嗯,有点不对劲。 GroupedData 并不是真正为数据访问而设计的。它只是描述了分组标准并提供了聚合方法。有关详细信息,请参阅我对 Using groupBy in Spark and getting back to a DataFrame 的回复。
这个想法的另一个问题是选择N random samples。如果没有对数据进行心理分组,这是一项很难并行实现的任务,而且当您 call groupBy 在 DataFrame 上时不会发生这种情况:
至少有两种方法可以解决这个问题:
-
转换为 RDD,groupBy 并执行本地采样
import random
n = 3
def sample(iter, n):
rs = random.Random() # We should probably use os.urandom as a seed
return rs.sample(list(iter), n)
df = sqlContext.createDataFrame(
[(x, y, random.random()) for x in (1, 2, 3) for y in "abcdefghi"],
("teamId", "x1", "x2"))
grouped = df.rdd.map(lambda row: (row.teamId, row)).groupByKey()
sampled = sqlContext.createDataFrame(
grouped.flatMap(lambda kv: sample(kv[1], n)))
sampled.show()
## +------+---+-------------------+
## |teamId| x1| x2|
## +------+---+-------------------+
## | 1| g| 0.81921738561455|
## | 1| f| 0.8563875814036598|
## | 1| a| 0.9010425238735935|
## | 2| c| 0.3864428179837973|
## | 2| g|0.06233470405822805|
## | 2| d|0.37620872770129155|
## | 3| f| 0.7518901502732027|
## | 3| e| 0.5142305439671874|
## | 3| d| 0.6250620479303716|
## +------+---+-------------------+
-
使用窗口函数
from pyspark.sql import Window
from pyspark.sql.functions import col, rand, rowNumber
w = Window.partitionBy(col("teamId")).orderBy(col("rnd_"))
sampled = (df
.withColumn("rnd_", rand()) # Add random numbers column
.withColumn("rn_", rowNumber().over(w)) # Add rowNumber over windw
.where(col("rn_") <= n) # Take n observations
.drop("rn_") # drop helper columns
.drop("rnd_"))
sampled.show()
## +------+---+--------------------+
## |teamId| x1| x2|
## +------+---+--------------------+
## | 1| f| 0.8563875814036598|
## | 1| g| 0.81921738561455|
## | 1| i| 0.8173912535268248|
## | 2| h| 0.10862995810038856|
## | 2| c| 0.3864428179837973|
## | 2| a| 0.6695356657072442|
## | 3| b|0.012329360826023095|
## | 3| a| 0.6450777858109182|
## | 3| e| 0.5142305439671874|
## +------+---+--------------------+
但恐怕两者都会相当昂贵。如果各个组的大小是平衡的并且相对较大,我会简单地使用DataFrame.randomSplit。
如果组的数量相对较少,可以尝试其他方法:
from pyspark.sql.functions import count, udf
from pyspark.sql.types import BooleanType
from operator import truediv
counts = (df
.groupBy(col("teamId"))
.agg(count("*").alias("n"))
.rdd.map(lambda r: (r.teamId, r.n))
.collectAsMap())
# This defines fraction of observations from a group which should
# be taken to get n values
counts_bd = sc.broadcast({k: truediv(n, v) for (k, v) in counts.items()})
to_take = udf(lambda k, rnd: rnd <= counts_bd.value.get(k), BooleanType())
sampled = (df
.withColumn("rnd_", rand())
.where(to_take(col("teamId"), col("rnd_")))
.drop("rnd_"))
sampled.show()
## +------+---+--------------------+
## |teamId| x1| x2|
## +------+---+--------------------+
## | 1| d| 0.14815204548854788|
## | 1| f| 0.8563875814036598|
## | 1| g| 0.81921738561455|
## | 2| a| 0.6695356657072442|
## | 2| d| 0.37620872770129155|
## | 2| g| 0.06233470405822805|
## | 3| b|0.012329360826023095|
## | 3| h| 0.9022527556458557|
## +------+---+--------------------+
在 Spark 1.5+ 中,您可以将 udf 替换为对 sampleBy 方法的调用:
df.sampleBy("teamId", counts_bd.value)
它不会为您提供确切的观察次数,但只要每组的观察次数足够大以获取适当的样本,大多数情况下就足够了。您也可以以类似的方式在 RDD 上使用sampleByKey。