工作流管理平台Airflow

Airflow

1. 引言

Airflow是Airbnb开源的一个用Python写就的工作流管理平台（workflow management platform）。在前一篇文章中，介绍了如何用Crontab管理数据流，但是缺点也是显而易见。针对于Crontab的缺点，灵活可扩展的Airflow具有以下特点：

• 工作流依赖关系的可视化；
• 日志追踪；
• （Python脚本）易于扩展

对比Java系的Oozie，Airflow奉行“Configuration as code”哲学，对于描述工作流、判断触发条件等全部采用Python，使得你编写工作流就像在写脚本一样；能debug工作流（test backfill命令），更好地判别是否有错误；能更快捷地在线上做功能扩展。Airflow充分利用Python的灵巧轻便，相比之下Oozie则显得笨重厚拙太多（其实我没在黑Java～～）。《What makes Airflow great?》介绍了更多关于Airflow的优良特性；其他有关于安装、介绍的文档在这里、还有这里。

下表给出Airflow（基于1.7版本）与Oozie（基于4.0版本）对比情况：

功能	Airflow	Oozie
工作流描述	Python	xml
数据触发	Sensor	datasets, input-events
工作流节点	operator	action
完整工作流	DAG	workflow
定期调度	DAG schedule_interval	coordinator frequency
任务依赖	>>, <<	<ok to>
内置函数、变量	template macros	EL function, EL constants

之前我曾提及Oozie没有能力表达复杂的DAG，是因为Oozie只能指定下流依赖（downstream）而不能指定上流依赖（upstream）。与之相比，Airflow就能表示复杂的DAG。Airflow没有像Oozie一样区分workflow与coordinator，而是把触发条件、工作流节点都看作一个operator，operator组成一个DAG。

2. 实战

下面将给出如何用Airflow完成data pipeline任务。

首先简要地介绍下背景：定时（每周）检查Hive表的partition的任务是否有生成，若有则触发Hive任务写Elasticsearch；然后等Hive任务完后，执行Python脚本查询Elasticsearch发送报表。但是，Airflow对Python3支持有问题（依赖包为Python2编写）；因此不得不自己写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连接Hive server2的采用的是impyla模块，HivePartitionSensor用于判断Hive表的partition是否存在。写自定义的operator，有点像写Hive、Pig的UDF；写好的operator需要放在目录~/airflow/dags，以便于DAG调用。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