轻量级OLAP（一）：Cube计算

有一个数据多维分析的任务：

• 日志的周UV；
• APP的收集量及标注量，TOP 20 APP（周UV)，TOP 20 APP标注分类（周UV)；
• 手机机型的收集量及标注量，TOP 20 机型（周UV)，TOP 20 手机厂商（周UV)；

初始的解决方案：Spark读取数据日志，然后根据分析需求逐一进行map、distinct、reduceByKey得到分析结果。但是，这种方案存在着非常大的缺点——重复扫描数据源多次。

1. Pig

Pig提供cube关键字做OLAP，将dimension分为了两类：

• normal，对应于cube operation，\(n\)个该维度的组合数为\(2^n\)；
• hierarchical ordering，对应于rollup operation， \(n\)个该维度的组合数为\(n+1\)。

官方doc例子如下：

salesinp = LOAD '/pig/data/salesdata' USING PigStorage(',') AS
    (product:chararray, year:int, region:chararray, state:chararray, city:chararray, sales:long);
cubedinp = CUBE salesinp BY CUBE(product,year);
result = FOREACH cubedinp GENERATE FLATTEN(group), SUM(cube.sales) AS totalsales;

salesinp = LOAD '/pig/data/salesdata' USING PigStorage(',') AS
    (product:chararray, year:int, region:chararray, state:chararray, city:chararray, sales:long);
rolledup = CUBE salesinp BY ROLLUP(region,state,city);
result = FOREACH rolledup GENERATE FLATTEN(group), SUM(cube.sales) AS totalsales

在例子中，cube的操作相当于按维度组合对每一record进行展开并group by Dimensions，与下一句foreach语句构成了Dimensions + Measure的数据输出格式。

2. Spark

朴素多维分析

从上面介绍的pig OLAP方案中，我们得到灵感——面对开篇的多维分析需求，也可以每一条记录按Dimensions + Measure的规则进行展开：

/**
 * @param e (uid, LogFact)
 * @return Array[((dimension order No, dimension), measure)]
 */
def flatAppDvc(e: (String, CaseClasses.LogFact)): Array[((String, String), String)] = {
  val source = (("00", e._2.source), e._1)
  val appName = (("11", e._2.appName), e._1)
  val appTag = (("12", e._2.appTag), e._1)
  val appAll = (("13", "a"), e._1)
  val appCollect = (("14", "a"), e._2.appName)
  val appLabel = e._2.appTag match {
    case "EMPTY" => (("15", "a"), "useless")
    case _ => (("15", "a"), e._2.appName)
  }
  val dvcModel = (("21", e._2.dvcModelLabel), e._1)
  val vendor = (("22", e._2.vendor), e._1)
  val (osAll, osCollect) = ((("23", e._2.osType), e._1), (("24", e._2.osType), e._2.dvcModel))
  val osLabel = e._2.dvcModelLabel match {
    case "EMPTY" => (("25", e._2.osType), "useless")
    case _ => (("25", e._2.osType), e._2.dvcModel)
  }

  Array(source, appName, appTag, appAll, appCollect, appLabel, dvcModel, vendor,
    osAll, osCollect, osLabel).filter(_._2 != "useless")
}

为了区别不同的维度组合，代码中采取了比较low的方式——为每个维度组合进行编号以示区别。Spark提供flatMap API将一行展开为多行，完美地满足了维度展开的需求；然后通过一把group by key + distinct count即可得到结果：

val flatRdd = logRdd.flatMap(flatAppDvc)
val result = flatRdd.distinct()
  .mapValues(_ => 1)
  .reduceByKey(_ + _)

多Measure

前面的分析需求比较简单，measure均为distinct count；因而可以不必对齐Dimensions + Measure。然而，对于比较复杂的分析需求：

• （整体上）广告物料的收集量、标注量、PV；
• （广告物料的）二级标注类别的广告物料数、UV、PV；
• （广告物料的）一级标注类别的广告物料数、UV、PV；

measure既有distinct count (UV) 也有count (PV)，这时需要Dimensions + Measure的对齐，维度flatMap如下：

/**
 * @param e ((adid, 2nd ad-category, 1st ad-category, uid)
 * @return Array[((dimension order No, dimension), measure:(adid, uid or adid, 1)]
 */
def flatAd(e: ((String, String, String), String)) = {
  val all = e._1._2 match {
    case "EMPTY" => (("0", "all"), (e._1._1, "non", 0))
    case _ => (("0", "all"), (e._1._1, e._1._1, 1))
  }
  val adCate = (("1", e._1._2), (e._1._1, e._2, 1))
  val adParent = (("2", e._1._3), (e._1._1, e._2, 1))

  Array(all, adCate, adParent)
}

尔后，计算每一维度的measure（其中distinct count采用HyperLogLogPlus算法的stream lib实现）：

val createHLL = (v: String) => {
  val hll = new HyperLogLogPlus(14, 0) // relative-SD = 0.01
  hll.offer(v)
  hll
}

def computeAdDimention(rdd: RDD[((String, String), (String, String, Int))]) = {
  rdd.combineByKey[(HyperLogLogPlus, HyperLogLogPlus, Int)](
    (v: (String, String, Int)) => (createHLL(v._1), createHLL(v._2), 1),
    (m: (HyperLogLogPlus, HyperLogLogPlus, Int), v: (String, String, Int)) => {
      m._1.offer(v._1)
      m._2.offer(v._2)
      val pv = m._3 + v._3
      (m._1, m._2, pv)
    },
    (m1: (HyperLogLogPlus, HyperLogLogPlus, Int),
     m2: (HyperLogLogPlus, HyperLogLogPlus, Int)) => {
      m1._1.addAll(m2._1)
      m1._2.addAll(m2._2)
      val pv = m1._3 + m2._3
      (m1._1, m1._2, pv)
    }
  )
    .mapValues(t => (t._1.cardinality().toInt, t._2.cardinality().toInt, t._3))
}

---

其实，本文有点标题党～～只是借了OLAP的壳做数据多维分析，距离真正的OLAP还是很远滴……