python - PySpark - RDD 中对象的时间重叠

Question

我的目标是根据时间重叠对对象进行分组。

我的rdd中的每个对象都包含start_time和end_time。

我的处理效率可能很低，但我计划做的是根据每个对象是否与任何其他对象有任何时间重叠来为每个对象分配一个重叠 ID。我有时间重叠的逻辑。然后，我希望按该 overlap_id 进行分组。

首先，

mapped_rdd = rdd.map(assign_overlap_id)
final_rdd = mapped_rdd.reduceByKey(combine_objects)

现在我的问题来了。我该如何编写 allocate_overlap_id 函数？

def assign_overlap_id(x):
  ...
  ...
  return (overlap_id, x)

Answer 1

使用 Spark SQL 和数据框架的简单解决方案:

斯卡拉:

import org.apache.spark.sql.functions.udf

case class Interval(start_time: Long, end_time: Long)

val rdd = sc.parallelize(
    Interval(0, 3) :: Interval(1, 4) ::
    Interval(2, 5) :: Interval(3, 4) ::
    Interval(5, 8) :: Interval(7, 10) :: Nil
)

val df = sqlContext.createDataFrame(rdd)

// Simple check if a given intervals overlap
def overlaps(start_first: Long, end_first: Long,
        start_second: Long, end_second: Long):Boolean = {
    (start_second > start_first & start_second < end_first) |
    (end_second > start_first & end_second < end_first) 
}

// Register udf and data frame aliases
// It look like Spark SQL doesn't support
// aliases in FROM clause [1] so we have to
// register df twice
sqlContext.udf.register("overlaps", overlaps)
df.registerTempTable("df1")
df.registerTempTable("df2")

// Join and filter
sqlContext.sql("""
     SELECT * FROM df1 JOIN df2
     WHERE overlaps(df1.start_time, df1.end_time, df2.start_time, df2.end_time)
""").show

使用 PySpark 也能实现同样的效果

from pyspark.sql.functions import udf
from pyspark.sql.types import BooleanType

rdd = sc.parallelize([
    (0, 3), (1, 4), 
    (2, 5), (3, 4),
    (5, 8), (7, 10)
])

df = sqlContext.createDataFrame(rdd, ('start_time', 'end_time'))

def overlaps(start_first, end_first, start_second, end_second):
    return ((start_first < start_second < end_first) or
        (start_first < end_second < end_first))

sqlContext.registerFunction('overlaps', overlaps, BooleanType())
df.registerTempTable("df1")
df.registerTempTable("df2")

sqlContext.sql("""
     SELECT * FROM df1 JOIN df2
     WHERE overlaps(df1.start_time, df1.end_time, df2.start_time, df2.end_time)
""").show()

按窗口分组的低级转换

更聪明的方法是使用某个指定宽度的窗口生成候选对。这是一个相当简化的解决方案:

斯卡拉:

// Generates list of "buckets" for a given interval
def genRange(interval: Interval) = interval match {
    case Interval(start_time, end_time) => {
      (start_time / 10L * 10L) to (((end_time / 10) + 1) * 10) by 1
    }
}


// For each interval generate pairs (bucket, interval)
val pairs = rdd.flatMap( (i: Interval) => genRange(i).map((r) => (r, i)))

// Join (in the worst case scenario it is still O(n^2)
// But in practice should be better than a naive
// Cartesian product
val candidates = pairs.
    join(pairs).
    map({
        case (k, (Interval(s1, e1), Interval(s2, e2))) => (s1, e1, s2, e2)
   }).distinct


// For each candidate pair check if there is overlap
candidates.filter { case (s1, e1, s2, e2) => overlaps(s1, e1, s2, e2) }

Python:

def genRange(start_time, end_time):
    return xrange(start_time / 10L * 10L, ((end_time / 10) + 1) * 10)

pairs = rdd.flatMap(lambda (s, e): ((r, (s, e)) for r in genRange(s, e)))
candidates = (pairs
    .join(pairs)
    .map(lambda (k, ((s1, e1), (s2, e2))): (s1, e1, s2, e2))
    .distinct())

candidates.filter(lambda (s1, e1, s2, e2): overlaps(s1, e1, s2, e2))

虽然在某些数据集上它足以用于生产就绪的解决方案，但您应该考虑实现一些最先进的算法，例如 NCList .

1. http://docs.datastax.com/en/datastax_enterprise/4.6/datastax_enterprise/spark/sparkSqlSupportedSyntax.html