def test_job_not_blocking_inventory_update(default_instance_group, job_template_factory, inventory_source_factory): objects = job_template_factory('jt', organization='org1', project='proj', inventory='inv', credential='cred', jobs=["job"]) job = objects.jobs["job"] job.instance_group = default_instance_group job.status = "running" job.save() with mock.patch("awx.main.scheduler.TaskManager.start_task"): task_manager = TaskManager() task_manager._schedule() inv = objects.inventory inv_source = inventory_source_factory("ec2") inv_source.source = "ec2" inv.inventory_sources.add(inv_source) inventory_update = inv_source.create_inventory_update() inventory_update.instance_group = default_instance_group inventory_update.status = "pending" inventory_update.save() assert not task_manager.is_job_blocked(inventory_update) dependency_graph = DependencyGraph(None) dependency_graph.add_job(job) assert not dependency_graph.is_job_blocked(inventory_update)
def test_job_not_blocking_project_update(default_instance_group, job_template_factory): objects = job_template_factory('jt', organization='org1', project='proj', inventory='inv', credential='cred', jobs=["job"]) job = objects.jobs["job"] job.instance_group = default_instance_group job.status = "running" job.save() with mock.patch("awx.main.scheduler.TaskManager.start_task"): task_manager = TaskManager() task_manager._schedule() proj = objects.project project_update = proj.create_project_update() project_update.instance_group = default_instance_group project_update.status = "pending" project_update.save() assert not task_manager.job_blocked_by(project_update) dependency_graph = DependencyGraph() dependency_graph.add_job(job) assert not dependency_graph.task_blocked_by(project_update)
class TaskManager: def __init__(self): """ Do NOT put database queries or other potentially expensive operations in the task manager init. The task manager object is created every time a job is created, transitions state, and every 30 seconds on each tower node. More often then not, the object is destroyed quickly because the NOOP case is hit. The NOOP case is short-circuit logic. If the task manager realizes that another instance of the task manager is already running, then it short-circuits and decides not to run. """ self.graph = dict() # start task limit indicates how many pending jobs can be started on this # .schedule() run. Starting jobs is expensive, and there is code in place to reap # the task manager after 5 minutes. At scale, the task manager can easily take more than # 5 minutes to start pending jobs. If this limit is reached, pending jobs # will no longer be started and will be started on the next task manager cycle. self.start_task_limit = settings.START_TASK_LIMIT self.time_delta_job_explanation = timedelta(seconds=30) def after_lock_init(self): """ Init AFTER we know this instance of the task manager will run because the lock is acquired. """ instances = Instance.objects.filter(hostname__isnull=False, enabled=True).exclude(node_type='hop') self.real_instances = {i.hostname: i for i in instances} self.controlplane_ig = None self.dependency_graph = DependencyGraph() instances_partial = [ SimpleNamespace( obj=instance, node_type=instance.node_type, remaining_capacity=instance.remaining_capacity, capacity=instance.capacity, jobs_running=instance.jobs_running, hostname=instance.hostname, ) for instance in instances ] instances_by_hostname = {i.hostname: i for i in instances_partial} for rampart_group in InstanceGroup.objects.prefetch_related('instances'): if rampart_group.name == settings.DEFAULT_CONTROL_PLANE_QUEUE_NAME: self.controlplane_ig = rampart_group self.graph[rampart_group.name] = dict( instances=[ instances_by_hostname[instance.hostname] for instance in rampart_group.instances.all() if instance.hostname in instances_by_hostname ], ) def job_blocked_by(self, task): # TODO: I'm not happy with this, I think blocking behavior should be decided outside of the dependency graph # in the old task manager this was handled as a method on each task object outside of the graph and # probably has the side effect of cutting down *a lot* of the logic from this task manager class blocked_by = self.dependency_graph.task_blocked_by(task) if blocked_by: return blocked_by if not task.dependent_jobs_finished(): blocked_by = task.dependent_jobs.first() if blocked_by: return blocked_by return None def get_tasks(self, status_list=('pending', 'waiting', 'running')): jobs = [j for j in Job.objects.filter(status__in=status_list).prefetch_related('instance_group')] inventory_updates_qs = ( InventoryUpdate.objects.filter(status__in=status_list).exclude(source='file').prefetch_related('inventory_source', 'instance_group') ) inventory_updates = [i for i in inventory_updates_qs] # Notice the job_type='check': we want to prevent implicit project updates from blocking our jobs. project_updates = [p for p in ProjectUpdate.objects.filter(status__in=status_list, job_type='check').prefetch_related('instance_group')] system_jobs = [s for s in SystemJob.objects.filter(status__in=status_list).prefetch_related('instance_group')] ad_hoc_commands = [a for a in AdHocCommand.objects.filter(status__in=status_list).prefetch_related('instance_group')] workflow_jobs = [w for w in WorkflowJob.objects.filter(status__in=status_list)] all_tasks = sorted(jobs + project_updates + inventory_updates + system_jobs + ad_hoc_commands + workflow_jobs, key=lambda task: task.created) return all_tasks def get_running_workflow_jobs(self): graph_workflow_jobs = [wf for wf in WorkflowJob.objects.filter(status='running')] return graph_workflow_jobs def get_inventory_source_tasks(self, all_sorted_tasks): inventory_ids = set() for task in all_sorted_tasks: if isinstance(task, Job): inventory_ids.add(task.inventory_id) return [invsrc for invsrc in InventorySource.objects.filter(inventory_id__in=inventory_ids, update_on_launch=True)] def spawn_workflow_graph_jobs(self, workflow_jobs): for workflow_job in workflow_jobs: if workflow_job.cancel_flag: logger.debug('Not spawning jobs for %s because it is pending cancelation.', workflow_job.log_format) continue dag = WorkflowDAG(workflow_job) spawn_nodes = dag.bfs_nodes_to_run() if spawn_nodes: logger.debug('Spawning jobs for %s', workflow_job.log_format) else: logger.debug('No nodes to spawn for %s', workflow_job.log_format) for spawn_node in spawn_nodes: if spawn_node.unified_job_template is None: continue kv = spawn_node.get_job_kwargs() job = spawn_node.unified_job_template.create_unified_job(**kv) spawn_node.job = job spawn_node.save() logger.debug('Spawned %s in %s for node %s', job.log_format, workflow_job.log_format, spawn_node.pk) can_start = True if isinstance(spawn_node.unified_job_template, WorkflowJobTemplate): workflow_ancestors = job.get_ancestor_workflows() if spawn_node.unified_job_template in set(workflow_ancestors): can_start = False logger.info( 'Refusing to start recursive workflow-in-workflow id={}, wfjt={}, ancestors={}'.format( job.id, spawn_node.unified_job_template.pk, [wa.pk for wa in workflow_ancestors] ) ) display_list = [spawn_node.unified_job_template] + workflow_ancestors job.job_explanation = gettext_noop( "Workflow Job spawned from workflow could not start because it " "would result in recursion (spawn order, most recent first: {})" ).format(', '.join(['<{}>'.format(tmp) for tmp in display_list])) else: logger.debug( 'Starting workflow-in-workflow id={}, wfjt={}, ancestors={}'.format( job.id, spawn_node.unified_job_template.pk, [wa.pk for wa in workflow_ancestors] ) ) if not job._resources_sufficient_for_launch(): can_start = False job.job_explanation = gettext_noop( "Job spawned from workflow could not start because it " "was missing a related resource such as project or inventory" ) if can_start: if workflow_job.start_args: start_args = json.loads(decrypt_field(workflow_job, 'start_args')) else: start_args = {} can_start = job.signal_start(**start_args) if not can_start: job.job_explanation = gettext_noop( "Job spawned from workflow could not start because it " "was not in the right state or required manual credentials" ) if not can_start: job.status = 'failed' job.save(update_fields=['status', 'job_explanation']) job.websocket_emit_status('failed') # TODO: should we emit a status on the socket here similar to tasks.py awx_periodic_scheduler() ? # emit_websocket_notification('/socket.io/jobs', '', dict(id=)) def process_finished_workflow_jobs(self, workflow_jobs): result = [] for workflow_job in workflow_jobs: dag = WorkflowDAG(workflow_job) status_changed = False if workflow_job.cancel_flag: workflow_job.workflow_nodes.filter(do_not_run=False, job__isnull=True).update(do_not_run=True) logger.debug('Canceling spawned jobs of %s due to cancel flag.', workflow_job.log_format) cancel_finished = dag.cancel_node_jobs() if cancel_finished: logger.info('Marking %s as canceled, all spawned jobs have concluded.', workflow_job.log_format) workflow_job.status = 'canceled' workflow_job.start_args = '' # blank field to remove encrypted passwords workflow_job.save(update_fields=['status', 'start_args']) status_changed = True else: workflow_nodes = dag.mark_dnr_nodes() for n in workflow_nodes: n.save(update_fields=['do_not_run']) is_done = dag.is_workflow_done() if not is_done: continue has_failed, reason = dag.has_workflow_failed() logger.debug('Marking %s as %s.', workflow_job.log_format, 'failed' if has_failed else 'successful') result.append(workflow_job.id) new_status = 'failed' if has_failed else 'successful' logger.debug("Transitioning {} to {} status.".format(workflow_job.log_format, new_status)) update_fields = ['status', 'start_args'] workflow_job.status = new_status if reason: logger.info(f'Workflow job {workflow_job.id} failed due to reason: {reason}') workflow_job.job_explanation = gettext_noop("No error handling paths found, marking workflow as failed") update_fields.append('job_explanation') workflow_job.start_args = '' # blank field to remove encrypted passwords workflow_job.save(update_fields=update_fields) status_changed = True if status_changed: workflow_job.websocket_emit_status(workflow_job.status) # Operations whose queries rely on modifications made during the atomic scheduling session workflow_job.send_notification_templates('succeeded' if workflow_job.status == 'successful' else 'failed') if workflow_job.spawned_by_workflow: schedule_task_manager() return result def start_task(self, task, rampart_group, dependent_tasks=None, instance=None): self.start_task_limit -= 1 if self.start_task_limit == 0: # schedule another run immediately after this task manager schedule_task_manager() from awx.main.tasks.system import handle_work_error, handle_work_success dependent_tasks = dependent_tasks or [] task_actual = { 'type': get_type_for_model(type(task)), 'id': task.id, } dependencies = [{'type': get_type_for_model(type(t)), 'id': t.id} for t in dependent_tasks] task.status = 'waiting' (start_status, opts) = task.pre_start() if not start_status: task.status = 'failed' if task.job_explanation: task.job_explanation += ' ' task.job_explanation += 'Task failed pre-start check.' task.save() # TODO: run error handler to fail sub-tasks and send notifications else: if type(task) is WorkflowJob: task.status = 'running' task.send_notification_templates('running') logger.debug('Transitioning %s to running status.', task.log_format) schedule_task_manager() # at this point we already have control/execution nodes selected for the following cases else: task.instance_group = rampart_group execution_node_msg = f' and execution node {task.execution_node}' if task.execution_node else '' logger.debug( f'Submitting job {task.log_format} controlled by {task.controller_node} to instance group {rampart_group.name}{execution_node_msg}.' ) with disable_activity_stream(): task.celery_task_id = str(uuid.uuid4()) task.save() task.log_lifecycle("waiting") def post_commit(): if task.status != 'failed' and type(task) is not WorkflowJob: # Before task is dispatched, ensure that job_event partitions exist create_partition(task.event_class._meta.db_table, start=task.created) task_cls = task._get_task_class() task_cls.apply_async( [task.pk], opts, queue=task.get_queue_name(), uuid=task.celery_task_id, callbacks=[{'task': handle_work_success.name, 'kwargs': {'task_actual': task_actual}}], errbacks=[{'task': handle_work_error.name, 'args': [task.celery_task_id], 'kwargs': {'subtasks': [task_actual] + dependencies}}], ) task.websocket_emit_status(task.status) # adds to on_commit connection.on_commit(post_commit) def process_running_tasks(self, running_tasks): for task in running_tasks: self.dependency_graph.add_job(task) def create_project_update(self, task): project_task = Project.objects.get(id=task.project_id).create_project_update(_eager_fields=dict(launch_type='dependency')) # Project created 1 seconds behind project_task.created = task.created - timedelta(seconds=1) project_task.status = 'pending' project_task.save() logger.debug('Spawned {} as dependency of {}'.format(project_task.log_format, task.log_format)) return project_task def create_inventory_update(self, task, inventory_source_task): inventory_task = InventorySource.objects.get(id=inventory_source_task.id).create_inventory_update(_eager_fields=dict(launch_type='dependency')) inventory_task.created = task.created - timedelta(seconds=2) inventory_task.status = 'pending' inventory_task.save() logger.debug('Spawned {} as dependency of {}'.format(inventory_task.log_format, task.log_format)) # inventory_sources = self.get_inventory_source_tasks([task]) # self.process_inventory_sources(inventory_sources) return inventory_task def capture_chain_failure_dependencies(self, task, dependencies): with disable_activity_stream(): task.dependent_jobs.add(*dependencies) for dep in dependencies: # Add task + all deps except self dep.dependent_jobs.add(*([task] + [d for d in dependencies if d != dep])) def get_latest_inventory_update(self, inventory_source): latest_inventory_update = InventoryUpdate.objects.filter(inventory_source=inventory_source).order_by("-created") if not latest_inventory_update.exists(): return None return latest_inventory_update.first() def should_update_inventory_source(self, job, latest_inventory_update): now = tz_now() if latest_inventory_update is None: return True ''' If there's already a inventory update utilizing this job that's about to run then we don't need to create one ''' if latest_inventory_update.status in ['waiting', 'pending', 'running']: return False timeout_seconds = timedelta(seconds=latest_inventory_update.inventory_source.update_cache_timeout) if (latest_inventory_update.finished + timeout_seconds) < now: return True if latest_inventory_update.inventory_source.update_on_launch is True and latest_inventory_update.status in ['failed', 'canceled', 'error']: return True return False def get_latest_project_update(self, job): latest_project_update = ProjectUpdate.objects.filter(project=job.project, job_type='check').order_by("-created") if not latest_project_update.exists(): return None return latest_project_update.first() def should_update_related_project(self, job, latest_project_update): now = tz_now() if latest_project_update is None: return True if latest_project_update.status in ['failed', 'canceled']: return True ''' If there's already a project update utilizing this job that's about to run then we don't need to create one ''' if latest_project_update.status in ['waiting', 'pending', 'running']: return False ''' If the latest project update has a created time == job_created_time-1 then consider the project update found. This is so we don't enter an infinite loop of updating the project when cache timeout is 0. ''' if ( latest_project_update.project.scm_update_cache_timeout == 0 and latest_project_update.launch_type == 'dependency' and latest_project_update.created == job.created - timedelta(seconds=1) ): return False ''' Normal Cache Timeout Logic ''' timeout_seconds = timedelta(seconds=latest_project_update.project.scm_update_cache_timeout) if (latest_project_update.finished + timeout_seconds) < now: return True return False def generate_dependencies(self, undeped_tasks): created_dependencies = [] for task in undeped_tasks: task.log_lifecycle("acknowledged") dependencies = [] if not type(task) is Job: continue # TODO: Can remove task.project None check after scan-job-default-playbook is removed if task.project is not None and task.project.scm_update_on_launch is True: latest_project_update = self.get_latest_project_update(task) if self.should_update_related_project(task, latest_project_update): project_task = self.create_project_update(task) created_dependencies.append(project_task) dependencies.append(project_task) else: dependencies.append(latest_project_update) # Inventory created 2 seconds behind job try: start_args = json.loads(decrypt_field(task, field_name="start_args")) except ValueError: start_args = dict() for inventory_source in [invsrc for invsrc in self.all_inventory_sources if invsrc.inventory == task.inventory]: if "inventory_sources_already_updated" in start_args and inventory_source.id in start_args['inventory_sources_already_updated']: continue if not inventory_source.update_on_launch: continue latest_inventory_update = self.get_latest_inventory_update(inventory_source) if self.should_update_inventory_source(task, latest_inventory_update): inventory_task = self.create_inventory_update(task, inventory_source) created_dependencies.append(inventory_task) dependencies.append(inventory_task) else: dependencies.append(latest_inventory_update) if len(dependencies) > 0: self.capture_chain_failure_dependencies(task, dependencies) UnifiedJob.objects.filter(pk__in=[task.pk for task in undeped_tasks]).update(dependencies_processed=True) return created_dependencies def process_pending_tasks(self, pending_tasks): running_workflow_templates = {wf.unified_job_template_id for wf in self.get_running_workflow_jobs()} tasks_to_update_job_explanation = [] for task in pending_tasks: if self.start_task_limit <= 0: break blocked_by = self.job_blocked_by(task) if blocked_by: task.log_lifecycle("blocked", blocked_by=blocked_by) job_explanation = gettext_noop(f"waiting for {blocked_by._meta.model_name}-{blocked_by.id} to finish") if task.job_explanation != job_explanation: if task.created < (tz_now() - self.time_delta_job_explanation): task.job_explanation = job_explanation tasks_to_update_job_explanation.append(task) continue found_acceptable_queue = False preferred_instance_groups = task.preferred_instance_groups if isinstance(task, WorkflowJob): if task.unified_job_template_id in running_workflow_templates: if not task.allow_simultaneous: logger.debug("{} is blocked from running, workflow already running".format(task.log_format)) continue else: running_workflow_templates.add(task.unified_job_template_id) self.start_task(task, None, task.get_jobs_fail_chain(), None) continue # Determine if there is control capacity for the task if task.capacity_type == 'control': control_impact = task.task_impact + settings.AWX_CONTROL_NODE_TASK_IMPACT else: control_impact = settings.AWX_CONTROL_NODE_TASK_IMPACT control_instance = InstanceGroup.fit_task_to_most_remaining_capacity_instance( task, self.graph[settings.DEFAULT_CONTROL_PLANE_QUEUE_NAME]['instances'], impact=control_impact, capacity_type='control' ) if not control_instance: self.task_needs_capacity(task, tasks_to_update_job_explanation) logger.debug(f"Skipping task {task.log_format} in pending, not enough capacity left on controlplane to control new tasks") continue task.controller_node = control_instance.hostname # All task.capacity_type == 'control' jobs should run on control plane, no need to loop over instance groups if task.capacity_type == 'control': task.execution_node = control_instance.hostname control_instance.remaining_capacity = max(0, control_instance.remaining_capacity - control_impact) control_instance.jobs_running += 1 self.dependency_graph.add_job(task) execution_instance = self.real_instances[control_instance.hostname] task.log_lifecycle("controller_node_chosen") task.log_lifecycle("execution_node_chosen") self.start_task(task, self.controlplane_ig, task.get_jobs_fail_chain(), execution_instance) found_acceptable_queue = True continue for rampart_group in preferred_instance_groups: if rampart_group.is_container_group: control_instance.jobs_running += 1 self.dependency_graph.add_job(task) self.start_task(task, rampart_group, task.get_jobs_fail_chain(), None) found_acceptable_queue = True break # TODO: remove this after we have confidence that OCP control nodes are reporting node_type=control if settings.IS_K8S and task.capacity_type == 'execution': logger.debug("Skipping group {}, task cannot run on control plane".format(rampart_group.name)) continue # at this point we know the instance group is NOT a container group # because if it was, it would have started the task and broke out of the loop. execution_instance = InstanceGroup.fit_task_to_most_remaining_capacity_instance( task, self.graph[rampart_group.name]['instances'], add_hybrid_control_cost=True ) or InstanceGroup.find_largest_idle_instance(self.graph[rampart_group.name]['instances'], capacity_type=task.capacity_type) if execution_instance: task.execution_node = execution_instance.hostname # If our execution instance is a hybrid, prefer to do control tasks there as well. if execution_instance.node_type == 'hybrid': control_instance = execution_instance task.controller_node = execution_instance.hostname control_instance.remaining_capacity = max(0, control_instance.remaining_capacity - settings.AWX_CONTROL_NODE_TASK_IMPACT) task.log_lifecycle("controller_node_chosen") if control_instance != execution_instance: control_instance.jobs_running += 1 execution_instance.remaining_capacity = max(0, execution_instance.remaining_capacity - task.task_impact) execution_instance.jobs_running += 1 task.log_lifecycle("execution_node_chosen") logger.debug( "Starting {} in group {} instance {} (remaining_capacity={})".format( task.log_format, rampart_group.name, execution_instance.hostname, execution_instance.remaining_capacity ) ) execution_instance = self.real_instances[execution_instance.hostname] self.dependency_graph.add_job(task) self.start_task(task, rampart_group, task.get_jobs_fail_chain(), execution_instance) found_acceptable_queue = True break else: logger.debug( "No instance available in group {} to run job {} w/ capacity requirement {}".format( rampart_group.name, task.log_format, task.task_impact ) ) if not found_acceptable_queue: self.task_needs_capacity(task, tasks_to_update_job_explanation) UnifiedJob.objects.bulk_update(tasks_to_update_job_explanation, ['job_explanation']) def task_needs_capacity(self, task, tasks_to_update_job_explanation): task.log_lifecycle("needs_capacity") job_explanation = gettext_noop("This job is not ready to start because there is not enough available capacity.") if task.job_explanation != job_explanation: if task.created < (tz_now() - self.time_delta_job_explanation): # Many launched jobs are immediately blocked, but most blocks will resolve in a few seconds. # Therefore we should only update the job_explanation after some time has elapsed to # prevent excessive task saves. task.job_explanation = job_explanation tasks_to_update_job_explanation.append(task) logger.debug("{} couldn't be scheduled on graph, waiting for next cycle".format(task.log_format)) def timeout_approval_node(self): workflow_approvals = WorkflowApproval.objects.filter(status='pending') now = tz_now() for task in workflow_approvals: approval_timeout_seconds = timedelta(seconds=task.timeout) if task.timeout == 0: continue if (now - task.created) >= approval_timeout_seconds: timeout_message = _("The approval node {name} ({pk}) has expired after {timeout} seconds.").format( name=task.name, pk=task.pk, timeout=task.timeout ) logger.warning(timeout_message) task.timed_out = True task.status = 'failed' task.send_approval_notification('timed_out') task.websocket_emit_status(task.status) task.job_explanation = timeout_message task.save(update_fields=['status', 'job_explanation', 'timed_out']) def reap_jobs_from_orphaned_instances(self): # discover jobs that are in running state but aren't on an execution node # that we know about; this is a fairly rare event, but it can occur if you, # for example, SQL backup an awx install with running jobs and restore it # elsewhere for j in UnifiedJob.objects.filter( status__in=['pending', 'waiting', 'running'], ).exclude(execution_node__in=Instance.objects.exclude(node_type='hop').values_list('hostname', flat=True)): if j.execution_node and not j.is_container_group_task: logger.error(f'{j.execution_node} is not a registered instance; reaping {j.log_format}') reap_job(j, 'failed') def process_tasks(self, all_sorted_tasks): running_tasks = [t for t in all_sorted_tasks if t.status in ['waiting', 'running']] self.process_running_tasks(running_tasks) pending_tasks = [t for t in all_sorted_tasks if t.status == 'pending'] undeped_tasks = [t for t in pending_tasks if not t.dependencies_processed] dependencies = self.generate_dependencies(undeped_tasks) self.process_pending_tasks(dependencies) self.process_pending_tasks(pending_tasks) def _schedule(self): finished_wfjs = [] all_sorted_tasks = self.get_tasks() self.after_lock_init() if len(all_sorted_tasks) > 0: # TODO: Deal with # latest_project_updates = self.get_latest_project_update_tasks(all_sorted_tasks) # self.process_latest_project_updates(latest_project_updates) # latest_inventory_updates = self.get_latest_inventory_update_tasks(all_sorted_tasks) # self.process_latest_inventory_updates(latest_inventory_updates) self.all_inventory_sources = self.get_inventory_source_tasks(all_sorted_tasks) running_workflow_tasks = self.get_running_workflow_jobs() finished_wfjs = self.process_finished_workflow_jobs(running_workflow_tasks) previously_running_workflow_tasks = running_workflow_tasks running_workflow_tasks = [] for workflow_job in previously_running_workflow_tasks: if workflow_job.status == 'running': running_workflow_tasks.append(workflow_job) else: logger.debug('Removed %s from job spawning consideration.', workflow_job.log_format) self.spawn_workflow_graph_jobs(running_workflow_tasks) self.timeout_approval_node() self.reap_jobs_from_orphaned_instances() self.process_tasks(all_sorted_tasks) return finished_wfjs def schedule(self): # Lock with advisory_lock('task_manager_lock', wait=False) as acquired: with transaction.atomic(): if acquired is False: logger.debug("Not running scheduler, another task holds lock") return logger.debug("Starting Scheduler") with task_manager_bulk_reschedule(): self._schedule() logger.debug("Finishing Scheduler")
class TaskManager(TaskBase): def __init__(self): """ Do NOT put database queries or other potentially expensive operations in the task manager init. The task manager object is created every time a job is created, transitions state, and every 30 seconds on each tower node. More often then not, the object is destroyed quickly because the NOOP case is hit. The NOOP case is short-circuit logic. If the task manager realizes that another instance of the task manager is already running, then it short-circuits and decides not to run. """ # start task limit indicates how many pending jobs can be started on this # .schedule() run. Starting jobs is expensive, and there is code in place to reap # the task manager after 5 minutes. At scale, the task manager can easily take more than # 5 minutes to start pending jobs. If this limit is reached, pending jobs # will no longer be started and will be started on the next task manager cycle. self.time_delta_job_explanation = timedelta(seconds=30) super().__init__(prefix="task_manager") def after_lock_init(self): """ Init AFTER we know this instance of the task manager will run because the lock is acquired. """ self.dependency_graph = DependencyGraph() self.instances = TaskManagerInstances(self.all_tasks) self.instance_groups = TaskManagerInstanceGroups( instances_by_hostname=self.instances) self.controlplane_ig = self.instance_groups.controlplane_ig def job_blocked_by(self, task): # TODO: I'm not happy with this, I think blocking behavior should be decided outside of the dependency graph # in the old task manager this was handled as a method on each task object outside of the graph and # probably has the side effect of cutting down *a lot* of the logic from this task manager class blocked_by = self.dependency_graph.task_blocked_by(task) if blocked_by: return blocked_by for dep in task.dependent_jobs.all(): if dep.status in ACTIVE_STATES: return dep # if we detect a failed or error dependency, go ahead and fail this # task. The errback on the dependency takes some time to trigger, # and we don't want the task to enter running state if its # dependency has failed or errored. elif dep.status in ("error", "failed"): task.status = 'failed' task.job_explanation = 'Previous Task Failed: {"job_type": "%s", "job_name": "%s", "job_id": "%s"}' % ( get_type_for_model(type(dep)), dep.name, dep.id, ) task.save(update_fields=['status', 'job_explanation']) task.websocket_emit_status('failed') return dep return None @timeit def start_task(self, task, instance_group, dependent_tasks=None, instance=None): self.dependency_graph.add_job(task) self.subsystem_metrics.inc(f"{self.prefix}_tasks_started", 1) self.start_task_limit -= 1 if self.start_task_limit == 0: # schedule another run immediately after this task manager ScheduleTaskManager().schedule() from awx.main.tasks.system import handle_work_error, handle_work_success # update capacity for control node and execution node if task.controller_node: self.instances[task.controller_node].consume_capacity( settings.AWX_CONTROL_NODE_TASK_IMPACT) if task.execution_node: self.instances[task.execution_node].consume_capacity( task.task_impact) dependent_tasks = dependent_tasks or [] task_actual = { 'type': get_type_for_model(type(task)), 'id': task.id, } dependencies = [{ 'type': get_type_for_model(type(t)), 'id': t.id } for t in dependent_tasks] task.status = 'waiting' (start_status, opts) = task.pre_start() if not start_status: task.status = 'failed' if task.job_explanation: task.job_explanation += ' ' task.job_explanation += 'Task failed pre-start check.' task.save() # TODO: run error handler to fail sub-tasks and send notifications else: if type(task) is WorkflowJob: task.status = 'running' task.send_notification_templates('running') logger.debug('Transitioning %s to running status.', task.log_format) # Call this to ensure Workflow nodes get spawned in timely manner ScheduleWorkflowManager().schedule() # at this point we already have control/execution nodes selected for the following cases else: task.instance_group = instance_group execution_node_msg = f' and execution node {task.execution_node}' if task.execution_node else '' logger.debug( f'Submitting job {task.log_format} controlled by {task.controller_node} to instance group {instance_group.name}{execution_node_msg}.' ) with disable_activity_stream(): task.celery_task_id = str(uuid.uuid4()) task.save() task.log_lifecycle("waiting") def post_commit(): if task.status != 'failed' and type(task) is not WorkflowJob: # Before task is dispatched, ensure that job_event partitions exist create_partition(task.event_class._meta.db_table, start=task.created) task_cls = task._get_task_class() task_cls.apply_async( [task.pk], opts, queue=task.get_queue_name(), uuid=task.celery_task_id, callbacks=[{ 'task': handle_work_success.name, 'kwargs': { 'task_actual': task_actual } }], errbacks=[{ 'task': handle_work_error.name, 'args': [task.celery_task_id], 'kwargs': { 'subtasks': [task_actual] + dependencies } }], ) task.websocket_emit_status(task.status) # adds to on_commit connection.on_commit(post_commit) @timeit def process_running_tasks(self, running_tasks): for task in running_tasks: if type(task) is WorkflowJob: ScheduleWorkflowManager().schedule() self.dependency_graph.add_job(task) @timeit def process_pending_tasks(self, pending_tasks): tasks_to_update_job_explanation = [] for task in pending_tasks: if self.start_task_limit <= 0: break if self.timed_out(): logger.warning( "Task manager has reached time out while processing pending jobs, exiting loop early" ) break blocked_by = self.job_blocked_by(task) if blocked_by: self.subsystem_metrics.inc(f"{self.prefix}_tasks_blocked", 1) task.log_lifecycle("blocked", blocked_by=blocked_by) job_explanation = gettext_noop( f"waiting for {blocked_by._meta.model_name}-{blocked_by.id} to finish" ) if task.job_explanation != job_explanation: if task.created < (tz_now() - self.time_delta_job_explanation): task.job_explanation = job_explanation tasks_to_update_job_explanation.append(task) continue if isinstance(task, WorkflowJob): # Previously we were tracking allow_simultaneous blocking both here and in DependencyGraph. # Double check that using just the DependencyGraph works for Workflows and Sliced Jobs. self.start_task(task, None, task.get_jobs_fail_chain(), None) continue found_acceptable_queue = False preferred_instance_groups = self.instance_groups.get_instance_groups_from_task_cache( task) # Determine if there is control capacity for the task if task.capacity_type == 'control': control_impact = task.task_impact + settings.AWX_CONTROL_NODE_TASK_IMPACT else: control_impact = settings.AWX_CONTROL_NODE_TASK_IMPACT control_instance = self.instance_groups.fit_task_to_most_remaining_capacity_instance( task, instance_group_name=settings.DEFAULT_CONTROL_PLANE_QUEUE_NAME, impact=control_impact, capacity_type='control') if not control_instance: self.task_needs_capacity(task, tasks_to_update_job_explanation) logger.debug( f"Skipping task {task.log_format} in pending, not enough capacity left on controlplane to control new tasks" ) continue task.controller_node = control_instance.hostname # All task.capacity_type == 'control' jobs should run on control plane, no need to loop over instance groups if task.capacity_type == 'control': task.execution_node = control_instance.hostname execution_instance = self.instances[ control_instance.hostname].obj task.log_lifecycle("controller_node_chosen") task.log_lifecycle("execution_node_chosen") self.start_task(task, self.controlplane_ig, task.get_jobs_fail_chain(), execution_instance) found_acceptable_queue = True continue for instance_group in preferred_instance_groups: if instance_group.is_container_group: self.start_task(task, instance_group, task.get_jobs_fail_chain(), None) found_acceptable_queue = True break # TODO: remove this after we have confidence that OCP control nodes are reporting node_type=control if settings.IS_K8S and task.capacity_type == 'execution': logger.debug( "Skipping group {}, task cannot run on control plane". format(instance_group.name)) continue # at this point we know the instance group is NOT a container group # because if it was, it would have started the task and broke out of the loop. execution_instance = self.instance_groups.fit_task_to_most_remaining_capacity_instance( task, instance_group_name=instance_group.name, add_hybrid_control_cost=True ) or self.instance_groups.find_largest_idle_instance( instance_group_name=instance_group.name, capacity_type=task.capacity_type) if execution_instance: task.execution_node = execution_instance.hostname # If our execution instance is a hybrid, prefer to do control tasks there as well. if execution_instance.node_type == 'hybrid': control_instance = execution_instance task.controller_node = execution_instance.hostname task.log_lifecycle("controller_node_chosen") task.log_lifecycle("execution_node_chosen") logger.debug( "Starting {} in group {} instance {} (remaining_capacity={})" .format(task.log_format, instance_group.name, execution_instance.hostname, execution_instance.remaining_capacity)) execution_instance = self.instances[ execution_instance.hostname].obj self.start_task(task, instance_group, task.get_jobs_fail_chain(), execution_instance) found_acceptable_queue = True break else: logger.debug( "No instance available in group {} to run job {} w/ capacity requirement {}" .format(instance_group.name, task.log_format, task.task_impact)) if not found_acceptable_queue: self.task_needs_capacity(task, tasks_to_update_job_explanation) UnifiedJob.objects.bulk_update(tasks_to_update_job_explanation, ['job_explanation']) def task_needs_capacity(self, task, tasks_to_update_job_explanation): task.log_lifecycle("needs_capacity") job_explanation = gettext_noop( "This job is not ready to start because there is not enough available capacity." ) if task.job_explanation != job_explanation: if task.created < (tz_now() - self.time_delta_job_explanation): # Many launched jobs are immediately blocked, but most blocks will resolve in a few seconds. # Therefore we should only update the job_explanation after some time has elapsed to # prevent excessive task saves. task.job_explanation = job_explanation tasks_to_update_job_explanation.append(task) logger.debug( "{} couldn't be scheduled on graph, waiting for next cycle".format( task.log_format)) def reap_jobs_from_orphaned_instances(self): # discover jobs that are in running state but aren't on an execution node # that we know about; this is a fairly rare event, but it can occur if you, # for example, SQL backup an awx install with running jobs and restore it # elsewhere for j in UnifiedJob.objects.filter(status__in=[ 'pending', 'waiting', 'running' ], ).exclude(execution_node__in=Instance.objects.exclude( node_type='hop').values_list('hostname', flat=True)): if j.execution_node and not j.is_container_group_task: logger.error( f'{j.execution_node} is not a registered instance; reaping {j.log_format}' ) reap_job(j, 'failed') def process_tasks(self): running_tasks = [ t for t in self.all_tasks if t.status in ['waiting', 'running'] ] self.process_running_tasks(running_tasks) self.subsystem_metrics.inc(f"{self.prefix}_running_processed", len(running_tasks)) pending_tasks = [t for t in self.all_tasks if t.status == 'pending'] self.process_pending_tasks(pending_tasks) self.subsystem_metrics.inc(f"{self.prefix}_pending_processed", len(pending_tasks)) def timeout_approval_node(self, task): if self.timed_out(): logger.warning( "Task manager has reached time out while processing approval nodes, exiting loop early" ) # Do not process any more workflow approval nodes. Stop here. # Maybe we should schedule another TaskManager run return timeout_message = _( "The approval node {name} ({pk}) has expired after {timeout} seconds." ).format(name=task.name, pk=task.pk, timeout=task.timeout) logger.warning(timeout_message) task.timed_out = True task.status = 'failed' task.send_approval_notification('timed_out') task.websocket_emit_status(task.status) task.job_explanation = timeout_message task.save(update_fields=['status', 'job_explanation', 'timed_out']) def get_expired_workflow_approvals(self): # timeout of 0 indicates that it never expires qs = WorkflowApproval.objects.filter(status='pending').exclude( timeout=0).filter(expires__lt=tz_now()) return qs @timeit def _schedule(self): self.get_tasks( dict(status__in=["pending", "waiting", "running"], dependencies_processed=True)) self.after_lock_init() self.reap_jobs_from_orphaned_instances() if len(self.all_tasks) > 0: self.process_tasks() for workflow_approval in self.get_expired_workflow_approvals(): self.timeout_approval_node(workflow_approval)