工作流管理平台Airflow

Airflow

1. 引言

Airflow是Airbnb開源的一個用Python寫就的工作流管理平台（workflow management platform）。在前一篇文章中，介紹了如何用Crontab管理數據流，但是缺點也是顯而易見。針對於Crontab的缺點，靈活可擴展的Airflow具有以下特點：

• 工作流依賴關系的可視化；
• 日志追蹤；
• （Python腳本）易於擴展

對比Java系的Oozie，Airflow奉行“Configuration as code”哲學，對於描述工作流、判斷觸發條件等全部采用Python，使得你編寫工作流就像在寫腳本一樣；能debug工作流（test backfill命令），更好地判別是否有錯誤；能更快捷地在線上做功能擴展。Airflow充分利用Python的靈巧輕便，相比之下Oozie則顯得笨重厚拙太多（其實我沒在黑Java～～）。《What makes Airflow great?》介紹了更多關於Airflow的優良特性；其他有關於安裝、介紹的文檔在這里、還有這里。

下表給出Airflow（基於1.7版本）與Oozie（基於4.0版本）對比情況：

功能	Airflow	Oozie
工作流描述	Python	xml
數據觸發	Sensor	datasets, input-events
工作流節點	operator	action
完整工作流	DAG	workflow
定期調度	DAG schedule_interval	coordinator frequency
任務依賴	>>, <<	<ok to>
內置函數、變量	template macros	EL function, EL constants

之前我曾提及Oozie沒有能力表達復雜的DAG，是因為Oozie只能指定下流依賴（downstream）而不能指定上流依賴（upstream）。與之相比，Airflow就能表示復雜的DAG。Airflow沒有像Oozie一樣區分workflow與coordinator，而是把觸發條件、工作流節點都看作一個operator，operator組成一個DAG。

2. 實戰

下面將給出如何用Airflow完成data pipeline任務。

首先簡要地介紹下背景：定時（每周）檢查Hive表的partition的任務是否有生成，若有則觸發Hive任務寫Elasticsearch；然后等Hive任務完后，執行Python腳本查詢Elasticsearch發送報表。但是，Airflow對Python3支持有問題（依賴包為Python2編寫）；因此不得不自己寫HivePartitionSensor：

# -*- coding: utf-8 -*- # @Time : 2016/11/29 # @Author : rain from airflow.operators import BaseSensorOperator from airflow.utils.decorators import apply_defaults from impala.dbapi import connect import logging class HivePartitionSensor(BaseSensorOperator): """  Waits for a partition to show up in Hive.  :param host, port: the host and port of hiveserver2  :param table: The name of the table to wait for, supports the dot notation (my_database.my_table)  :type table: string  :param partition: The partition clause to wait for. This is passed as  is to the metastore Thrift client,and apparently supports SQL like  notation as in ``ds='2016-12-01'``.  :type partition: string  """ template_fields = ('table', 'partition',) ui_color = '#2b2d42'  @apply_defaults def __init__( self, conn_host, conn_port, table, partition="ds='{{ ds }}'", poke_interval=60 * 3, *args, **kwargs): super(HivePartitionSensor, self).__init__( poke_interval=poke_interval, *args, **kwargs) if not partition: partition = "ds='{{ ds }}'" self.table = table self.partition = partition self.conn_host = conn_host self.conn_port = conn_port self.conn = connect(host=self.conn_host, port=self.conn_port, auth_mechanism='PLAIN') def poke(self, context): logging.info( 'Poking for table {self.table}, ' 'partition {self.partition}'.format(**locals())) cursor = self.conn.cursor() cursor.execute("show partitions {}".format(self.table)) partitions = cursor.fetchall() partitions = [i[0] for i in partitions] if self.partition in partitions: return True else: return False

Python3連接Hive server2的采用的是impyla模塊，HivePartitionSensor用於判斷Hive表的partition是否存在。寫自定義的operator，有點像寫Hive、Pig的UDF；寫好的operator需要放在目錄~/airflow/dags，以便於DAG調用。那么，完整的工作流DAG如下：

# tag cover analysis, based on Airflow v1.7.1.3 from airflow.operators import BashOperator from operatorUD.HivePartitionSensor import HivePartitionSensor from airflow.models import DAG from datetime import datetime, timedelta from impala.dbapi import connect conn = connect(host='192.168.72.18', port=10000, auth_mechanism='PLAIN') def latest_hive_partition(table): cursor = conn.cursor() cursor.execute("show partitions {}".format(table)) partitions = cursor.fetchall() partitions = [i[0] for i in partitions] return partitions[-1].split("=")[1] log_partition_value = """{{ macros.ds_add(ds, -2)}}""" tag_partition_value = latest_hive_partition('tag.dmp') args = { 'owner': 'jyzheng', 'depends_on_past': False, 'start_date': datetime.strptime('2016-12-06', '%Y-%m-%d') } # execute every Tuesday dag = DAG( dag_id='tag_cover', default_args=args, schedule_interval='@weekly', dagrun_timeout=timedelta(minutes=10)) ad_sensor = HivePartitionSensor( task_id='ad_sensor', conn_host='192.168.72.18', conn_port=10000, table='ad.ad_log', partition="day_time={}".format(log_partition_value), dag=dag ) ad_hive_task = BashOperator( task_id='ad_hive_task', bash_command='hive -f /path/to/cron/cover/ad_tag.hql --hivevar LOG_PARTITION={} ' '--hivevar TAG_PARTITION={}'.format(log_partition_value, tag_partition_value), dag=dag ) ad2_hive_task = BashOperator( task_id='ad2_hive_task', bash_command='hive -f /path/to/cron/cover/ad2_tag.hql --hivevar LOG_PARTITION={} ' '--hivevar TAG_PARTITION={}'.format(log_partition_value, tag_partition_value), dag=dag ) report_task = BashOperator( task_id='report_task', bash_command='sleep 5m; python3 /path/to/cron/report/tag_cover.py {}'.format(log_partition_value), dag=dag ) ad_sensor >> ad_hive_task >> report_task ad_sensor >> ad2_hive_task >> report_task