apache-spark - 在嵌套字段上加入 PySpark DataFrames

Question

我想在这两个 PySpark DataFrame 之间执行连接:

from pyspark import SparkContext
from pyspark.sql.functions import col

sc = SparkContext()

df1 = sc.parallelize([
    ['owner1', 'obj1', 0.5],
    ['owner1', 'obj1', 0.2],
    ['owner2', 'obj2', 0.1]
]).toDF(('owner', 'object', 'score'))

df2 = sc.parallelize(
          [Row(owner=u'owner1',
           objects=[Row(name=u'obj1', value=Row(fav=True, ratio=0.3))])]).toDF()

必须在对象名称上执行连接，即 df2 的对象和 df1 的对象内部的字段名称。

我能够对嵌套字段执行 SELECT，如

df2.where(df2.owner == 'owner1').select(col("objects.value.ratio")).show()

但我无法运行此连接:

df2.alias('u').join(df1.alias('s'), col('u.objects.name') == col('s.object'))

返回错误

pyspark.sql.utils.AnalysisException: u"cannot resolve '(objects.name = cast(object as double))' due to data type mismatch: differing types in '(objects.name = cast(object as double))' (array and double).;"

任何想法如何解决这个问题？

Answer 1

由于您想匹配和提取特定元素，最简单的方法是 explode行:

matches = df2.withColumn("object", explode(col("objects"))).alias("u").join(
  df1.alias("s"),
  col("s.object") == col("u.object.name")
)

matches.show()
## +-------------------+------+-----------------+------+------+-----+
## |            objects| owner|           object| owner|object|score|
## +-------------------+------+-----------------+------+------+-----+
## |[[obj1,[true,0.3]]]|owner1|[obj1,[true,0.3]]|owner1|  obj1|  0.5|
## |[[obj1,[true,0.3]]]|owner1|[obj1,[true,0.3]]|owner1|  obj1|  0.2|
## +-------------------+------+-----------------+------+------+-----+

另一种但效率很低的方法是使用 array_contains :

matches_contains = df1.alias("s").join(
  df2.alias("u"), expr("array_contains(objects.name, object)"))

它是无效的，因为它将扩展为笛卡尔积:

matches_contains.explain()
## == Physical Plan ==
## Filter array_contains(objects#6.name,object#4)
## +- CartesianProduct
##    :- Scan ExistingRDD[owner#3,object#4,score#5] 
##    +- Scan ExistingRDD[objects#6,owner#7]

如果数组的大小相对较小，则可以生成 array_contains 的优化版本。正如我在这里展示的: Filter by whether column value equals a list in spark