PySpark和廣播加入例如

Question

我使用的Spark 1.3

# Read from text file, parse it and then do some basic filtering to get data1 
data1.registerTempTable('data1') 

# Read from text file, parse it and then do some basic filtering to get data1 
data2.registerTempTable('data2') 

# Perform join 
data_joined = data1.join(data2, data1.id == data2.id); 

我的數據是相當扭曲和數據2（幾KB）< < DATA1（GB的10秒），表現相當不錯。我正在閱讀有關廣播連接，但不知道如何使用Python API執行相同操作。

Answer 1

Spark 1.3不支持使用DataFrame的廣播連接。在星火> = 1.5.0，您可以使用broadcast功能應用廣播聯接：

from pyspark.sql.functions import broadcast 

data1.join(broadcast(data2), data1.id == data2.id) 

對於舊版本的唯一的選擇是要轉換爲RDD並應用相同的邏輯在其他語言。大致是這樣的：

from pyspark.sql import Row 
from pyspark.sql.types import StructType 

# Create a dictionary where keys are join keys 
# and values are lists of rows 
data2_bd = sc.broadcast(
    data2.map(lambda r: (r.id, r)).groupByKey().collectAsMap()) 


# Define a new row with fields from both DFs 
output_row = Row(*data1.columns + data2.columns) 

# And an output schema 
output_schema = StructType(data1.schema.fields + data2.schema.fields) 

# Given row x, extract a list of corresponding rows from broadcast 
# and output a list of merged rows 
def gen_rows(x): 
    return [output_row(*x + y) for y in data2_bd.value.get(x.id, [])] 

# flatMap and create a new data frame 
joined = data1.rdd.flatMap(lambda row: gen_rows(row)).toDF(output_schema) 

Answer 2

此代碼工作在火花2.0.2彬hadoop2.7版本

from pyspark.sql import SparkSession 

from pyspark.sql.functions import broadcast 

spark = SparkSession.builder.appName("Python Spark SQL basic example").config("spark.some.config.option", "some-value").getOrCreate() 

df2 = spark.read.csv("D:\\trans_mar.txt",sep="^"); 

df1=spark.read.csv("D:\\trans_feb.txt",sep="^"); 

print(df1.join(broadcast(df2),df2._c77==df1._c77).take(10))