1.pandas df 与 spark df的相互转换

df_s=spark.createDataFrame(df_p)

df_p=df_s.toPandas()

1 2 3

import pandas as pd import numpy as np arr = np.arange(6).reshape(-1,3)

1 2	df_p=pd.DataFrame(arr) df_p

1 2	df_p.columns=['a','b','c'] df_p

1 2	df_s=spark.createDataFrame(df_p) df_s.show()

 

1	df_s.collect()

1	df_s.toPandas()

2. Spark与Pandas中DataFrame对比

http://www.lining0806.com/spark%E4%B8%8Epandas%E4%B8%ADdataframe%E5%AF%B9%E6%AF%94/

 

3.1 利用反射机制推断RDD模式

• sc创建RDD
• 转换成Row元素，列名=值
• spark.createDataFrame生成df
• df.show(), df.printSchema()

1 2 3 4

from pyspark.sql import Row people = spark.sparkContext.textFile('file:///usr/local/spark/examples/src/main/resources/people.txt').map(lambda line:line.split(',')).map(lambda w:Row(name=w[0],age=int(w[1]))) sPeople = spark.createDataFrame(people) sPeople.createOrReplaceTempView('people')

1 2 3

personDF = spark.sql('select name,age from people where age>20') personRDD = personDF.rdd.map(lambda p:"Name:"+p.name+","+"Age:"+str(p.age)) personRDD.foreach(print)

 

1	sPeople.show()

1	sPeople.printSchema()

3.2 使用编程方式定义RDD模式

• 生成“表头”
  • fields = [StructField(field_name, StringType(), True) ,...]
  • schema = StructType(fields)

1 2 3 4

from pyspark.sql.types import * from pyspark.sql import Row schemaString = 'name age' fields = [StructField(field_name,StringType(),True) for field_name in schemaString.split(' ')]<br>schema = StructType(fields)

• 生成“表中的记录”
  • 创建RDD
  • 转换成Row元素，列名=值

1 2 3 4

lines = spark.sparkContext.textFile('file:///usr/local/spark/examples/src/main/resources/people.txt') part = lines.map(lambda w:w.split(",")) peoples = part.map(lambda p:Row(p[0],p[1].strip())) peoples.collect()

• 把“表头”和“表中的记录”拼装在一起
  • = spark.createDataFrame(RDD, schema)

1 2 3

schemaPeople = spark.createDataFrame(people,schema) schemaPeople.show() schemaPeople.printSchema()

 4. DataFrame保存为文件

df.write.json(dir)

1	schemaPeople.write.json('file:///home/hadoop/schema_out')