有一個數據多維分析的任務：

• 日誌的周UV；
• APP的收集量及標註量，TOP 20 APP（周UV)，TOP 20 APP標註分類（周UV)；
• 手機機型的收集量及標註量，TOP 20 機型（周UV)，TOP 20 手機廠商（周UV)；

初始的解決方案：Spark讀取資料日誌，然後根據分析需求逐一進行map、distinct、reduceByKey得到分析結果。但是，這種方案存在著非常大的缺點——重複掃描資料來源多次。

1. Pig

Pig提供cube關鍵字做OLAP，將dimension分為了兩類：

• normal，對應於cube operation，\(n\)個該維度的組合數為\(2^n\)；
• hierarchical ordering，對應於rollup operation， \(n\)
  個該維度的組合數為\(n+1\)。

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的操作相當於按維度組合對每一record進行展開並group by Dimensions，與下一句foreach語句構成了Dimensions + Measure的資料輸出格式。

2. Spark

樸素多維分析

從上面介紹的pig OLAP方案中，我們得到靈感——面對開篇的多維分析需求，也可以每一條記錄按Dimensions + Measure的規則進行展開：

/**
 * @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")
}
 

為了區別不同的維度組合，程式碼中採取了比較low的方式——為每個維度組合進行編號以示區別。Spark提供flatMap API將一行展開為多行，完美地滿足了維度展開的需求；然後通過一把group by key + distinct count即可得到結果：

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

多Measure

前面的分析需求比較簡單，measure均為distinct count；因而可以不必對齊Dimensions + Measure。然而，對於比較複雜的分析需求：

• （整體上）廣告物料的收集量、標註量、PV；
• （廣告物料的）二級標註類別的廣告物料數、UV、PV；
• （廣告物料的）一級標註類別的廣告物料數、UV、PV；

measure既有distinct count (UV) 也有count (PV)，這時需要Dimensions + Measure的對齊，維度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)
}

爾後，計算每一維度的measure（其中distinct count採用HyperLogLogPlus演算法的stream lib實現）：

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))
}

其實，本文有點標題黨～～只是借了OLAP的殼做資料多維分析，距離真正的OLAP還是很遠滴……