如何理解spark窗口函数的结果

Question

我有以下代码来尝试spark窗口函数：

test("spark sql time window 2") {
    val spark = SparkSession.builder().master("local").appName("SparkSQLWindowTest").getOrCreate()
    import spark.implicits._
    import org.apache.spark.sql.functions._
    val ds = Seq(
      SaleRecord("2017-10-11 09:01:12", 1),
      SaleRecord("2017-10-11 09:01:18", 6),
      SaleRecord("2017-10-11 10:11:12", 2),
      SaleRecord("2017-10-11 10:18:13", 5),
      SaleRecord("2017-10-11 10:22:13", 3),
      SaleRecord("2017-10-11 10:22:22", 6),
      SaleRecord("2017-10-11 10:34:56", 2),
      SaleRecord("2017-10-11 10:48:22", 6),
      SaleRecord("2017-10-11 11:52:23", 4),
      SaleRecord("2017-10-11 12:56:24", 2)).toDS

    val ds2 = ds.groupBy(window($"Time", "20 minutes", "9 minutes")).agg(sum("revenue")).orderBy("window.start")
    ds2.show(truncate = false)

    /*
+---------------------------------------------+------------+
|window                                       |sum(revenue)|
+---------------------------------------------+------------+
|[2017-10-11 08:45:00.0,2017-10-11 09:05:00.0]|7.0         |
|[2017-10-11 08:54:00.0,2017-10-11 09:14:00.0]|7.0         |
|[2017-10-11 09:57:00.0,2017-10-11 10:17:00.0]|2.0         |
|[2017-10-11 10:06:00.0,2017-10-11 10:26:00.0]|16.0        |
|[2017-10-11 10:15:00.0,2017-10-11 10:35:00.0]|16.0        |
|[2017-10-11 10:24:00.0,2017-10-11 10:44:00.0]|2.0         |
|[2017-10-11 10:33:00.0,2017-10-11 10:53:00.0]|8.0         |
|[2017-10-11 10:42:00.0,2017-10-11 11:02:00.0]|6.0         |
|[2017-10-11 11:36:00.0,2017-10-11 11:56:00.0]|4.0         |
|[2017-10-11 11:45:00.0,2017-10-11 12:05:00.0]|4.0         |
|[2017-10-11 12:39:00.0,2017-10-11 12:59:00.0]|2.0         |
|[2017-10-11 12:48:00.0,2017-10-11 13:08:00.0]|2.0         |
+---------------------------------------------+------------+


     */
  }

SaleRecord的定义是一个简单的案例类：

case class SaleRecord(time: String, revenue: Double)

我无法理解前三行是如何在结果中生成的？

为什么第一个窗口是[2017-10-11 08:45:00.0,2017-10-11 09:05:00.0]

Answer 1

window(timeColumn, windowDuration, slideDuration=None, startTime=None)

首先，window函数将创建一个时间模板，这意味着：

zero = 1970-01-01 00:00:00 UTC

[zero + startTime + slideDuration * n, zero + startTime + slideDuration * n + windowDuration)

例如：

window('ts', '5 seconds', '3 seconds', '2 seconds')
# is equal to :
['1970-01-01 00:00:02', '1970-01-01 00:00:07'),
['1970-01-01 00:00:05', '1970-01-01 00:00:10'),
['1970-01-01 00:00:08', '1970-01-01 00:00:13'),
['1970-01-01 00:00:11', '1970-01-01 00:00:16'),
...

然后，DataFrame中的每一行都将“落入”时间模板，根据timeColumn的说法。一行可以属于多个时间模板单元格。

最后，删除所有空时间模板单元并执行agg。

Answer 2

使用windows功能，星火将计算来自unix时代时间1970-01-01 00:00:00 UTC的帧。因为您将淤泥持续时间设置为9 minutes，所以第一个框架包含来自Time列的值[2017-10-11 08:45:00.0,2017-10-11 09:05:00.0]。

为清晰起见

$ date --date="2017-10-11 08:45:00" +"%s
1507686300
$ echo $[1507686300%180]
0