python之excel自动报表

一、自动报表前期工作:

  需要安装XlsxWriter模块,可以从github上下载以后解压、setup.py install。

二、程序:

#!/usr/bin/env python
# -*- coding=utf8 -*- import os
import re
import xlsxwriter
import sys
import collections # salt 172.18.144.69 cmd.run "salt-call hdfs_capacity.hdfs_cap decare"
def check_capacity():
value_dicts=collections.OrderedDict()
ns_list=str(os.popen("cat /software/servers/hadoop-2.7.1/etc/hadoop/hdfs-site.xml|grep -A 1 dfs.nameservices|grep -v dfs.nameservices| awk -F '>|<' '{print $3}'").read().strip().split("\n")).strip("['']").split(",")
for ns in ns_list:
result=os.popen("su - hadp -c 'hadoop dfs -count -q hdfs://%s/user/*'"%ns).read()
lists=[]
for line in result.strip().split("\n"):
temp_list = [i for i in re.split('\|+',line.strip().replace(" ","|")) if i !=' ']
temp_value = str(int(temp_list[6])/1024/1024/1024)
temp_list[6] = temp_value
lists.append(temp_list[4:])
value_dicts[ns]=sorted(lists, key=lambda temp:int(temp[2]),reverse=True)
return value_dicts def xml(value_dicts,cluster_name):
workbook = xlsxwriter.Workbook ('/tmp/%s_hdfs.xlsx'%cluster_name)
worksheet = workbook.add_worksheet ()
worksheet.set_column ('E:E', 40)
worksheet.set_column ('B:D', 15)
worksheet.set_row(0,20)
#设置格式
header_format = workbook.add_format ({'bold': True,'bg_color':'#7CCD7C','border':1,'align': 'center','valign': 'vcenter'})
value_format = workbook.add_format ({'border':1})
menu_format = workbook.add_format ({'bold': True,'border':1,'bg_color':'red','align': 'center','valign': 'vcenter'})
worksheet.write ('A1', u'ns',header_format)
worksheet.write ('B1', u'目录个数',header_format)
worksheet.write ('C1', u'文件个数',header_format)
worksheet.write ('D1', u'容量(单位GB)',header_format)
worksheet.write ('E1', u'用户目录',header_format)
ind_num=2
num=0
for index,keys in enumerate(value_dicts.keys()):
ind_num+=int(num)
num=0
for indexs,value in enumerate(value_dicts[keys]):
worksheet.write ('B%s'%str(int(indexs)+ind_num), value[0],value_format)
worksheet.write ('C%s'%str(int(indexs)+ind_num), value[1],value_format)
worksheet.write ('D%s'%str(int(indexs)+ind_num), value[2],value_format)
worksheet.write ('E%s'%str(int(indexs)+ind_num), value[3],value_format)
num=int(indexs)+1
if num-1 == 0:
worksheet.write ('A%s'%str(ind_num), keys,menu_format)
else:
#合并单元格
worksheet.merge_range(ind_num-1,0, ind_num+num-2,0,keys,menu_format)
workbook.close() def hdfs_cap(cluster_name):
try:
xml(check_capacity(),cluster_name)
except Exception,e:
return False
else:
return True if __name__ == "__main__":
cluster_name=sys.argv[1]
xml(check_capacity(),cluster_name)

三、后续:

  考虑到本次程序是单一的,以后可以考虑到放到django中,通过salt-api来调用,并从页面出口。

上一篇:PHP无限极分类


下一篇:GPFS 安装测试