.. swf_tut: :Authors: Slawek "oozie" Ligus , Brad Morris =============================== Amazon Simple Workflow Tutorial =============================== This tutorial focuses on boto's interface to AWS SimpleWorkflow service. .. _SimpleWorkflow: http://aws.amazon.com/swf/ What is a workflow? ------------------- A workflow is a sequence of multiple activities aimed at accomplishing a well-defined objective. For instance, booking an airline ticket as a workflow may encompass multiple activities, such as selection of itinerary, submission of personal details, payment validation and booking confirmation. Except for the start and completion of a workflow, each step has a well-defined predecessor and successor. With that - on successful completion of an activity the workflow can progress with its execution, - when one of workflow's activities fails it can be retried, - and when it keeps failing repeatedly the workflow may regress to the previous step to gather alternative inputs or it may simply fail at that stage. Why use workflows? ------------------ Modelling an application on a workflow provides a useful abstraction layer for writing highly-reliable programs for distributed systems, as individual responsibilities can be delegated to a set of redundant, independent and non-critical processing units. How does Amazon SWF help you accomplish this? --------------------------------------------- Amazon SimpleWorkflow service defines an interface for workflow orchestration and provides state persistence for workflow executions. Amazon SWF applications involve communication between the following entities: - The Amazon Simple Workflow Service - providing centralized orchestration and workflow state persistence, - Workflow Executors - some entity starting workflow executions, typically through an action taken by a user or from a cronjob. - Deciders - a program codifying the business logic, i.e. a set of instructions and decisions. Deciders take decisions based on initial set of conditions and outcomes from activities. - Activity Workers - their objective is very straightforward: to take inputs, execute the tasks and return a result to the Service. The Workflow Executor contacts SWF Service and requests instantiation of a workflow. A new workflow is created and its state is stored in the service. The next time a decider contacts SWF service to ask for a decision task, it will be informed about a new workflow execution is taking place and it will be asked to advise SWF service on what the next steps should be. The decider then instructs the service to dispatch specific tasks to activity workers. At the next activity worker poll, the task is dispatched, then executed and the results reported back to the SWF, which then passes them onto the deciders. This exchange keeps happening repeatedly until the decider is satisfied and instructs the service to complete the execution. Prerequisites ------------- You need a valid access and secret key. The examples below assume that you have exported them to your environment, as follows: .. code-block:: bash bash$ export AWS_ACCESS_KEY_ID= bash$ export AWS_SECRET_ACCESS_KEY= Before workflows and activities can be used, they have to be registered with SWF service: .. code-block:: python # register.py import boto.swf.layer2 as swf from boto.swf.exceptions import SWFTypeAlreadyExistsError, SWFDomainAlreadyExistsError DOMAIN = 'boto_tutorial' VERSION = '1.0' registerables = [] registerables.append(swf.Domain(name=DOMAIN)) for workflow_type in ('HelloWorkflow', 'SerialWorkflow', 'ParallelWorkflow', 'SubWorkflow'): registerables.append(swf.WorkflowType(domain=DOMAIN, name=workflow_type, version=VERSION, task_list='default')) for activity_type in ('HelloWorld', 'ActivityA', 'ActivityB', 'ActivityC'): registerables.append(swf.ActivityType(domain=DOMAIN, name=activity_type, version=VERSION, task_list='default')) for swf_entity in registerables: try: swf_entity.register() print swf_entity.name, 'registered successfully' except (SWFDomainAlreadyExistsError, SWFTypeAlreadyExistsError): print swf_entity.__class__.__name__, swf_entity.name, 'already exists' Execution of the above should produce no errors. .. code-block:: bash bash$ python -i register.py Domain boto_tutorial already exists WorkflowType HelloWorkflow already exists SerialWorkflow registered successfully ParallelWorkflow registered successfully ActivityType HelloWorld already exists ActivityA registered successfully ActivityB registered successfully ActivityC registered successfully >>> HelloWorld ---------- This example is an implementation of a minimal Hello World workflow. Its execution should unfold as follows: #. A workflow execution is started. #. The SWF service schedules the initial decision task. #. A decider polls for decision tasks and receives one. #. The decider requests scheduling of an activity task. #. The SWF service schedules the greeting activity task. #. An activity worker polls for activity task and receives one. #. The worker completes the greeting activity. #. The SWF service schedules a decision task to inform about work outcome. #. The decider polls and receives a new decision task. #. The decider schedules workflow completion. #. The workflow execution finishes. Workflow logic is encoded in the decider: .. code-block:: python # hello_decider.py import boto.swf.layer2 as swf DOMAIN = 'boto_tutorial' ACTIVITY = 'HelloWorld' VERSION = '1.0' TASKLIST = 'default' class HelloDecider(swf.Decider): domain = DOMAIN task_list = TASKLIST version = VERSION def run(self): history = self.poll() if 'events' in history: # Find workflow events not related to decision scheduling. workflow_events = [e for e in history['events'] if not e['eventType'].startswith('Decision')] last_event = workflow_events[-1] decisions = swf.Layer1Decisions() if last_event['eventType'] == 'WorkflowExecutionStarted': decisions.schedule_activity_task('saying_hi', ACTIVITY, VERSION, task_list=TASKLIST) elif last_event['eventType'] == 'ActivityTaskCompleted': decisions.complete_workflow_execution() self.complete(decisions=decisions) return True The activity worker is responsible for printing the greeting message when the activity task is dispatched to it by the service: .. code-block:: python import boto.swf.layer2 as swf DOMAIN = 'boto_tutorial' VERSION = '1.0' TASKLIST = 'default' class HelloWorker(swf.ActivityWorker): domain = DOMAIN version = VERSION task_list = TASKLIST def run(self): activity_task = self.poll() if 'activityId' in activity_task: print 'Hello, World!' self.complete() return True With actors implemented we can spin up a workflow execution: .. code-block:: bash $ python >>> import boto.swf.layer2 as swf >>> execution = swf.WorkflowType(name='HelloWorkflow', domain='boto_tutorial', version='1.0', task_list='default').start() >>> From separate terminals run an instance of a worker and a decider to carry out a workflow execution (the worker and decider may run from two independent machines). .. code-block:: bash $ python -i hello_decider.py >>> while HelloDecider().run(): pass ... .. code-block:: bash $ python -i hello_worker.py >>> while HelloWorker().run(): pass ... Hello, World! Great. Now, to see what just happened, go back to the original terminal from which the execution was started, and read its history. .. code-block:: bash >>> execution.history() [{'eventId': 1, 'eventTimestamp': 1381095173.2539999, 'eventType': 'WorkflowExecutionStarted', 'workflowExecutionStartedEventAttributes': {'childPolicy': 'TERMINATE', 'executionStartToCloseTimeout': '3600', 'parentInitiatedEventId': 0, 'taskList': {'name': 'default'}, 'taskStartToCloseTimeout': '300', 'workflowType': {'name': 'HelloWorkflow', 'version': '1.0'}}}, {'decisionTaskScheduledEventAttributes': {'startToCloseTimeout': '300', 'taskList': {'name': 'default'}}, 'eventId': 2, 'eventTimestamp': 1381095173.2539999, 'eventType': 'DecisionTaskScheduled'}, {'decisionTaskStartedEventAttributes': {'scheduledEventId': 2}, 'eventId': 3, 'eventTimestamp': 1381095177.5439999, 'eventType': 'DecisionTaskStarted'}, {'decisionTaskCompletedEventAttributes': {'scheduledEventId': 2, 'startedEventId': 3}, 'eventId': 4, 'eventTimestamp': 1381095177.855, 'eventType': 'DecisionTaskCompleted'}, {'activityTaskScheduledEventAttributes': {'activityId': 'saying_hi', 'activityType': {'name': 'HelloWorld', 'version': '1.0'}, 'decisionTaskCompletedEventId': 4, 'heartbeatTimeout': '600', 'scheduleToCloseTimeout': '3900', 'scheduleToStartTimeout': '300', 'startToCloseTimeout': '3600', 'taskList': {'name': 'default'}}, 'eventId': 5, 'eventTimestamp': 1381095177.855, 'eventType': 'ActivityTaskScheduled'}, {'activityTaskStartedEventAttributes': {'scheduledEventId': 5}, 'eventId': 6, 'eventTimestamp': 1381095179.427, 'eventType': 'ActivityTaskStarted'}, {'activityTaskCompletedEventAttributes': {'scheduledEventId': 5, 'startedEventId': 6}, 'eventId': 7, 'eventTimestamp': 1381095179.6989999, 'eventType': 'ActivityTaskCompleted'}, {'decisionTaskScheduledEventAttributes': {'startToCloseTimeout': '300', 'taskList': {'name': 'default'}}, 'eventId': 8, 'eventTimestamp': 1381095179.6989999, 'eventType': 'DecisionTaskScheduled'}, {'decisionTaskStartedEventAttributes': {'scheduledEventId': 8}, 'eventId': 9, 'eventTimestamp': 1381095179.7420001, 'eventType': 'DecisionTaskStarted'}, {'decisionTaskCompletedEventAttributes': {'scheduledEventId': 8, 'startedEventId': 9}, 'eventId': 10, 'eventTimestamp': 1381095180.026, 'eventType': 'DecisionTaskCompleted'}, {'eventId': 11, 'eventTimestamp': 1381095180.026, 'eventType': 'WorkflowExecutionCompleted', 'workflowExecutionCompletedEventAttributes': {'decisionTaskCompletedEventId': 10}}] Serial Activity Execution ------------------------- The following example implements a basic workflow with activities executed one after another. The business logic, i.e. the serial execution of activities, is encoded in the decider: .. code-block:: python # serial_decider.py import time import boto.swf.layer2 as swf class SerialDecider(swf.Decider): domain = 'boto_tutorial' task_list = 'default_tasks' version = '1.0' def run(self): history = self.poll() if 'events' in history: # Get a list of non-decision events to see what event came in last. workflow_events = [e for e in history['events'] if not e['eventType'].startswith('Decision')] decisions = swf.Layer1Decisions() # Record latest non-decision event. last_event = workflow_events[-1] last_event_type = last_event['eventType'] if last_event_type == 'WorkflowExecutionStarted': # Schedule the first activity. decisions.schedule_activity_task('%s-%i' % ('ActivityA', time.time()), 'ActivityA', self.version, task_list='a_tasks') elif last_event_type == 'ActivityTaskCompleted': # Take decision based on the name of activity that has just completed. # 1) Get activity's event id. last_event_attrs = last_event['activityTaskCompletedEventAttributes'] completed_activity_id = last_event_attrs['scheduledEventId'] - 1 # 2) Extract its name. activity_data = history['events'][completed_activity_id] activity_attrs = activity_data['activityTaskScheduledEventAttributes'] activity_name = activity_attrs['activityType']['name'] # 3) Optionally, get the result from the activity. result = last_event['activityTaskCompletedEventAttributes'].get('result') # Take the decision. if activity_name == 'ActivityA': decisions.schedule_activity_task('%s-%i' % ('ActivityB', time.time()), 'ActivityB', self.version, task_list='b_tasks', input=result) if activity_name == 'ActivityB': decisions.schedule_activity_task('%s-%i' % ('ActivityC', time.time()), 'ActivityC', self.version, task_list='c_tasks', input=result) elif activity_name == 'ActivityC': # Final activity completed. We're done. decisions.complete_workflow_execution() self.complete(decisions=decisions) return True The workers only need to know which task lists to poll. .. code-block:: python # serial_worker.py import time import boto.swf.layer2 as swf class MyBaseWorker(swf.ActivityWorker): domain = 'boto_tutorial' version = '1.0' task_list = None def run(self): activity_task = self.poll() if 'activityId' in activity_task: # Get input. # Get the method for the requested activity. try: print 'working on activity from tasklist %s at %i' % (self.task_list, time.time()) self.activity(activity_task.get('input')) except Exception as error: self.fail(reason=str(error)) raise error return True def activity(self, activity_input): raise NotImplementedError class WorkerA(MyBaseWorker): task_list = 'a_tasks' def activity(self, activity_input): self.complete(result="Now don't be givin him sambuca!") class WorkerB(MyBaseWorker): task_list = 'b_tasks' def activity(self, activity_input): self.complete() class WorkerC(MyBaseWorker): task_list = 'c_tasks' def activity(self, activity_input): self.complete() Spin up a workflow execution and run the decider: .. code-block:: bash $ python >>> import boto.swf.layer2 as swf >>> execution = swf.WorkflowType(name='SerialWorkflow', domain='boto_tutorial', version='1.0', task_list='default_tasks').start() >>> .. code-block:: bash $ python -i serial_decider.py >>> while SerialDecider().run(): pass ... Run the workers. The activities will be executed in order: .. code-block:: bash $ python -i serial_worker.py >>> while WorkerA().run(): pass ... working on activity from tasklist a_tasks at 1382046291 .. code-block:: bash $ python -i serial_worker.py >>> while WorkerB().run(): pass ... working on activity from tasklist b_tasks at 1382046541 .. code-block:: bash $ python -i serial_worker.py >>> while WorkerC().run(): pass ... working on activity from tasklist c_tasks at 1382046560 Looks good. Now, do the following to inspect the state and history of the execution: .. code-block:: python >>> execution.describe() {'executionConfiguration': {'childPolicy': 'TERMINATE', 'executionStartToCloseTimeout': '3600', 'taskList': {'name': 'default_tasks'}, 'taskStartToCloseTimeout': '300'}, 'executionInfo': {'cancelRequested': False, 'closeStatus': 'COMPLETED', 'closeTimestamp': 1382046560.901, 'execution': {'runId': '12fQ1zSaLmI5+lLXB8ux+8U+hLOnnXNZCY9Zy+ZvXgzhE=', 'workflowId': 'SerialWorkflow-1.0-1382046514'}, 'executionStatus': 'CLOSED', 'startTimestamp': 1382046514.994, 'workflowType': {'name': 'SerialWorkflow', 'version': '1.0'}}, 'latestActivityTaskTimestamp': 1382046560.632, 'openCounts': {'openActivityTasks': 0, 'openChildWorkflowExecutions': 0, 'openDecisionTasks': 0, 'openTimers': 0}} >>> execution.history() ... Parallel Activity Execution --------------------------- When activities are independent from one another, their execution may be scheduled in parallel. The decider schedules all activities at once and marks progress until all activities are completed, at which point the workflow is completed. .. code-block:: python # parallel_decider.py import boto.swf.layer2 as swf import time SCHED_COUNT = 5 class ParallelDecider(swf.Decider): domain = 'boto_tutorial' task_list = 'default' def run(self): decision_task = self.poll() if 'events' in decision_task: decisions = swf.Layer1Decisions() # Decision* events are irrelevant here and can be ignored. workflow_events = [e for e in decision_task['events'] if not e['eventType'].startswith('Decision')] # Record latest non-decision event. last_event = workflow_events[-1] last_event_type = last_event['eventType'] if last_event_type == 'WorkflowExecutionStarted': # At start, kickoff SCHED_COUNT activities in parallel. for i in range(SCHED_COUNT): decisions.schedule_activity_task('activity%i' % i, 'ActivityA', '1.0', task_list=self.task_list) elif last_event_type == 'ActivityTaskCompleted': # Monitor progress. When all activities complete, complete workflow. completed_count = sum([1 for a in decision_task['events'] if a['eventType'] == 'ActivityTaskCompleted']) print '%i/%i' % (completed_count, SCHED_COUNT) if completed_count == SCHED_COUNT: decisions.complete_workflow_execution() self.complete(decisions=decisions) return True Again, the only bit of information a worker needs is which task list to poll. .. code-block:: python # parallel_worker.py import time import boto.swf.layer2 as swf class ParallelWorker(swf.ActivityWorker): domain = 'boto_tutorial' task_list = 'default' def run(self): """Report current time.""" activity_task = self.poll() if 'activityId' in activity_task: print 'working on', activity_task['activityId'] self.complete(result=str(time.time())) return True Spin up a workflow execution and run the decider: .. code-block:: bash $ python -i parallel_decider.py >>> execution = swf.WorkflowType(name='ParallelWorkflow', domain='boto_tutorial', version='1.0', task_list='default').start() >>> while ParallelDecider().run(): pass ... 1/5 2/5 4/5 5/5 Run two or more workers to see how the service partitions work execution in parallel. .. code-block:: bash $ python -i parallel_worker.py >>> while ParallelWorker().run(): pass ... working on activity1 working on activity3 working on activity4 .. code-block:: bash $ python -i parallel_worker.py >>> while ParallelWorker().run(): pass ... working on activity2 working on activity0 As seen above, the work was partitioned between the two running workers. Sub-Workflows ------------- Sometimes it's desired or necessary to break the process up into multiple workflows. Since the decider is stateless, it's up to you to determine which workflow is being used and which action you would like to take. .. code-block:: python import boto.swf.layer2 as swf class SubWorkflowDecider(swf.Decider): domain = 'boto_tutorial' task_list = 'default' version = '1.0' def run(self): history = self.poll() events = [] if 'events' in history: events = history['events'] # Collect the entire history if there are enough events to become paginated while 'nextPageToken' in history: history = self.poll(next_page_token=history['nextPageToken']) if 'events' in history: events = events + history['events'] workflow_type = history['workflowType']['name'] # Get all of the relevent events that have happened since the last decision task was started workflow_events = [e for e in events if e['eventId'] > history['previousStartedEventId'] and not e['eventType'].startswith('Decision')] decisions = swf.Layer1Decisions() for event in workflow_events: last_event_type = event['eventType'] if last_event_type == 'WorkflowExecutionStarted': if workflow_type == 'SerialWorkflow': decisions.start_child_workflow_execution('SubWorkflow', self.version, "subworkflow_1", task_list=self.task_list, input="sub_1") elif workflow_type == 'SubWorkflow': for i in range(2): decisions.schedule_activity_task("activity_%d" % i, 'ActivityA', self.version, task_list='a_tasks') else: decisions.fail_workflow_execution(reason="Unknown workflow %s" % workflow_type) break elif last_event_type == 'ChildWorkflowExecutionCompleted': decisions.schedule_activity_task("activity_2", 'ActivityB', self.version, task_list='b_tasks') elif last_event_type == 'ActivityTaskCompleted': attrs = event['activityTaskCompletedEventAttributes'] activity = events[attrs['scheduledEventId'] - 1] activity_name = activity['activityTaskScheduledEventAttributes']['activityType']['name'] if activity_name == 'ActivityA': completed_count = sum([1 for a in events if a['eventType'] == 'ActivityTaskCompleted']) if completed_count == 2: # Complete the child workflow decisions.complete_workflow_execution() elif activity_name == 'ActivityB': # Complete the parent workflow decisions.complete_workflow_execution() self.complete(decisions=decisions) return True Misc ---- Some of these things are not obvious by reading the API documents, so hopefully they help you avoid some time-consuming pitfalls. Pagination ========== When the decider polls for new tasks, the maximum number of events it will return at a time is 100 (configurable to a smaller number, but not larger). When running a workflow, this number gets quickly exceeded. If it does, the decision task will contain a key ``nextPageToken`` which can be submit to the ``poll()`` call to get the next page of events. .. code-block:: python decision_task = self.poll() events = [] if 'events' in decision_task: events = decision_task['events'] while 'nextPageToken' in decision_task: decision_task = self.poll(next_page_token=decision_task['nextPageToken']) if 'events' in decision_task: events += decision_task['events'] Depending on your workflow logic, you might not need to aggregate all of the events. Decision Tasks ============== When first running deciders and activities, it may seem that the decider gets called for every event that an activity triggers; however, this is not the case. More than one event can happen between decision tasks. The decision task will contain a key ``previousStartedEventId`` that lets you know the ``eventId`` of the last DecisionTaskStarted event that was processed. Your script will need to handle all of the events that have happened since then, not just the last activity. .. code-block:: python workflow_events = [e for e in events if e['eventId'] > decision_task['previousStartedEventId']] You may also wish to still filter out tasks that start with 'Decision' or filter it in some other way that fulfills your needs. You will now have to iterate over the workflow_events list and respond to each event, as it may contain multiple events. Filtering Events ================ When running many tasks in parallel, a common task is searching through the history to see how many events of a particular activity type started, completed, and/or failed. Some basic list comprehension makes this trivial. .. code-block:: python def filter_completed_events(self, events, type): completed = [e for e in events if e['eventType'] == 'ActivityTaskCompleted'] orig = [events[e['activityTaskCompletedEventAttributes']['scheduledEventId']-1] for e in completed] return [e for e in orig if e['activityTaskScheduledEventAttributes']['activityType']['name'] == type] .. _Amazon SWF API Reference: http://docs.aws.amazon.com/amazonswf/latest/apireference/Welcome.html .. _StackOverflow questions: http://stackoverflow.com/questions/tagged/amazon-swf .. _Miscellaneous Blog Articles: http://log.ooz.ie/search/label/SimpleWorkflow