Question

我有2个数据框，比如df1和df2。

df1数据来自数据库，df2是我从客户端收到的新数据。我需要处理新数据，并根据它是新记录还是要更新的现有记录来执行UPSERTs。

示例数据输出：

df1= sqlContext.createDataFrame([("xxx1","81A01","TERR NAME 01","NJ"),("xxx2","81A01","TERR NAME 01","NJ"),("xxx3","81A01","TERR NAME 01","NJ"),("xxx4","81A01","TERR NAME 01","CA"),("xx5","81A01","TERR NAME 01","ME")], ["zip_code","territory_code","territory_name","state"])
df2= sqlContext.createDataFrame([("xxx1","81A01","TERR NAME 55","NY"),("xxx2","81A01","TERR NAME 55","NY"),("x103","81A01","TERR NAME 01","NJ")], ["zip_code","territory_code","territory_name","state"])

df1.show()
+--------+--------------+--------------+-----+
|zip_code|territory_code|territory_name|state|
+--------+--------------+--------------+-----+
|    xxx1|         81A01|  TERR NAME 01|   NJ|
|    xxx2|         81A01|  TERR NAME 01|   NJ|
|    xxx3|         81A01|  TERR NAME 01|   NJ|
|    xxx4|         81A01|  TERR NAME 01|   CA|
|    xxx5|         81A01|  TERR NAME 01|   ME|
+---------------------------------------------

# Print out information about this data
df2.show()
+--------+--------------+--------------+-----+
|zip_code|territory_code|territory_name|state|
+--------+--------------+--------------+-----+
|    xxx1|         81A01|  TERR NAME 55|   NY|
|    xxx2|         81A01|  TERR NAME 55|   NY|
|    x103|         81A01|  TERR NAME 01|   NJ|
+---------------------------------------------

预期结果：我需要将 df2 数据框与 df1 进行比较。根据上述比较创建2个新数据集，即要更新的记录和要附加/插入数据库的记录。

如果 zip_code ＆amp; territory_code 是相同的，那么它是UPDATE，否则它是对数据库的INSERT。

例如： INSERT的新数据框输出：

 +--------+--------------+--------------+-----+
 |zip_code|territory_code|territory_name|state|
 +--------+--------------+--------------+-----+
 |    x103|         81A01|  TERR NAME 01|   NJ|
 +---------------------------------------------

UPDATE的新数据框：

+--------+--------------+--------------+-----+
|zip_code|territory_code|territory_name|state|
+--------+--------------+--------------+-----+
|    xxx1|         81A01|  TERR NAME 55|   NY|
|    xxx2|         81A01|  TERR NAME 55|   NY|
+---------------------------------------------

有人可以帮帮我吗？我正在使用AWS Glue。

更新：解决方案（使用连接和减法）

df3 = df1.join(df2, (df1.zip_code == df2.zip_code_new) & (df1.territory_code == df2.territory_code_new))
df5=df3.drop("zip_code", "territory_code", "territory_name", "state")
df5.show()

+------------+------------------+------------------+---------+
|zip_code_new|territory_code_new|territory_name_new|state_new|
+------------+------------------+------------------+---------+
|        x103|             81A01|      TERR NAME 01|       NJ|
+------------+------------------+------------------+---------+

df4=df2.subtract(df5)
df4.show()

+------------+------------------+------------------+---------+
|zip_code_new|territory_code_new|territory_name_new|state_new|
+------------+------------------+------------------+---------+
|    xxx1    |         81A01    |  TERR NAME 55    |   NY    |
|    xxx2    |         81A01    |  TERR NAME 55    |   NY    |
+------------------------------------------------------------+

对于RDS数据库更新，我使用pymysql / Mysqldb：

db = MySQLdb.connect("xxxx.rds.amazonaws.com", "username", "password", "databasename")
cursor = db.cursor()

#cursor.execute("REPLACE INTO table SELECT * FROM table_stg")
insertQry = "INSERT INTO table VALUES('xxx1','81A01','TERR NAME 55','NY') ON DUPLICATE KEY UPDATE territory_name='TERR NAME 55', state='NY'"
n=cursor.execute(insertQry)
db.commit()
cursor.fetchall()
db.close()

由于

Answer 1

这是一个解决方案草图：

在您的唯一键（zip_code和地区）上投射两个帧
使用spark dataframe api计算两个数据帧之间的交集和差异。请看这个链接：How to obtain the difference between two DataFrames?
对键的交集进行更新
为差异做一个插入（在新数据框内，而不是在现有数据中）

在scala中，这看起来像这样 - 在python中它应该非常相似：

import org.apache.spark.sql.SparkSession

case class ZipTerr(zip_code: String, territory_code: String, 
    territory_name: String, state:String)

case class Key(zip_code: String, territory_code: String)

val spark: SparkSession

val newData = spark.createDataFrame(List(
  ZipTerr("xxx1", "81A01", "TERR NAME 01", "NJ"),
  ZipTerr("xxx2", "81A01", "TERR NAME 01", "NJ"),
  ZipTerr("xxx3", "81A01", "TERR NAME 01", "NJ"),
  ZipTerr("xxx4", "81A01", "TERR NAME 01", "CA"),
  ZipTerr("xx5","81A01","TERR NAME 01","ME")
))

val oldData = spark.createDataFrame(List(
  ZipTerr("xxx1","81A01","TERR NAME 55","NY"),
  ZipTerr("xxx2","81A01","TERR NAME 55","NY"),
  ZipTerr("x103","81A01","TERR NAME 01","NJ")
))

val newKeys = newData.map(z => Key(z.getAs("zip_code"), z.getAs("territory_code")))
val oldKeys = oldData.map(z => Key(z.getAs("zip_code"), z.getAs("territory_code")))

val keysToInsert = newKeys.except(oldKeys)
val keysToUpdate = newKeys.intersect(oldKeys)

这有帮助吗？

注意：变量名称表明您正在使用胶水动态帧。然而，您正在使用sqlContext.createDataFrame函数为它们分配纯火花数据帧。

Answer 2

为了清楚起见，我在这里使用代码片段重现解决方案：

df1= sqlContext.createDataFrame([("xxx1","81A01","TERR NAME 01","NJ"),("xxx2","81A01","TERR NAME 01","NJ"),("xxx3","81A01","TERR NAME 01","NJ"),("xxx4","81A01","TERR NAME 01","CA"),("xx5","81A01","TERR NAME 01","ME")], ["zip_code","territory_code","territory_name","state"])
df2= sqlContext.createDataFrame([("xxx1","81A01","TERR NAME 55","NY"),("xxx2","81A01","TERR NAME 55","NY"),("x103","81A01","TERR NAME 01","NJ")], ["zip_code_new","territory_code_new","territory_name_new","state"])

df1.show()
+--------+--------------+--------------+-----+
|zip_code|territory_code|territory_name|state|
+--------+--------------+--------------+-----+
|    xxx1|         81A01|  TERR NAME 01|   NJ|
|    xxx2|         81A01|  TERR NAME 01|   NJ|
|    xxx3|         81A01|  TERR NAME 01|   NJ|
|    xxx4|         81A01|  TERR NAME 01|   CA|
|    xxx5|         81A01|  TERR NAME 01|   ME|
+---------------------------------------------

# Print out information about this data
df2.show()
+------------+------------------+------------------+---------+
|zip_code_new|territory_code_new|territory_name_new|state_new|
+------------+------------------+------------------+---------+
|    xxx1    |         81A01    |  TERR NAME 55    |   NY    |
|    xxx2    |         81A01    |  TERR NAME 55    |   NY    |
|    x103    |         81A01    |  TERR NAME 01    |   NJ    |
+------------------------------------------------------------+

获取新记录，可以使用＆＃34; 附加＆＃34;将其插入到mysql中。操作

df3 = df1.join(df2, (df1.zip_code == df2.zip_code_new) & (df1.territory_code == df2.territory_code_new))
df5=df3.drop("zip_code", "territory_code", "territory_name", "state")
df5.show()

+------------+------------------+------------------+---------+
|zip_code_new|territory_code_new|territory_name_new|state_new|
+------------+------------------+------------------+---------+
|        x103|             81A01|      TERR NAME 01|       NJ|
+------------+------------------+------------------+---------+

然后获取需要更新到mysql数据库的其余记录。在纯python需要的情况下，我们可以使用arr = df1.collect()，然后是for r in arr:，否则，我们可以使用pandas迭代器处理每个记录。

df4=df2.subtract(df5)
df4.show()

+------------+------------------+------------------+---------+
|zip_code_new|territory_code_new|territory_name_new|state_new|
+------------+------------------+------------------+---------+
|    xxx1    |         81A01    |  TERR NAME 55    |   NY    |
|    xxx2    |         81A01    |  TERR NAME 55    |   NY    |
+------------------------------------------------------------+

希望这可以帮助有需要的人。如果在上述场景中有更好的数据帧迭代方法，请告诉我。感谢

Pyspark - 基于数据框

2 个答案: