2.4之前的有效udf（注意它必须返回一些东西

from pyspark.sql.functions import udf

@udf("boolean")
def contains_all(x, y):
    if x is not None and y is not None:
        return set(y).issubset(set(x))

在2.4或更高版本中，不需要udf：

from pyspark.sql.functions import array_intersect, size

def contains_all(x, y):
    return size(array_intersect(x, y)) == size(y)

用途：

from pyspark.sql.functions import col, sum as sum_, when

df1 = spark.createDataFrame(
   [(1, [1, 2, 3, 5]), (2, [1, 2, 3, 6]), (3, [1, 2, 9, 8]), (4, [1, 2, 5, 6])],
   ("id", "transactions")
)

df2 = spark.createDataFrame(
    [([1], 1.0), ([2], 1.0), ([2, 1], 2.0), ([6, 1], 2.0)],
    ("items", "cost")
)

(df1
    .crossJoin(df2).groupBy("id", "transactions")
    .agg(sum_(when(
        contains_all("transactions", "items"), col("cost")
    )).alias("score"))
    .show())

结果：

+---+------------+-----+                                                        
| id|transactions|score|
+---+------------+-----+
|  1|[1, 2, 3, 5]|  4.0|
|  4|[1, 2, 5, 6]|  6.0|
|  2|[1, 2, 3, 6]|  6.0|
|  3|[1, 2, 9, 8]|  4.0|
+---+------------+-----+

如果df2很小，最好将其用作局部变量：

items = sc.broadcast([
    (set(items), cost) for items, cost in df2.select("items", "cost").collect()
])

def score(y):
    @udf("double")
    def _(x):
        if x is not None:
            transactions = set(x)
            return sum(
                cost for items, cost in y.value 
                if items.issubset(transactions))
    return _

df1.withColumn("score", score(items)("transactions")).show()

+---+------------+-----+
| id|transactions|score|
+---+------------+-----+
|  1|[1, 2, 3, 5]|  4.0|
|  2|[1, 2, 3, 6]|  6.0|
|  3|[1, 2, 9, 8]|  4.0|
|  4|[1, 2, 5, 6]|  6.0|
+---+------------+-----+

最后才有可能爆款加入

from pyspark.sql.functions import explode

costs = (df1
    # Explode transactiosn
    .select("id", explode("transactions").alias("item"))
    .join(
        df2 
            # Add id so we can later use it to identify source
            .withColumn("_id", monotonically_increasing_id().alias("_id"))
             # Explode items
            .select(
                "_id", explode("items").alias("item"), 
                # We'll need size of the original items later
                size("items").alias("size"), "cost"), 
         ["item"])
     # Count matches in groups id, items
     .groupBy("_id", "id", "size", "cost")
     .count()
     # Compute cost
     .groupBy("id")
     .agg(sum_(when(col("size") == col("count"), col("cost"))).alias("score")))

costs.show()

+---+-----+                                                                      
| id|score|
+---+-----+
|  1|  4.0|
|  3|  4.0|
|  2|  6.0|
|  4|  6.0|
+---+-----+

然后将结果与原始的df1连接起来，

df1.join(costs, ["id"])

但这不是一个简单的解决方案，需要多次 Shuffle 。它可能仍然优于笛卡尔积（crossJoin），但它将取决于实际数据。

**Spark 3.0+**多了一个使用forall的选项

F.expr("forall(look_for, x -> array_contains(look_in, x))")

Spark 3.1+ -F.forall('look_for', lambda x: F.array_contains('look_in', x))的替代语法
将其与选项（Spark 2.4中的array_intersect）进行比较

F.size(F.array_intersect('look_for', 'look_in')) == F.size('look_for')

• 它们在处理重复和空值方面有所不同。*

from pyspark.sql import functions as F
df = spark.createDataFrame(
    [(['a', 'b', 'c'], ['a']),
     (['a', 'b', 'c'], ['d']),
     (['a', 'b', 'c'], ['a', 'b']),
     (['a', 'b', 'c'], ['c', 'd']),
     (['a', 'b', 'c'], ['a', 'b', 'c']),
     (['a', 'b', 'c'], ['a', None]),
     (['a', 'b',None], ['a', None]),
     (['a', 'b',None], ['a']),
     (['a', 'b',None], [None]),
     (['a', 'b', 'c'], None),
     (None, ['a']),
     (None, None),
     (['a', 'b', 'c'], ['a', 'a']),
     (['a', 'a', 'a'], ['a']),
     (['a', 'a', 'a'], ['a', 'a', 'a']),
     (['a', 'a', 'a'], ['a', 'a',None]),
     (['a', 'a',None], ['a', 'a', 'a']),
     (['a', 'a',None], ['a', 'a',None])],
    ['look_in', 'look_for'])
df = df.withColumn('spark_3_0', F.expr("forall(look_for, x -> array_contains(look_in, x))"))
df = df.withColumn('spark_2_4', F.size(F.array_intersect('look_for', 'look_in')) == F.size('look_for'))

在某些情况下，从数组内部删除空值可能很有用，使用Spark 3.4+中的array_compact最简单。