def prepare_config(self):
for key in self._pop_config_keys:
config.pop(key, None)
config.update(
{
# Default to a uniform retry delay of one second
"agent_http_retry_delay_offset": 0.1,
"agent_http_retry_delay_factor": 0,
"shutting_down": False,
"jobtypes": {},
"current_assignments": {},
"agent_id": uuid.uuid4(),
"agent_http_persistent_connections": False,
"agent_shutdown_timeout": 3,
"master": PYFARM_AGENT_MASTER,
"agent_hostname": os.urandom(self.RAND_LENGTH).encode("hex"),
"agent_ram": memory.total_ram(),
"agent_cpus": cpu.total_cpus(),
"agent_api_port": randint(10000, 50000),
"free_ram": memory.free_ram(),
"agent_time_offset": randint(-50, 50),
"state": choice(AgentState),
"start": time.time(),
"agent_pretty_json": False,
"agent_html_template_reload": True,
"agent_master_reannounce": randint(5, 15),
}
)
def get(self, **_):
# Get counts for child processes and grandchild processes
process = psutil.Process()
direct_child_processes = len(process.children(recursive=False))
all_child_processes = len(process.children(recursive=True))
grandchild_processes = all_child_processes - direct_child_processes
# Determine the last time we talked to the master (if ever)
contacted = config.master_contacted(update=False)
if isinstance(contacted, datetime): # pragma: no cover
contacted = datetime.utcnow() - contacted
# Determine the last time we announced ourselves to the
# master (if ever)
last_announce = config.get("last_announce", None)
if isinstance(last_announce, datetime): # pragma: no cover
last_announce = datetime.utcnow() - last_announce
data = {"state": config["state"],
"agent_hostname": config["agent_hostname"],
"free_ram": memory.free_ram(),
"agent_process_ram": memory.process_memory(),
"consumed_ram": memory.total_consumption(),
"child_processes": direct_child_processes,
"grandchild_processes": grandchild_processes,
"pids": config["pids"],
"agent_id": config["agent_id"],
"last_master_contact": contacted,
"last_announce": last_announce,
"agent_lock_file": config["agent_lock_file"],
"uptime": total_seconds(
timedelta(seconds=time.time() - config["start"])),
"jobs": list(config["jobtypes"].keys())}
if config["farm_name"]:
data["farm_name"] = config["farm_name"]
return dumps(data)
def callback_agent_id_set(
self, change_type, key, new_value, old_value, shutdown_events=True):
"""
When `agent_id` is created we need to:
* Register a shutdown event so that when the agent is told to
shutdown it will notify the master of a state change.
* Star the scheduled task manager
"""
if key == "agent_id" and change_type == config.CREATED \
and not self.register_shutdown_events:
if shutdown_events:
self.register_shutdown_events = True
# set the initial free_ram
config["free_ram"] = memory.free_ram()
config.master_contacted()
svclog.debug(
"`%s` was %s, adding system event trigger for shutdown",
key, change_type)
self.repeating_call(
config["agent_master_reannounce"], self.reannounce)
def post_shutdown_to_master(self):
"""
This method is called before the reactor shuts down and lets the
master know that the agent's state is now ``offline``
"""
# We're under the assumption that something's wrong with
# our code if we try to call this method before self.shutting_down
# is set.
assert self.shutting_down
yield self.post_shutdown_lock.acquire()
svclog.info("Informing master of shutdown")
# Because post_shutdown_to_master is blocking and needs to
# stop the reactor from finishing we perform the retry in-line
data = None
tries = 0
num_retry_errors = 0
response = None
timed_out = False
while True:
tries += 1
try:
response = yield post_direct(
self.agent_api(),
data={
"state": AgentState.OFFLINE,
"free_ram": memory.free_ram(),
"current_assignments": config["current_assignments"]})
except (ResponseNeverReceived, RequestTransmissionFailed) as error:
num_retry_errors += 1
if num_retry_errors > config["broken_connection_max_retry"]:
svclog.error(
"Failed to post shutdown to the master, "
"caught try-again errors %s times in a row.",
num_retry_errors)
break
elif self.shutdown_timeout < datetime.utcnow():
svclog.error("While posting shutdown to master, caught "
"%s. Shutdown timeout has been reached, not "
"retrying.",
error.__class__.__name__)
break
else:
svclog.debug("While posting shutdown to master, caught "
"%s. Retrying immediately.",
error.__class__.__name__)
# When we get a hard failure it could be an issue with the
# server, although it's unlikely, so we retry. Only retry
# for a set period of time though since the shutdown as a timeout
except Exception as failure:
if self.shutdown_timeout > datetime.utcnow():
delay = http_retry_delay()
svclog.warning(
"State update failed due to unhandled error: %s. "
"Retrying in %s seconds",
failure, delay)
# Wait for 'pause' to fire, introducing a delay
pause = Deferred()
reactor.callLater(delay, pause.callback, None)
yield pause
else:
timed_out = True
svclog.warning(
"State update failed due to unhandled error: %s. "
"Shutdown timeout reached, not retrying.",
failure)
break
else:
data = yield treq.json_content(response)
if response.code == NOT_FOUND:
svclog.warning(
"Agent %r no longer exists, cannot update state.",
config["agent_id"])
break
elif response.code == OK:
svclog.info(
"Agent %r has POSTed shutdown state change "
"successfully.",
config["agent_id"])
break
elif response.code >= INTERNAL_SERVER_ERROR:
if self.shutdown_timeout > datetime.utcnow():
delay = http_retry_delay()
svclog.warning(
"State update failed due to server error: %s. "
"Retrying in %s seconds.",
data, delay)
# Wait for 'pause' to fire, introducing a delay
pause = Deferred()
reactor.callLater(delay, pause.callback, None)
yield pause
else:
timed_out = True
svclog.warning(
"State update failed due to server error: %s. "
"Shutdown timeout reached, not retrying.",
data)
break
yield self.post_shutdown_lock.release()
extra_data = {
"response": response,
"timed_out": timed_out,
"tries": tries,
"retry_errors": num_retry_errors
}
if isinstance(data, dict):
data.update(extra_data)
else:
data = extra_data
returnValue(data)
def system_data(self, requery_timeoffset=False):
"""
Returns a dictionary of data containing information about the
agent. This is the information that is also passed along to
the master.
"""
# query the time offset and then cache it since
# this is typically a blocking operation
if config["agent_time_offset"] == "auto":
config["agent_time_offset"] = None
if requery_timeoffset or config["agent_time_offset"] is None:
ntplog.info(
"Querying ntp server %r for current time",
config["agent_ntp_server"])
ntp_client = NTPClient()
try:
pool_time = ntp_client.request(
config["agent_ntp_server"],
version=config["agent_ntp_server_version"])
except Exception as e:
ntplog.warning("Failed to determine network time: %s", e)
else:
config["agent_time_offset"] = \
int(pool_time.tx_time - time.time())
# format the offset for logging purposes
utcoffset = datetime.utcfromtimestamp(pool_time.tx_time)
iso_timestamp = utcoffset.isoformat()
ntplog.debug(
"network time: %s (local offset: %r)",
iso_timestamp, config["agent_time_offset"])
if config["agent_time_offset"] != 0:
ntplog.warning(
"Agent is %r second(s) off from ntp server at %r",
config["agent_time_offset"],
config["agent_ntp_server"])
data = {
"id": config["agent_id"],
"hostname": config["agent_hostname"],
"version": config.version,
"os_class": system.operating_system(),
"os_fullname": platform(),
"ram": int(config["agent_ram"]),
"cpus": config["agent_cpus"],
"cpu_name": cpu.cpu_name(),
"port": config["agent_api_port"],
"free_ram": memory.free_ram(),
"time_offset": config["agent_time_offset"] or 0,
"state": config["state"],
"mac_addresses": list(network.mac_addresses()),
"current_assignments": config.get(
"current_assignments", {}), # may not be set yet
"disks": disks.disks(as_dict=True)
}
try:
gpu_names = graphics.graphics_cards()
data["gpus"] = gpu_names
except graphics.GPULookupError:
pass
if "remote_ip" in config:
data.update(remote_ip=config["remote_ip"])
if config["farm_name"]:
data["farm_name"] = config["farm_name"]
return data
def reannounce(self, force=False):
"""
Method which is used to periodically contact the master. This
method is generally called as part of a scheduled task.
"""
# Attempt to acquire the reannounce lock but fail after 70%
# of the total time between reannouncements elapses. This should
# help prevent an accumulation of requests in the event the master
# is having issues.
try:
yield self.reannounce_lock.acquire(
config["agent_master_reannounce"] * .70
)
except utility.LockTimeoutError:
svclog.debug("Timed out while waiting to acquire reannounce_lock")
returnValue(None)
if not self.should_reannounce() and not force:
yield self.reannounce_lock.release()
returnValue(None)
svclog.debug("Announcing %s to master", config["agent_hostname"])
data = None
num_retry_errors = 0
while True: # for retries
try:
response = yield post_direct(
self.agent_api(),
data={
"state": config["state"],
"current_assignments": config.get(
"current_assignments", {} # may not be set yet
),
"free_ram": memory.free_ram(),
"disks": disks.disks(as_dict=True)
}
)
except (ResponseNeverReceived, RequestTransmissionFailed) as error:
num_retry_errors += 1
if num_retry_errors > config["broken_connection_max_retry"]:
svclog.error(
"Failed to announce self to the master, "
"caught try-again type errors %s times in a row.",
num_retry_errors)
break
else:
svclog.debug("While announcing self to master, caught "
"%s. Retrying immediately.",
error.__class__.__name__)
except Exception as error:
if force:
delay = http_retry_delay()
svclog.error(
"Failed to announce self to the master: %s. Will "
"retry in %s seconds.", error, delay)
deferred = Deferred()
reactor.callLater(delay, deferred.callback, None)
yield deferred
else:
# Don't retry because reannounce is called periodically
svclog.error(
"Failed to announce self to the master: %s. This "
"request will not be retried.", error)
break
else:
data = yield treq.json_content(response)
if response.code == OK:
config.master_contacted(announcement=True)
svclog.info("Announced self to the master server.")
break
elif response.code >= INTERNAL_SERVER_ERROR:
if not self.shutting_down:
delay = http_retry_delay()
svclog.warning(
"Could not announce self to the master server, "
"internal server error: %s. Retrying in %s "
"seconds.", data, delay)
deferred = Deferred()
reactor.callLater(delay, deferred.callback, None)
yield deferred
else:
svclog.warning(
"Could not announce to master. Not retrying "
"because of pending shutdown.")
break
elif response.code == NOT_FOUND:
svclog.warning("The master says it does not know about our "
"agent id. Posting as a new agent.")
yield self.post_agent_to_master()
break
# If this is a client problem retrying the request
# is unlikely to fix the issue so we stop here
elif response.code >= BAD_REQUEST:
svclog.error(
"Failed to announce self to the master, bad "
"request: %s. This request will not be retried.",
data)
break
else:
svclog.error(
"Unhandled error when posting self to the "
"master: %s (code: %s). This request will not be "
"retried.", data, response.code)
break
yield self.reannounce_lock.release()
returnValue(data)
def test_ramfree(self):
v1 = convert.bytetomb(psutil.virtual_memory().available)
v2 = memory.free_ram()
self.assertApproximates(v1, v2, 5)
def test_ramused(self):
v1 = memory.total_ram() - memory.free_ram()
v2 = memory.used_ram()
self.assertApproximates(v1, v2, 5)
def post(self, **kwargs):
if request_from_master(kwargs["request"]):
config.master_contacted()
request = kwargs["request"]
request_data = kwargs["data"]
# First, get the resources we have *right now*. In some cases
# this means using the functions in pyfarm.core.sysinfo because
# entries in `config` could be slightly out of sync with the system.
memory_free = free_ram()
cpus = config["agent_cpus"]
requires_ram = request_data["job"].get("ram")
requires_cpus = request_data["job"].get("cpus")
if ("agent_id" in request_data and
request_data["agent_id"] != config["agent_id"]):
logger.error("Wrong agent_id in assignment: %s. Our id is %s",
request_data["agent_id"], config["agent_id"])
return (
dumps({"error": "You have the wrong agent. "
"I am %s." % config["agent_id"],
"agent_id": config["agent_id"]}),
BAD_REQUEST
)
elif self.agent.reannounce_lock.locked:
logger.warning("Temporarily rejecting assignment because we "
"are in the middle of a reannounce.")
return (
dumps({"error": "Agent cannot accept assignments because of a "
"reannounce in progress. Try again shortly."}),
SERVICE_UNAVAILABLE
)
elif self.agent.shutting_down:
logger.error("Rejecting assignment because the agent is in the "
"process of shutting down.")
return (
dumps({"error": "Agent cannot accept assignments because it is "
"shutting down."}),
SERVICE_UNAVAILABLE
)
elif "restart_requested" in config \
and config["restart_requested"] is True:
logger.error("Rejecting assignment because of scheduled restart.")
return (
dumps({"error": "Agent cannot accept assignments because of a "
"pending restart."}),
SERVICE_UNAVAILABLE
)
elif "agent_id" not in config:
logger.error(
"Agent has not yet connected to the master or `agent_id` "
"has not been set yet.")
return (
dumps({"error": "agent_id has not been set in the config"}),
SERVICE_UNAVAILABLE
)
# Do we have enough ram?
elif requires_ram is not None and requires_ram > memory_free:
logger.error(
"Task %s requires %sMB of ram, this agent has %sMB free. "
"Rejecting Task %s.",
request_data["job"]["id"], requires_ram, memory_free,
request_data["job"]["id"])
config["free_ram"] = memory_free
return (
dumps({"error": "Not enough ram",
"agent_ram": memory_free,
"requires_ram": requires_ram}),
BAD_REQUEST
)
# Do we have enough cpus (count wise)?
elif requires_cpus is not None and requires_cpus > cpus:
logger.error(
"Task %s requires %s CPUs, this agent has %s CPUs. "
"Rejecting Task %s.",
request_data["job"]["id"], requires_cpus, cpus,
request_data["job"]["id"])
return (
dumps({"error": "Not enough cpus",
"agent_cpus": cpus,
"requires_cpus": requires_cpus}),
BAD_REQUEST
)
new_task_ids = set(task["id"] for task in request_data["tasks"])
for assignment in config["current_assignments"].itervalues():
existing_task_ids = set(x["id"] for x in assignment["tasks"])
# If the assignment is identical to one we already have
if existing_task_ids == new_task_ids:
logger.debug(
"Ignoring repeated assignment of the same batch")
return dumps({"id": assignment["id"]}), ACCEPTED
# If there is only a partial overlap
elif existing_task_ids & new_task_ids:
logger.error("Rejecting assignment with partial overlap with "
"existing assignment.")
unknown_task_ids = new_task_ids - existing_task_ids
return (
dumps({"error": "Partial overlap of tasks",
"rejected_task_ids": list(unknown_task_ids)}),
CONFLICT
)
if not config["agent_allow_sharing"]:
for jobtype in config["jobtypes"].itervalues():
num_finished_tasks = (len(jobtype.finished_tasks) +
len(jobtype.failed_tasks))
if len(jobtype.assignment["tasks"]) > num_finished_tasks:
logger.error("Rejecting an assignment that would require "
"agent sharing")
return (
dumps({
"error": "Agent does not allow multiple "
"assignments",
"rejected_task_ids": list(new_task_ids)}),
CONFLICT
)
assignment_uuid = uuid4()
request_data.update(id=assignment_uuid)
config["current_assignments"][assignment_uuid] = request_data
logger.debug("Accepted assignment %s: %r",
assignment_uuid, request_data)
logger.info("Accept assignment from job %s with %s tasks",
request_data["job"]["title"], len(request_data["tasks"]))
def assignment_failed(result, assign_id):
logger.error(
"Assignment %s failed, result: %r, removing.", assign_id, result)
logger.error(result.getTraceback())
if (len(config["current_assignments"]) <= 1 and
not self.agent.shutting_down):
config["state"] = AgentState.ONLINE
self.agent.reannounce(force=True)
# Do not mark the assignment as failed if the reason for failing
# was that we ran out of disk space
failed = not isinstance(result.value, InsufficientSpaceError)
assignment = config["current_assignments"].pop(assign_id)
if "jobtype" in assignment:
jobtype_id = assignment["jobtype"].pop("id", None)
if jobtype_id:
instance = config["jobtypes"].pop(jobtype_id, None)
instance.stop(
assignment_failed=failed,
avoid_reassignment=not failed,
error="Error in jobtype: %r. "
"Traceback: %s" % (result,
traceback.format_exc()))
def assignment_started(_, assign_id):
logger.debug("Assignment %s has started", assign_id)
config["state"] = AgentState.RUNNING
self.agent.reannounce(force=True)
def remove_assignment(_, assign_id):
assignment = config["current_assignments"].pop(assign_id)
if "jobtype" in assignment:
jobtype_id = assignment["jobtype"].pop("id", None)
if jobtype_id:
config["jobtypes"].pop(jobtype_id, None)
def assignment_stopped(_, assign_id):
logger.debug("Assignment %s has stopped", assign_id)
if (len(config["current_assignments"]) <= 1 and
not self.agent.shutting_down):
config["state"] = AgentState.ONLINE
self.agent.reannounce(force=True)
assignment = config["current_assignments"][assign_id]
if "jobtype" in assignment:
jobtype_id = assignment["jobtype"].pop("id", None)
if jobtype_id:
jobtype = config["jobtypes"].pop(jobtype_id, None)
updates_deferred = DeferredList(
jobtype.task_update_deferreds)
updates_deferred.addBoth(remove_assignment, assign_id)
else:
config["current_assignments"].pop(assign_id)
def restart_if_necessary(_): # pragma: no cover
if "restart_requested" in config and config["restart_requested"]:
stopping = config["agent"].stop()
stopping.addCallbacks(lambda _: reactor.stop(),
lambda _: reactor.stop())
def load_jobtype_failed(result, assign_id):
logger.error(
"Loading jobtype for assignment %s failed, removing.", assign_id)
traceback = result.getTraceback()
logger.debug("Got traceback")
logger.error(traceback)
assignment = config["current_assignments"].pop(assign_id)
# Mark all tasks as failed on master and set an error message
logger.debug("Marking tasks in assignment as failed")
def post_update(post_url, post_data, task, delay=0):
post_func = partial(post, post_url, data=post_data,
callback=lambda x: result_callback(
post_url, post_data, task, x),
errback=lambda x: error_callback(
post_url, post_data, task, x))
reactor.callLater(delay, post_func)
def result_callback(cburl, cbdata, task, response):
if 500 <= response.code < 600:
logger.error(
"Error while marking task %s as failed on master, "
"retrying", task["id"])
post_update(cburl, cbdata, task, delay=http_retry_delay())
elif response.code != OK:
logger.error(
"Could not mark task %s as failed, server response "
"code was %s", task["id"], response.code)
else:
logger.info(
"Marked task %s as failed on master", task["id"])
def error_callback(cburl, cbdata, task, failure_reason):
logger.error(
"Error while marking task %s as failed, retrying",
task["id"], failure_reason)
post_update(cburl, cbdata, task, delay=http_retry_delay())
for task in assignment["tasks"]:
url = "%s/jobs/%s/tasks/%s" % (
config["master_api"], assignment["job"]["id"], task["id"])
data = {
"state": WorkState.FAILED,
"last_error": traceback}
post_update(url, data, task)
# If the loading was partially successful for some reason, there
# might already be an entry for this jobtype in the config.
# Remove it if it exists.
if "jobtype" in assignment:
jobtype_id = assignment["jobtype"].pop("id", None)
if jobtype_id:
config["jobtypes"].pop(jobtype_id, None)
def loaded_jobtype(jobtype_class, assign_id):
# TODO: report error to master
if hasattr(jobtype_class, "getTraceback"):
logger.error(jobtype_class.getTraceback())
return
# TODO: add call to prepare_for_job
# TODO: add call to spawn_persistent_process
# Instance the job type and pass in the assignment data.
instance = jobtype_class(request_data)
if not isinstance(instance, JobType):
raise TypeError(
"Expected a subclass of "
"pyfarm.jobtypes.core.jobtype.JobType")
# TODO: add callback to cleanup_after_job
# TODO: add callback to stop persistent process
try:
started_deferred, stopped_deferred = instance._start()
started_deferred.addCallback(assignment_started, assign_id)
started_deferred.addErrback(assignment_failed, assign_id)
stopped_deferred.addCallback(assignment_stopped, assign_id)
stopped_deferred.addErrback(assignment_failed, assign_id)
stopped_deferred.addBoth(restart_if_necessary)
stopped_deferred.addBoth(
lambda *args: instance._remove_tempdirs())
stopped_deferred.addBoth(
lambda *args: instance._close_logs())
stopped_deferred.addBoth(
lambda *args: instance._upload_logfile())
except Exception as e:
logger.error("Error on starting jobtype, stopping it now. "
"Error was: %r. Traceback: %s", e,
traceback.format_exc())
instance.stop(assignment_failed=True,
error="Error while loading jobtype: %r. "
"Traceback: %s" %
(e, traceback.format_exc()))
assignment = config["current_assignments"].pop(assign_id)
if "jobtype" in assignment:
jobtype_id = assignment["jobtype"].pop("id", None)
if jobtype_id:
config["jobtypes"].pop(jobtype_id, None)
# Load the job type then pass the class along to the
# callback. No errback here because all the errors
# are handled internally in this case.
jobtype_loader = JobType.load(request_data)
jobtype_loader.addCallback(loaded_jobtype, assignment_uuid)
jobtype_loader.addErrback(load_jobtype_failed, assignment_uuid)
return dumps({"id": assignment_uuid}), ACCEPTED
