hadoop streaming编程小demo(python版)

大数据团队搞数据质量评测。自动化质检和监控平台是用django，MR也是通过python实现的。(后来发现有orc压缩问题，python不知道怎么解决，正在改成java版本)

这里展示一个python编写MR的例子吧。

抄一句话：Hadoop Streaming是Hadoop提供的一个编程工具，它允许用户使用任何可执行文件或者脚本文件作为Mapper和Reducer。

1、首先，先介绍一下背景，我们的数据是存放在hive里的。hive建表语句如下：

我们将会解析元数据，和HDFS上的数据进行merge，方便处理。这里的partition_key用的是year/month/day。

hive (gulfstream_ods)> desc g_order;
OK
col_name        data_type       comment
order_id                bigint                  订单id
driver_id               bigint                  司机id，司机抢单前该值为0
driver_phone            string                  司机电话
passenger_id            bigint                  乘客id
passenger_phone         string                  乘客电话
car_id                  int                     接驾车辆id
area                    int                     城市id
district                string                  城市区号
type                    int                     订单时效，0 实时  1预约
current_lng             decimal(19,6)           乘客发单时的经度
current_lat             decimal(19,6)           乘客发单时的纬度
starting_name           string                  起点名称
starting_lng            decimal(19,6)           起点经度
starting_lat            decimal(19,6)           起点纬度
dest_name               string                  终点名称
dest_lng                decimal(19,6)           终点经度
dest_lat                decimal(19,6)           终点纬度
driver_start_distance   int                     司机与出发地的路面距离，单位：米
start_dest_distance     int                     出发地与终点的路面距离，单位：米
departure_time          string                  出发时间（预约单的预约时间，实时单为发单时间）
strive_time             string                  抢单成功时间
consult_time            string                  协商时间
arrive_time             string                  司机点击‘我已到达’的时间
setoncar_time           string                  上车时间（暂时不用）
begin_charge_time       string                  司机点机‘开始计费’的时间
finish_time             string                  完成时间
year                    string
month                   string
day                     string

# Partition Information
# col_name              data_type               comment

year                    string
month                   string
day                     string

2、我们解析元数据

这里是解析元数据的过程。之后我们把元数据序列化后存入文件desc.gulfstream_ods.g_order，我们将会将此配置文件连同MR脚本一起上传到hadoop集群。

import subprocess
from subprocess import Popen

def desc_table(db, table):
process = Popen('hive -e "desc %s.%s"' % (db, table),
shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
stdout, stderr = process.communicate()
is_column = True
structure_list = list()
column_list = list()
for line in stdout.split('\n'):
value_list = list()
if not line or len(line.split()) < 2:
break
if is_column:
column_list = line.split()
is_column = False
continue
else:
value_list = line.split()
structure_dict = dict(zip(column_list, value_list))
structure_list.append(structure_dict)

return structure_list

3、下面是hadoop streaming执行脚本。

#!/bin/bash
source /etc/profile
source ~/.bash_profile

#hadoop目录
echo "HADOOP_HOME: "$HADOOP_HOME
HADOOP="$HADOOP_HOME/bin/hadoop"

DB=$1
TABLE=$2
YEAR=$3
MONTH=$4
DAY=$5
echo $DB--$TABLE--$YEAR--$MONTH--$DAY

if [ "$DB" = "gulfstream_ods" ]
then
DB_NAME="gulfstream"
else
DB_NAME=$DB
fi
TABLE_NAME=$TABLE

#输入路径
input_path="/user/xiaoju/data/bi/$DB_NAME/$TABLE_NAME/$YEAR/$MONTH/$DAY/*"
#标记文件后缀名
input_mark="_SUCCESS"
echo $input_path
#输出路径
output_path="/user/bigdata-t/QA/yangfan/$DB_NAME/$TABLE_NAME/$YEAR/$MONTH/$DAY"
output_mark="_SUCCESS"
echo $output_path
#性能约束参数
capacity_mapper=500
capacity_reducer=200
map_num=10
reducer_num=10
queue_name="root.dashujudidiyanjiuyuan-zhinengpingtaibu.datapolicy-develop"
#启动job name
job_name="DW_Monitor_${DB_NAME}_${TABLE_NAME}_${YEAR}${MONTH}${DAY}"
mapper="python mapper.py $DB $TABLE_NAME"
reducer="python reducer.py"

$HADOOP fs -rmr $output_path
$HADOOP jar $HADOOP_HOME/share/hadoop/tools/lib/hadoop-streaming-2.7.2.jar \
-jobconf mapred.job.name="$job_name" \
-jobconf mapred.job.queue.name=$queue_name \
-jobconf mapred.map.tasks=$map_num \
-jobconf mapred.reduce.tasks=$reducer_num \
-jobconf mapred.map.capacity=$capacity_mapper \
-jobconf mapred.reduce.capacity=$capacity_reducer \
-input $input_path \
-output $output_path \
-file ./mapper.py \
-file ./reducer.py \
-file ./utils.py \
-file ./"desc.${DB}.${TABLE_NAME}" \
-mapper "$mapper" \
-reducer "$reducer"
if [ $? -ne 0 ]; then
echo "$DB_NAME $TABLE_NAME $YEAR $MONTH $DAY run faild"
fi
$HADOOP fs -touchz "${output_path}/$output_mark"
rm -rf ./${DB_NAME}.${TABLE_NAME}.${YEAR}-${MONTH}-${DAY}
$HADOOP fs -get $output_path/part-00000 ./${DB_NAME}.${TABLE_NAME}.${YEAR}-${MONTH}-${DAY}

4、这里是Wordcount的进阶版本，第一个功能是分区域统计订单量，第二个功能是在一天中分时段统计订单量。

mapper脚本

# -*- coding:utf-8 -*-
#!/usr/bin/env python
import sys
import json
import pickle
reload(sys)
sys.setdefaultencoding('utf-8')

# 将字段和元数据匹配, 返回迭代器
def read_from_input(file, separator, columns):
for line in file:
if line is None or line == '':
continue
data_list = mapper_input(line, separator)
if not data_list:
continue
item = None
# 最后3列, 年月日作为partitionkey, 无用
if len(data_list) == len(columns) - 3:
item = dict(zip(columns, data_list))
elif len(data_list) == len(columns):
item = dict(zip(columns, data_list))
if not item:
continue
yield item

def index_columns(db, table):
with open('desc.%s.%s' % (db, table), 'r') as fr:
structure_list = deserialize(fr.read())
return [column.get('col_name') for column in structure_list]

# map入口
def main(separator, columns):
items = read_from_input(sys.stdin, separator, columns)
mapper_result = {}
for item in items:
mapper_plugin_1(item, mapper_result)
mapper_plugin_2(item, mapper_result)

def mapper_plugin_1(item, mapper_result):
# key在现实中可以是不同appkey, 是用来分发到不同的reducer上的, 相同的route用来分发到相同的reducer
key = 'route1'
area = item.get('area')
district = item.get('district')
order_id = item.get('order_id')
if not area or not district or not order_id:
return
mapper_output(key, {'area': area, 'district': district, 'order_id': order_id, 'count': 1})

def mapper_plugin_2(item, mapper_result):
key = 'route2'
strive_time = item.get('strive_time')
order_id = item.get('order_id')
if not strive_time or not order_id:
return
try:
day_hour = strive_time.split(':')[0]
mapper_output(key, {'order_id': order_id, 'strive_time': strive_time, 'count': 1, 'day_hour': day_hour})except Exception, ex:
pass

def serialize(data, type='json'):
if type == 'json':
try:
return json.dumps(data)
except Exception, ex:
return ''
elif type == 'pickle':
try:
return pickle.dumps(data)
except Exception, ex:
return ''
else:
return ''

def deserialize(data, type='json'):
if type == 'json':
try:
return json.loads(data)
except Exception, ex:
return []
elif type == 'pickle':
try:
return pickle.loads(data)
except Exception, ex:
return []
else:
return []

def mapper_input(line, separator='\t'):
try:
return line.split(separator)
except Exception, ex:
return None

def mapper_output(key, data, separator='\t'):
key = str(key)
data = serialize(data)
print '%s%s%s' % (key, separator, data)
# print >> sys.stderr, '%s%s%s' % (key, separator, data)

if __name__ == '__main__':
db = sys.argv[1]
table = sys.argv[2]
columns = index_columns(db, table)
main('||', columns)

reducer脚本

#!/usr/bin/env python
# vim: set fileencoding=utf-8
import sys
reload(sys)
sys.setdefaultencoding('utf-8')
import json
import pickle
from itertools import groupby
from operator import itemgetter

def read_from_mapper(file, separator):
for line in file:
yield reducer_input(line)

def main(separator='\t'):
reducer_result = {}
line_list = read_from_mapper(sys.stdin, separator)
for route_key, group in groupby(line_list, itemgetter(0)):
if route_key is None:
continue
reducer_result.setdefault(route_key, {})
if route_key == 'route1':
reducer_plugin_1(route_key, group, reducer_result)
reducer_output(route_key, reducer_result[route_key])
if route_key == 'route2':
reducer_plugin_2(route_key, group, reducer_result)
reducer_output(route_key, reducer_result[route_key])

def reducer_plugin_1(route_key, group, reducer_result):
for _, data in group:
if data is None or len(data) == 0:
continue
if not data.get('area') or not data.get('district') or not data.get('count'):
continue
key = '_'.join([data.get('area'), data.get('district')])
reducer_result[route_key].setdefault(key, 0)
reducer_result[route_key][key] += int(data.get('count'))
# print >> sys.stderr, '%s' % json.dumps(reducer_result[route_key])

def reducer_plugin_2(route_key, group, reducer_result):
for _, data in group:
if data is None or len(data) == 0:
continue
if not data.get('order_id') or not data.get('strive_time') or not data.get('count') or not data.get('day_hour'):
continue
key = data.get('day_hour')
reducer_result[route_key].setdefault(key, {})
reducer_result[route_key][key].setdefault('count', 0)
reducer_result[route_key][key].setdefault('order_list', [])
reducer_result[route_key][key]['count'] += int(data.get('count'))
if len(reducer_result[route_key][key]['order_list']) < 100:
reducer_result[route_key][key]['order_list'].append(data.get('order_id'))
# print >> sys.stderr, '%s' % json.dumps(reducer_result[route_key])

def serialize(data, type='json'):
if type == 'json':
try:
return json.dumps(data)
except Exception, ex:
return ''
elif type == 'pickle':
try:
return pickle.dumps(data)
except Exception, ex:
return ''
else:
return ''

def deserialize(data, type='json'):
if type == 'json':
try:
return json.loads(data)
except Exception, ex:
return []
elif type == 'pickle':
try:
return pickle.loads(data)
except Exception, ex:
return []
else:
return []

def reducer_input(data, separator='\t'):
data_list = data.strip().split(separator, 2)
key = data_list[0]
data = deserialize(data_list[1])
return [key, data]

def reducer_output(key, data, separator='\t'):
key = str(key)
data = serialize(data)
print '%s\t%s' % (key, data)
# print >> sys.stderr, '%s\t%s' % (key, data)

if __name__ == '__main__':
main()

5、上一个版本，遭遇了reduce慢的情况，原因有两个：一是因为route的设置，所有相同的route都将分发到同一个reducer，造成单个reducer处理压力大，性能下降。二是因为集群是搭建在虚拟机上的，性能本身就差。可以对这个问题进行改进。改进版本如下，方案是在mapper阶段先对数据进行初步的统计，缓解reducer的计算压力。

mapper脚本

# -*- coding:utf-8 -*-
#!/usr/bin/env python
import sys
import json
import pickle
reload(sys)
sys.setdefaultencoding('utf-8')

# 将字段和元数据匹配, 返回迭代器
def read_from_input(file, separator, columns):
for line in file:
if line is None or line == '':
continue
data_list = mapper_input(line, separator)
if not data_list:
continue
item = None
# 最后3列, 年月日作为partitionkey, 无用
if len(data_list) == len(columns) - 3:
item = dict(zip(columns, data_list))
elif len(data_list) == len(columns):
item = dict(zip(columns, data_list))
if not item:
continue
yield item

def index_columns(db, table):
with open('desc.%s.%s' % (db, table), 'r') as fr:
structure_list = deserialize(fr.read())
return [column.get('col_name') for column in structure_list]

# map入口
def main(separator, columns):
items = read_from_input(sys.stdin, separator, columns)
mapper_result = {}
for item in items:
mapper_plugin_1(item, mapper_result)
mapper_plugin_2(item, mapper_result)

for route_key, route_value in mapper_result.iteritems():
for key, value in route_value.iteritems():
ret_dict = dict()
ret_dict['route_key'] = route_key
ret_dict['key'] = key
ret_dict.update(value)
mapper_output('route_total', ret_dict)

def mapper_plugin_1(item, mapper_result):
# key在现实中可以是不同appkey, 是用来分发到不同的reducer上的, 相同的route用来分发到相同的reducer
key = 'route1'
area = item.get('area')
district = item.get('district')
order_id = item.get('order_id')
if not area or not district or not order_id:
returntry:
# total统计
mapper_result.setdefault(key, {})
mapper_result[key].setdefault('_'.join([area, district]), {})
mapper_result[key]['_'.join([area, district])].setdefault('count', 0)
mapper_result[key]['_'.join([area, district])].setdefault('order_id', [])
mapper_result[key]['_'.join([area, district])]['count'] += 1
if len(mapper_result[key]['_'.join([area, district])]['order_id']) < 10:
mapper_result[key]['_'.join([area, district])]['order_id'].append(order_id)
except Exception, ex:
pass

def mapper_plugin_2(item, mapper_result):
key = 'route2'
strive_time = item.get('strive_time')
order_id = item.get('order_id')
if not strive_time or not order_id:
return
try:
day_hour = strive_time.split(':')[0]# total统计
mapper_result.setdefault(key, {})
mapper_result[key].setdefault(day_hour, {})
mapper_result[key][day_hour].setdefault('count', 0)
mapper_result[key][day_hour].setdefault('order_id', [])
mapper_result[key][day_hour]['count'] += 1
if len(mapper_result[key][day_hour]['order_id']) < 10:
mapper_result[key][day_hour]['order_id'].append(order_id)
except Exception, ex:
pass

def serialize(data, type='json'):
if type == 'json':
try:
return json.dumps(data)
except Exception, ex:
return ''
elif type == 'pickle':
try:
return pickle.dumps(data)
except Exception, ex:
return ''
else:
return ''

def deserialize(data, type='json'):
if type == 'json':
try:
return json.loads(data)
except Exception, ex:
return []
elif type == 'pickle':
try:
return pickle.loads(data)
except Exception, ex:
return []
else:
return []

def mapper_input(line, separator='\t'):
try:
return line.split(separator)
except Exception, ex:
return None

def mapper_output(key, data, separator='\t'):
key = str(key)
data = serialize(data)
print '%s%s%s' % (key, separator, data)
# print >> sys.stderr, '%s%s%s' % (key, separator, data)

if __name__ == '__main__':
db = sys.argv[1]
table = sys.argv[2]
columns = index_columns(db, table)
main('||', columns)

reducer脚本

#!/usr/bin/env python
# vim: set fileencoding=utf-8
import sys
reload(sys)
sys.setdefaultencoding('utf-8')
import json
import pickle
from itertools import groupby
from operator import itemgetter

def read_from_mapper(file, separator):
for line in file:
yield reducer_input(line)

def main(separator='\t'):
reducer_result = {}
line_list = read_from_mapper(sys.stdin, separator)
for route_key, group in groupby(line_list, itemgetter(0)):
if route_key is None:
continue
reducer_result.setdefault(route_key, {})if route_key == 'route_total':
reducer_total(route_key, group, reducer_result)
reducer_output(route_key, reducer_result[route_key])

def reducer_total(route_key, group, reducer_result):
for _, data in group:
if data is None or len(data) == 0:
continue
if data.get('route_key') == 'route1':
reducer_result[route_key].setdefault(data.get('route_key'), {})
reducer_result[route_key][data.get('key')].setdefault('count', 0)
reducer_result[route_key][data.get('key')].setdefault('order_id', [])
reducer_result[route_key][data.get('key')]['count'] += data.get('count')
for order_id in data.get('order_id'):
if len(reducer_result[route_key][data.get('key')]['order_id']) <= 10:
reducer_result[route_key][data.get('key')]['order_id'].append(order_id)
elif data.get('route_key') == 'route2':
reducer_result[route_key].setdefault(data.get('route_key'), {})
reducer_result[route_key][data.get('key')].setdefault('count', 0)
reducer_result[route_key][data.get('key')].setdefault('order_id', [])
reducer_result[route_key][data.get('key')]['count'] += data.get('count')
for order_id in data.get('order_id'):
if len(reducer_result[route_key][data.get('key')]['order_id']) <= 10:
reducer_result[route_key][data.get('key')]['order_id'].append(order_id)
else:
pass

def serialize(data, type='json'):
if type == 'json':
try:
return json.dumps(data)
except Exception, ex:
return ''
elif type == 'pickle':
try:
return pickle.dumps(data)
except Exception, ex:
return ''
else:
return ''

def deserialize(data, type='json'):
if type == 'json':
try:
return json.loads(data)
except Exception, ex:
return []
elif type == 'pickle':
try:
return pickle.loads(data)
except Exception, ex:
return []
else:
return []

def reducer_input(data, separator='\t'):
data_list = data.strip().split(separator, 2)
key = data_list[0]
data = deserialize(data_list[1])
return [key, data]

def reducer_output(key, data, separator='\t'):
key = str(key)
data = serialize(data)
print '%s\t%s' % (key, data)
# print >> sys.stderr, '%s\t%s' % (key, data)

if __name__ == '__main__':
main()

遇到的问题：

1、The DiskSpace /user/bigdata/qa quota of is exceeded

在reducer结束后，遭遇如上问题，是因为HDFS 路径下的disk容量已经被沾满，释放容量即可;