable as the executable and the path to the Python script as the first element of your args tuple. If you have no other args, just pass the path as a string to args.

p.make_process('write3', sys.executable, args='executables/write_three_lines.py', working_dir=here, uid=os.environ['LOGNAME'], env=os.environ)

The final argument that needs mention is watch_immediately. If you pass True for this, papa will make the process and return a Watcher. This is effectively the same as doing p.make_process(name, ...) followed immediately by p.watch(name), but it has one fewer round-trip communication with the kernel. If all you want to do is launch an application and monitor its output, this is a good way to go.

p.remove_processes(*args)

If you do not care about retrieving the output or the exit code for a process, you can use remove_processes to tell the papa kernel to close the output buffers and automatically remove the process from the process list when it exits.

• p.remove_processes('circus.logger')

• p.remove_processes('circus.uwsgi', 'circus.nginx.*', 'circus.logger')

p.watch_processes(*args)

The watch_processes command returns a Watcher object for the specified process or processes. That object uses a separate socket to retrieve the output of the processes it is watching.

Optimization Note: Actually, it hijacks the socket of your Papa object. If you issue any other commands to the Papa object that require a connection to the kernel, the Papa object will silently create a new socket and connect up for the additional commands. If you close the Watcher and the Papa object has not already created a new connection, the socket will be returned to the Papa object. So if you launch an application, use watch to grab all of its output until it closes, then use the set command to update your saved status, all of that can occur with a single connection.

The Watcher object

When you use watch_processes or when you do make_process with watch_immediately=True, you get back a Watcher object.

You can use watchers manually or with a context manager. Here is an example without a context manager:

class MyLogger(object):
    def __init__(self, watcher):
        self.watcher = watcher

    def save_stuff(self):
        if self.watcher and self.watcher.ready:
            out, err, closed = self.watcher.read()
            ... save it ...
            self.watcher.acknowledge()  # remove it from the buffer

    def close(self):
        self.watcher.close()

If you are running your logger in a separate thread anyway, you might want to just use a context manager, like this:

with p.watch_processes('aack') as watcher:
    while watcher:
        out, err, closed = watcher.read()  # block until something arrives
        ... save it ...
        watcher.acknowledge()  # remove it from the buffer

The Watcher object has a fileno method, so it can be used with select.select, like this:

watchers = []

watchers.append(p.watch_processes('circus.uwsgi'))
watchers.append(p.watch_processes('circus.nginx'))
watchers.append(p.watch_processes('circus.mongos.*'))

while watchers:
    ready_watchers = select.select(watchers, [], [])[0]  # wait for one of these
    for watcher in ready_watchers:  # iterate through all that are ready
        out, err, closed = watcher.read()
        ... save it ...
        watcher.acknowledge()
        if not watcher:  # if it is done, remove this watcher from the list
            watcher.close()
            del watchers[watcher]

Of course, in the above example it would have been even more efficient to just use a single watcher, like this:

with p.watch_processes('circus.uwsgi', 'circus.nginx', 'circus.mongos.*') as watcher:
    while watcher:
        out, err, closed = watcher.read()
        ... save it ...
        # watcher.acknowledge() - no need since watcher.read will do it for us

w.ready

This property is True if the Watcher has data available to read on the socket.

w.read()

Read will grab all waiting output from the Watcher and return a tuple of (out, err, closed). Each of these is an array of papa.ProcessOutput objects. An output object is actually a namedtuple with 3 values: name, timestamp, and data.

The name element is the name of the process. The timestamp is a float of when the data was captured by the papa kernel. The data is a binary string if found in either the out or err array. It is the exit code if found in the closed array. Using all of these elements, you can write proper timestamps into your logs, even if data was captured by the papa kernel minutes, hours or days earlier.

The read method will block if no data is ready to read. If you do not want to block, use either the ready property or a mechanism such as select.select before calling read.

w.acknowledge()

Just because your have read output from a process, the papa kernel cannot know that you successfully logged it. Maybe you crashed or were shutdown before you had the chance. So the papa kernel will hold onto the data until you acknowledge receipt. This can be done either by calling acknowledge, or by doing a subsequent read or a close.

w.close()

When you are done with a Watcher, be sure to close it. That will release the socket and potentially even return the socket back to the original Papa object. It will also send off a final acknowledge if necessary.

If you use a context manager, the close happens automatically.

if watcher:

A boolean check on the Watcher object will return True if it is still active and False if it has received and acknowledged a close message from all processes it is monitoring.

WARNING: There should be only one

You will get very screwy results if you have multiple watchers for the same process. Each will get the available data, then acknowledge receipt at some point, removing that data from the queue. Based on timing, both will get overlapping results, but neither is likely to get everything.

Shutting Down

Papa is meant to be a long-lived process and it is meant to be usable by multiple client apps. If you would like to shut Papa down, you can try p.exit_if_idle(). This call will only exit Papa if there are no processes, sockets or values. So if your app cleaned everything up and no other app is using Papa, exit_if_idle will allow Papa to die. It will return True if Papa has indicated that it will exit when the connection closes.

If you want to do a complete cleanup, kill all of your processes however you like, then do:

p.remove_processes('myapp.*')
p.remove_sockets('myapp.*')
p.remove_values('myapp.*')
if p.exit_if_idle():
    print('Papa says it will shutdown!')

WARNING: If another process connects to Papa before the connection is closed, Papa will remain open. The exit_if_idle command will drop the connection if it returns True so this is a very narrow window of opportunity for failure.