Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

552 lines (479 sloc) 29.389 kb
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Copyright (C) 2009-2012:
# Gabes Jean, naparuba@gmail.com
# Gerhard Lausser, Gerhard.Lausser@consol.de
# Gregory Starck, g.starck@gmail.com
# Hartmut Goebel, h.goebel@goebel-consult.de
#
# This file is part of Shinken.
#
# Shinken is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# Shinken is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
#
# You should have received a copy of the GNU Affero General Public License
# along with Shinken. If not, see <http://www.gnu.org/licenses/>.
"""
This is the class of the dispatcher. Its role is to dispatch
configurations to other elements like schedulers, reactionner,
pollers, receivers and brokers. It is responsible for hight avaibility part. If an
element dies and the element type has a spare, it sends the config of the
dead one to the spare
"""
import time
import random
import itertools
from shinken.util import alive_then_spare_then_deads
from shinken.log import logger
# Always initialize random :)
random.seed()
# Dispatcher Class
class Dispatcher:
# Load all elements, set them as not assigned
# and add them to elements, so loop will be easier :)
def __init__(self, conf, arbiter):
self.arbiter = arbiter
# Pointer to the whole conf
self.conf = conf
self.realms = conf.realms
# Direct pointer to importants elements for us
for sat_type in ('arbiters', 'schedulers', 'reactionners', 'brokers', 'receivers', 'pollers'):
setattr(self, sat_type, getattr(self.conf, sat_type))
# for each satellite, we look if current arbiter have a specific satellitemap value setted for this satellite
# if so, we give this map to the satellite (used to build satellite URI later)
if arbiter is None:
continue
key = sat_type[:-1] + '_name' # i.e: schedulers -> scheduler_name
for satellite in getattr(self, sat_type):
sat_name = getattr(satellite, key)
satellite.set_arbiter_satellitemap(arbiter.satellitemap.get(sat_name, {}))
self.dispatch_queue = {'schedulers': [], 'reactionners': [],
'brokers': [], 'pollers': [], 'receivers': []}
self.elements = [] # all elements, sched and satellites
self.satellites = [] # only satellites not schedulers
for cfg in self.conf.confs.values():
cfg.is_assigned = False
cfg.assigned_to = None
# We try to remeber each "push", so we
# can know with configuration ids+flavor
# if a satellite already got it or not :)
cfg.push_flavor = 0
# Add satellites in the good lists
self.elements.extend(self.schedulers)
# Others are in 2 lists
self.elements.extend(self.reactionners)
self.satellites.extend(self.reactionners)
self.elements.extend(self.pollers)
self.satellites.extend(self.pollers)
self.elements.extend(self.brokers)
self.satellites.extend(self.brokers)
self.elements.extend(self.receivers)
self.satellites.extend(self.receivers)
# Some flag about dispatch need or not
self.dispatch_ok = False
self.first_dispatch_done = False
# Prepare the satellites confs
for satellite in self.satellites:
satellite.prepare_for_conf()
# Some properties must be given to satellites from global
# configuration, like the max_plugins_output_length to pollers
parameters = {'max_plugins_output_length': self.conf.max_plugins_output_length}
for poller in self.pollers:
poller.add_global_conf_parameters(parameters)
# Reset need_conf for all schedulers.
for s in self.schedulers:
s.need_conf = True
# Same for receivers
for rec in self.receivers:
rec.need_conf = True
# checks alive elements
def check_alive(self):
for elt in self.elements:
#print "Updating elements", elt.get_name(), elt.__dict__
elt.update_infos()
# Not alive needs new need_conf
# and spare too if they do not have already a conf
# REF: doc/shinken-scheduler-lost.png (1)
if not elt.alive or hasattr(elt, 'conf') and elt.conf is None:
elt.need_conf = True
for arb in self.arbiters:
# If not me, but not the master too
if arb != self.arbiter and arb.spare:
arb.update_infos()
#print "Arb", arb.get_name(), "alive?", arb.alive, arb.__dict__
# Check if all active items are still alive
# the result goes into self.dispatch_ok
# TODO: finish need conf
def check_dispatch(self):
# Check if the other arbiter has a conf, but only if I am a master
for arb in self.arbiters:
# If not me and I'm a master
if arb != self.arbiter and self.arbiter and not self.arbiter.spare:
if not arb.have_conf(self.conf.magic_hash):
arb.put_conf(self.conf.whole_conf_pack)
else:
# Ok, it already has the conf. I remember that
# it does not have to run, I'm still alive!
arb.do_not_run()
# We check for confs to be dispatched on alive scheds. If not dispatched, need dispatch :)
# and if dipatch on a failed node, remove the association, and need a new disaptch
for r in self.realms:
for cfg_id in r.confs:
push_flavor = r.confs[cfg_id].push_flavor
sched = r.confs[cfg_id].assigned_to
if sched is None:
if self.first_dispatch_done:
logger.info("Scheduler configuration %d is unmanaged!!" % cfg_id)
self.dispatch_ok = False
else:
if not sched.alive:
self.dispatch_ok = False # so we ask a new dispatching
logger.warning("Scheduler %s had the configuration %d but is dead, I am not happy." % (sched.get_name(), cfg_id))
sched.conf.assigned_to = None
sched.conf.is_assigned = False
sched.conf.push_flavor = 0
sched.push_flavor = 0
sched.conf = None
# Maybe the scheduler restarts, so is alive but without the conf we think it was managing
# so ask it what it is really managing, and if not, put the conf unassigned
if not sched.do_i_manage(cfg_id, push_flavor):
self.dispatch_ok = False # so we ask a new dispatching
logger.warning("Scheduler %s did not managed its configuration %d,I am not happy." % (sched.get_name(), cfg_id))
if sched.conf:
sched.conf.assigned_to = None
sched.conf.is_assigned = False
sched.conf.push_flavor = 0
sched.push_flavor = 0
sched.need_conf = True
sched.conf = None
# Else: ok the conf is managed by a living scheduler
# Maybe satelite are alive, but do not have a cfg yet.
# I think so. It is not good. I ask a global redispatch for
# the cfg_id I think is not corectly dispatched.
for r in self.realms:
for cfg_id in r.confs:
push_flavor = r.confs[cfg_id].push_flavor
try:
for kind in ('reactionner', 'poller', 'broker', 'receiver'):
# We must have the good number of satellite or we are not happy
# So we are sure to raise a dispatch every loop a satellite is missing
if len(r.to_satellites_managed_by[kind][cfg_id]) < r.get_nb_of_must_have_satellites(kind):
logger.warning("Missing satellite %s for configuration %d:" % (kind, cfg_id))
# TODO: less violent! Must only resent to who need?
# must be caught by satellite who sees that it already has the conf (hash)
# and do nothing
self.dispatch_ok = False # so we will redispatch all
r.to_satellites_need_dispatch[kind][cfg_id] = True
r.to_satellites_managed_by[kind][cfg_id] = []
for satellite in r.to_satellites_managed_by[kind][cfg_id]:
# Maybe the sat was marked as not alive, but still in
# to_satellites_managed_by. That means that a new dispatch
# is needed
# Or maybe it is alive but I thought that this reactionner managed the conf
# and it doesn't. I ask a full redispatch of these cfg for both cases
# DBG:
#print "What I manage", satellite.get_name(), satellite.what_i_managed()
#try:
# satellite.reachable and cfg_id not in satellite.what_i_managed()
#except TypeError, exp:
# print "DBG: ERROR: (%s) for satellite %s" % (exp, satellite.__dict__)
# satellite.reachable = False
#wim = satellite.managed_confs # what_i_managed()
#print "%s [%s]Look at what manage the %s %s (alive? %s, reachable? %s): %s (look for %s)" % (int(time.time()), r.get_name(), kind, satellite.get_name(), satellite.alive, satellite.reachable, wim, cfg_id)
#print satellite.alive, satellite.reachable, satellite.do_i_manage(cfg_id, push_flavor)
if not satellite.alive or (satellite.reachable and not satellite.do_i_manage(cfg_id, push_flavor)):
logger.warning('[%s] The %s %s seems to be down, I must re-dispatch its role to someone else.' % (r.get_name(), kind, satellite.get_name()))
self.dispatch_ok = False # so we will redispatch all
r.to_satellites_need_dispatch[kind][cfg_id] = True
r.to_satellites_managed_by[kind][cfg_id] = []
# At the first pass, there is no cfg_id in to_satellites_managed_by
except KeyError:
pass
# Look for receivers. If they got conf, it's ok, if not, need a simple
# conf
for r in self.realms:
for rec in r.receivers:
# If the receiver does not have a conf, must got one :)
if rec.reachable and not rec.got_conf():
self.dispatch_ok = False # so we will redispatch all
rec.need_conf = True
# Imagine a world where... oh no, wait...
# Imagine a master got the conf and the network is down
# a spare takes it (good :) ). Like the Empire, the master
# strikes back! It was still alive! (like Elvis). It still got conf
# and is running! not good!
# Bad dispatch: a link that has a conf but I do not allow this
# so I ask it to wait a new conf and stop kidding.
def check_bad_dispatch(self):
for elt in self.elements:
if hasattr(elt, 'conf'):
# If element has a conf, I do not care, it's a good dispatch
# If dead: I do not ask it something, it won't respond..
if elt.conf is None and elt.reachable:
# print "Ask", elt.get_name() , 'if it got conf'
if elt.have_conf():
logger.warning("The element %s have a conf and should not have one! I ask it to idle now" % elt.get_name())
elt.active = False
elt.wait_new_conf()
# I do not care about order not send or not. If not,
# The next loop will resent it
# else:
# print "No conf"
# I ask satellites which sched_id they manage. If I do not agree, I ask
# them to remove it
for satellite in self.satellites:
kind = satellite.get_my_type()
if satellite.reachable:
cfg_ids = satellite.managed_confs # what_i_managed()
# I do not care about satellites that do nothing, they already
# do what I want :)
if len(cfg_ids) != 0:
id_to_delete = []
for cfg_id in cfg_ids:
# DBG print kind, ":", satellite.get_name(), "manage cfg id:", cfg_id
# Ok, we search for realms that have the conf
for r in self.realms:
if cfg_id in r.confs:
# Ok we've got the realm, we check its to_satellites_managed_by
# to see if reactionner is in. If not, we remove he sched_id for it
if not satellite in r.to_satellites_managed_by[kind][cfg_id]:
id_to_delete.append(cfg_id)
# Maybe we removed all cfg_id of this reactionner
# We can put it idle, no active and wait_new_conf
if len(id_to_delete) == len(cfg_ids):
satellite.active = False
logger.info("I ask %s to wait a new conf" % satellite.get_name())
satellite.wait_new_conf()
else:
# It is not fully idle, just less cfg
for id in id_to_delete:
logger.info("I ask to remove configuration N%d from %s" % (id, satellite.get_name()))
satellite.remove_from_conf(id)
# Make an ORDERED list of schedulers so we can
# send them conf in this order for a specific realm
def get_scheduler_ordered_list(self, r):
# get scheds, alive and no spare first
scheds = []
for s in r.schedulers:
scheds.append(s)
# now the spare scheds of higher realms
# they are after the sched of realm, so
# they will be used after the spare of
# the realm
for higher_r in r.higher_realms:
for s in higher_r.schedulers:
if s.spare:
scheds.append(s)
# Now we sort the scheds so we take master, then spare
# the dead, but we do not care about them
scheds.sort(alive_then_spare_then_deads)
scheds.reverse() # pop is last, I need first
print_sched = [s.get_name() for s in scheds]
print_sched.reverse()
return scheds
# Manage the dispatch
# REF: doc/shinken-conf-dispatching.png (3)
def dispatch(self):
# Ok, we pass at least one time in dispatch, so now errors are True errors
self.first_dispatch_done = True
# If no needed to dispatch, do not dispatch :)
if not self.dispatch_ok:
for r in self.realms:
conf_to_dispatch = [cfg for cfg in r.confs.values() if not cfg.is_assigned]
nb_conf = len(conf_to_dispatch)
if nb_conf > 0:
logger.info("Dispatching Realm %s" % r.get_name())
logger.info('[%s] Dispatching %d/%d configurations' % (r.get_name(), nb_conf, len(r.confs)))
# Now we get in scheds all scheduler of this realm and upper so
# we will send them conf (in this order)
scheds = self.get_scheduler_ordered_list(r)
if nb_conf > 0:
print_string = '[%s] Schedulers order: %s' % (r.get_name(), ','.join([s.get_name() for s in scheds]))
logger.info(print_string)
# Try to send only for alive members
scheds = [s for s in scheds if s.alive]
# Now we do the real job
# every_one_need_conf = False
for conf in conf_to_dispatch:
logger.info('[%s] Dispatching configuration %s' % (r.get_name(), conf.id))
# If there is no alive schedulers, not good...
if len(scheds) == 0:
logger.info('[%s] but there a no alive schedulers in this realm!' % r.get_name())
# we need to loop until the conf is assigned
# or when there are no more schedulers available
while True:
try:
sched = scheds.pop()
except IndexError: # No more schedulers.. not good, no loop
# need_loop = False
# The conf does not need to be dispatch
cfg_id = conf.id
for kind in ('reactionner', 'poller', 'broker', 'receiver'):
r.to_satellites[kind][cfg_id] = None
r.to_satellites_need_dispatch[kind][cfg_id] = False
r.to_satellites_managed_by[kind][cfg_id] = []
break
logger.info('[%s] Trying to send conf %d to scheduler %s' % (r.get_name(), conf.id, sched.get_name()))
if not sched.need_conf:
logger.info('[%s] The scheduler %s do not need conf, sorry' % (r.get_name(), sched.get_name()))
continue
# We tag conf with the instance_name = scheduler_name
instance_name = sched.scheduler_name
# We give this configuraton a new 'flavor'
conf.push_flavor = random.randint(1, 1000000)
# REF: doc/shinken-conf-dispatching.png (3)
# REF: doc/shinken-scheduler-lost.png (2)
override_conf = sched.get_override_configuration()
satellites_for_sched = r.get_satellites_links_for_scheduler()
s_conf = r.serialized_confs[conf.id]
# Prepare the conf before sending it
conf_package = {'conf': s_conf, 'override_conf': override_conf,
'modules': sched.modules, 'satellites': satellites_for_sched,
'instance_name': sched.scheduler_name, 'push_flavor': conf.push_flavor,
'skip_initial_broks': sched.skip_initial_broks,
}
t1 = time.time()
is_sent = sched.put_conf(conf_package)
logger.debug("Conf is sent in %d" % (time.time() - t1))
if not is_sent:
logger.warning('[%s] configuration dispatching error for scheduler %s' % (r.get_name(), sched.get_name()))
continue
logger.info('[%s] Dispatch OK of conf in scheduler %s' % (r.get_name(), sched.get_name()))
sched.conf = conf
sched.push_flavor = conf.push_flavor
sched.need_conf = False
conf.is_assigned = True
conf.assigned_to = sched
# We update all data for this scheduler
sched.managed_confs = {conf.id: conf.push_flavor}
# Now we generate the conf for satellites:
cfg_id = conf.id
for kind in ('reactionner', 'poller', 'broker', 'receiver'):
r.to_satellites[kind][cfg_id] = sched.give_satellite_cfg()
r.to_satellites_need_dispatch[kind][cfg_id] = True
r.to_satellites_managed_by[kind][cfg_id] = []
# Ok, the conf is dispatched, no more loop for this
# configuration
break
# We pop conf to dispatch, so it must be no more conf...
conf_to_dispatch = [cfg for cfg in self.conf.confs.values() if not cfg.is_assigned]
nb_missed = len(conf_to_dispatch)
if nb_missed > 0:
logger.warning("All schedulers configurations are not dispatched, %d are missing" % nb_missed)
else:
logger.info("OK, all schedulers configurations are dispatched :)")
self.dispatch_ok = True
# Sched without conf in a dispatch ok are set to no need_conf
# so they do not raise dispatch where no use
if self.dispatch_ok:
for sched in self.schedulers.items.values():
if sched.conf is None:
# print "Tagging sched", sched.get_name(), "so it do not ask anymore for conf"
sched.need_conf = False
arbiters_cfg = {}
for arb in self.arbiters:
arbiters_cfg[arb.id] = arb.give_satellite_cfg()
# We put the satellites conf with the "new" way so they see only what we want
for r in self.realms:
for cfg in r.confs.values():
cfg_id = cfg.id
# flavor if the push number of this configuration send to a scheduler
flavor = cfg.push_flavor
for kind in ('reactionner', 'poller', 'broker', 'receiver'):
if r.to_satellites_need_dispatch[kind][cfg_id]:
cfg_for_satellite_part = r.to_satellites[kind][cfg_id]
# make copies of potential_react list for sort
satellites = []
for satellite in r.get_potential_satellites_by_type(kind):
satellites.append(satellite)
satellites.sort(alive_then_spare_then_deads)
# Only keep alive Satellites and reachable ones
satellites = [s for s in satellites if s.alive and s.reachable]
# If we got a broker, we make the list to pop a new
# item first for each scheduler, so it will smooth the load
# But the spare must stay at the end ;)
if kind == "broker":
nospare = [s for s in satellites if not s.spare]
# Should look over the list, not over
if len(nospare) != 0:
idx = cfg_id % len(nospare)
#print "No spare", nospare
spares = [s for s in satellites if s.spare]
#print "Spare", spares
#print "Got 1", nospare[idx:]
#print "Got 2", nospare[:-idx+1]
new_satellites = nospare[idx:]
for _b in nospare[:-idx+1]:
if _b not in new_satellites:
new_satellites.append(_b)
#new_satellites.extend(nospare[:-idx+1])
#print "New satellites", cfg_id, new_satellites
#for s in new_satellites:
# print "New satellites", cfg_id, s.get_name()
satellites = new_satellites
satellites.extend(spares)
# Dump the order where we will send conf
satellite_string = "[%s] Dispatching %s satellite with order: " % (r.get_name(), kind)
for satellite in satellites:
satellite_string += '%s (spare:%s), ' % (satellite.get_name(), str(satellite.spare))
logger.info(satellite_string)
# Now we dispatch cfg to every one ask for it
nb_cfg_sent = 0
for satellite in satellites:
# Send only if we need, and if we can
if nb_cfg_sent < r.get_nb_of_must_have_satellites(kind) and satellite.alive:
satellite.cfg['schedulers'][cfg_id] = cfg_for_satellite_part
if satellite.manage_arbiters:
satellite.cfg['arbiters'] = arbiters_cfg
# Brokers should have poller/reactionners links too
if kind == "broker":
r.fill_broker_with_poller_reactionner_links(satellite)
is_sent = False
# Maybe this satellite already got this configuration, so skip it
if satellite.do_i_manage(cfg_id, flavor):
logger.info('[%s] Skipping configuration %d send to the %s %s: it already got it' % (r.get_name(), cfg_id, kind, satellite.get_name()))
is_sent = True
else: # ok, it really need it :)
logger.info('[%s] Trying to send configuration to %s %s' % (r.get_name(), kind, satellite.get_name()))
is_sent = satellite.put_conf(satellite.cfg)
if is_sent:
satellite.active = True
logger.info('[%s] Dispatch OK of configuration %s to %s %s' % (r.get_name(), cfg_id, kind, satellite.get_name()))
# We change the satellite configuration, update our data
satellite.known_conf_managed_push(cfg_id, flavor)
nb_cfg_sent += 1
r.to_satellites_managed_by[kind][cfg_id].append(satellite)
# If we got a broker, the conf_id must be sent to only ONE
# broker, so here it's done, we are happy.
if kind == "broker":
break
#If receiver, we must send the hostnames of this configuration
if kind == 'receiver':
hnames = [h.get_name() for h in cfg.hosts]
logger.debug("[%s] Sending %s hostnames to the receiver %s" % (r.get_name(), len(hnames), satellite.get_name()))
satellite.push_host_names(cfg_id, hnames)
# else:
# #I've got enough satellite, the next ones are considered spares
if nb_cfg_sent == r.get_nb_of_must_have_satellites(kind):
logger.info("[%s] OK, no more %s sent need" % (r.get_name(), kind))
r.to_satellites_need_dispatch[kind][cfg_id] = False
# And now we dispatch receivers. It's easier, they need ONE conf
# in all their life :)
for r in self.realms:
for rec in r.receivers:
if rec.need_conf:
logger.info('[%s] Trying to send configuration to receiver %s' % (r.get_name(), rec.get_name()))
is_sent = rec.put_conf(rec.cfg)
if is_sent:
rec.active = True
rec.need_conf = False
logger.info('[%s] Dispatch OK of configuration to receiver %s' % (r.get_name(), rec.get_name()))
else:
logger.error('[%s] Dispatching failed for receiver %s' % (r.get_name(), rec.get_name()))
Jump to Line
Something went wrong with that request. Please try again.