我想分享这个特别的Apache Spark与Python解决方案,因为它的文档相当贫乏。
我想计算K/V对(存储在Pairwise RDD中)的平均值,按KEY分组。以下是示例数据的样子:
>>> rdd1.take(10) # Show a small sample.
[(u'2013-10-09', 7.60117302052786),
(u'2013-10-10', 9.322709163346612),
(u'2013-10-10', 28.264462809917358),
(u'2013-10-07', 9.664429530201343),
(u'2013-10-07', 12.461538461538463),
(u'2013-10-09', 20.76923076923077),
(u'2013-10-08', 11.842105263157894),
(u'2013-10-13', 32.32514177693762),
(u'2013-10-13', 26.249999999999996),
(u'2013-10-13', 10.693069306930692)]
现在以下代码序列是一种不太理想的方式,但它确实可以工作。这就是我在找到更好的解决方案之前所做的。它不是很糟糕,但--如您将在回答部分看到的那样--有一种更简洁、高效的方法。
>>> import operator
>>> countsByKey = sc.broadcast(rdd1.countByKey()) # SAMPLE OUTPUT of countsByKey.value: {u'2013-09-09': 215, u'2013-09-08': 69, ... snip ...}
>>> rdd1 = rdd1.reduceByKey(operator.add) # Calculate the numerators (i.e. the SUMs).
>>> rdd1 = rdd1.map(lambda x: (x[0], x[1]/countsByKey.value[x[0]])) # Divide each SUM by it's denominator (i.e. COUNT)
>>> print(rdd1.collect())
[(u'2013-10-09', 11.235365503035176),
(u'2013-10-07', 23.39500642456595),
... snip ...
]