Tuesday, August 19, 2008

A Little Command Line Love

One of the things I do in my "spare" time is work on building web applications that will (hopefully) earn some spare money on the side without too much maintenance on my part. Those who have read The Four Hour Work Week will recognize this as my "muse" business.

In working on these web apps, I needed a place to host them, so I went with WebFaction due to their excellent support for TurboGears. I'm using a shared hosting environment with WebFaction, so my usual method of getting stuff up can't involve any scripts in /etc/init.d like I'd use at work, so I used to do the old nohup python start-appname.py >> output.log 2>&1 & to "daemonize" the process and then ps -furick446 to figure out which processes I needed to kill/restart when updating code. This was irritatingly verbose, so I figured I'd build a "userspace daemonizer" which I'll describe below.

The requirements of my daemonizer were pretty straightforward:

  • I should be able to add/remove services from the command line with a minimum amount of typing

  • I should be able to start/stop/restart any service just by listing it by name (no more ps -furick446)

  • It should do "real" daemonization (fork, fork, dup stuff if you're familiar with it) rather than the nohup garbage I'd been using

  • It should perform some sort of verification or status check on processes to make sure that they're successfully started/stopped/running/etc.

The first (and most questionable) decision I made was to go with SQLAlchemy as my service database. I say this is questionable because I ended up with only one table and one client that uses the database, so a fully relational model is really overkill here. Anyway, here's the SQLAlchemy setup code I used. It's pretty simple, and represents kind of the "lowest common denominator" of SQLAlchemy usage. (I've also included all the imports required for the whole shebang at the top of the file.)

#!/usr/bin/env python2.5
import os, sys, signal, shlex, time
from optparse import OptionParser

from sqlalchemy import *
from sqlalchemy.orm import *

os.path.join(HOME, 'server.sqlite')))

DBURI='sqlite:///' + DBFILE

engine = create_engine(DBURI)
metadata = MetaData(engine)
session = scoped_session(sessionmaker(bind=engine))

process = Table(
'process', metadata,
Column('name', String(255), primary_key=True),
Column('pid', Integer),
Column('command_line', String(255)),
Column('working_directory', String(255)),
Column('stdin', String(255), default='/dev/null'),
Column('stdout', String(255), default='/dev/null'),
Column('stderr', String(255), default='/dev/null'))

class Process(object):
def __repr__(self):
return '%s(PID %s, WD %s): %s < %s >> %s 2>> %s' % (
self.name, self.pid, self.working_directory,
self.command_line, self.stdin, self.stdout, self.stderr)
session.mapper(Process, process)

So, for all of you out there who wonder how to use SQLAlchemy outside of a web framework, there you go. The idea here is that every service has a row in the process table, and every service can optionally redirect its stdin/stdout/stderr to/from files. If a process is running, it will have a pid. You can also specify the startup directory of each process. That pretty much covers the data model.

My next task was to figure out how to invoke this from the command line. I decided to make all the commands of the form python server.py command [options] [service]. Actually, the only command that takes any options is the add command, so I definitely over-engineered here, but I wanted to be able to specify a global list of options shared by all commands to be used with an optparse option parser. So I built the following dictionary:

'working-directory':(['-w', '--working-directory'],
dict(dest='ws', default=HOME,
help='Set working directory to DIR',
'stdin':(['-i', '--stdin'],
dict(dest='stdin', default='/dev/null',
help='Use FILE as stdin', metavar='FILE')),
'stdout':(['-o', '--stdout'],
dict(dest='stdout', default='/dev/null',
help='Use FILE as stdout', metavar='FILE')),
'stderr':(['-e', '--stderr'],
dict(dest='stderr', default='/dev/null',
help='Use FILE as stderr', metavar='FILE')),

The next task was to specify the commands. I'm lazy and I like decorators, so I decided that new commands should be as easy as possible to write. My add command, for instance, is the most complex, and it looks like this:

def add(service, command, working_directory,
stdin, stdout, stderr):
p = Process(name=service,

That's pretty simple. Of course, as you might have guessed, the Command decorator isn't all that simple. Its responsibilities are as follows:

  • Add the function name to a command list so I can get a list of commands from the command line

  • Build an optparse parser based on the OPTPARSE_OPTIONS dict and the named arguments to the function

  • Wrap the function in a new function with a signature like function(args) so it can be called with sys.argv[2:]

  • In the wrapper function, use the optparse parser to initialize the argument list to the function and then call with a named argument dict

So with all that explanation, here's the code:

class Command(object):
commands = {}
def __init__(self, func):
Command.commands[func.__name__] = self
options = self.get_options(func)
self.func = func
self.optparse_options = []
self.optparse_option_names = []
self.positional_options = []
for o in options:
positional_string = ' '.join(
'<%s>' % o for o in self.positional_options)
self.parser = OptionParser('%%prog [options] %s' % positional_string,
for args, kwargs in self.optparse_options:
self.parser.add_option(*args, **kwargs)

def get_options(self, func):
code = func.func_code
return [ vn for vn in code.co_varnames[:code.co_argcount] ]

def run(klass, args):
if args and args[0] in klass.commands:
return klass.commands[args[0]](args[1:])
print 'Unrecognized command'
print 'Acceptable commands:'
for name in klass.commands:
print ' -', name

def __call__(self, args):
(opts,a) = self.parser.parse_args(args)
if len(a) != len(self.positional_options):
self.parser.error('Wrong number of arguments')
o = {}
for name in self.optparse_option_names:
o[name] = getattr(opts, name, None)
for name, value in zip(self.positional_options, a):
o[name] = value
return self.func(**o)

Once I've built the decorator, the commands are fairly straightforward (as the add command shows). Here are the "short" commands:

def initialize():
print 'Service database initialized'

def list():
q = Process.query()
if q.count():
for p in Process.query():
print p
print 'No processes'

def add(service, command, working_directory,
stdin, stdout, stderr):
p = Process(name=service,

def remove(service):
p = Process.query.get(service)
print 'Removed %s from the service list' % service

def status(service):
p = Process.query.get(service)
print p

Starting and stopping processes is a little more complex, but not too bad. First off, I needed a way to determine if a process was running. I didn't want to parse the ps -furick446 results, so I send a unix signal 0 to the PID (which doesn't do anything to the receiving process). If there's an exception, the process is either not running or not owned by me. So here's the is_running code:

def is_running(pid):
os.kill(pid, 0)
return True
except Exception, ex:
return False

I also need a daemonizer function that will start, daemonize, and set the PID of a process object:

def daemonize(p):
pid = os.fork()
if pid: return # exit first parent
pid = os.fork()
if pid:
Process.query.get(p.name).pid = pid
sys.exit(1) # exit second parent
si = open(p.stdin, 'r')
so = open(p.stdout, 'a+')
se = open(p.stderr, 'a+', 0)
os.dup2(si.fileno(), sys.stdin.fileno())
os.dup2(so.fileno(), sys.stdout.fileno())
os.dup2(se.fileno(), sys.stderr.fileno())
args = shlex.split(p.command_line.encode('utf-8'))
os.execvp(args[0], args)

Once these are defined, I can start, stop, and restart processes fairly simply:

def start(service):
print 'Starting %s ...' % service
p = Process.query.get(service)
if p.pid is not None:
if is_running(p.pid):
print '... %s already running with PID %s' % (
service, p.pid)
p.pid = 0
print '... started %s' % service

def stop(service):
print 'Stopping %s ...' % service
p = Process.query.get(service)
if not p.pid or not is_running(p.pid):
print '... service %s is already stopped' % service
p.pid = None
for retry in range(5):
print '... sending SIGTERM to %s' % p.pid
os.kill(p.pid, signal.SIGTERM)
if not is_running(p.pid):
print '... sending SIGKILL to %s' % p.pid
os.kill(p.pid, signal.SIGKILL)
if is_running(p.pid):
print '... process %s could not be killed' % p.pid
p.pid = None
print '... %s is stopped' % service

def restart(service):

Finally, I need to hook the script up and make sure it runs from the command line:

def main():

if __name__ == '__main__':

And there you have it! A userspace daemonizer that lets you manage an arbitrary number of services. It is definitely overkill in many ways, but hopefully the sharing the process of building it will be as educational to you as it was to me.

No comments:

Post a Comment