Spark SQL 实现分层抽样和分层随机抽样

缘起：TABLESAMPLE 的非随机性

最近需要实现一段 Spark SQL 逻辑，对数据集进行抽样指定的行数，才发现直接使用​​TABLESAMPLE​​函数抽样指定行数的方法其实是非随机的。

由于数据集较大，刚开始的逻辑是，取窗口函数随机排序后 row_number 的前 n 行。但运行速度较慢，所以想起了 TABLESAMLE 函数，支持直接取 Rows， 尝试后发现速度特别快，基本上几秒内就完成对亿级数据的采样。所以好奇就去查看文档和代码逻辑。

The TABLESAMPLE statement is used to sample the table. It supports the following sampling methods:

• TABLESAMPLE(x ROWS): Sample the table down to the given number of rows.
• TABLESAMPLE(x PERCENT): Sample the table down to the given percentage. Note that percentages are defined as a number between 0 and 100.
• TABLESAMPLE(BUCKET x OUT OF y): Sample the table down to a x out of y fraction.

Note: TABLESAMPLE returns the approximate number of rows or fraction requested.

文档中没有对实现逻辑有过多的说明，所以去代码中找问题。

源码中，匹配 SampleByRowsContext时，调用的方法是 Limit(expression(ctx.expression), query)，也就是说和 limit rows是一个逻辑。

而 SampleByPercentileContext实现的才是随机采样。

所以，如果对抽样的随机性有要求，还是老老实实用 SampleByPercentileContext，或者窗口函数。

总结：Spark SQL 随机抽样方法

随机抽样

抽取固定数量

使用窗口函数 + 随机排序进行抽样

WITH RankedData AS (
    SELECT *,
           row_number() OVER (ORDER BY rand(2077)) as rn
    FROM your_table
)
SELECT *
FROM RankedData
WHERE rn <= 1000 

抽取固定比例

直接使用TABLESAMPLE函数，实现对整体的固定比例抽样

SELECT *
FROM your_table
TABLESAMPLE (10 PERCENT)

分层随机抽样

分层抽样通常在数据科学中使用较多，为了保证样本的随机性，通常情况下，我们需要对 ​y​​标签进行分层抽样；如果考虑时间因素的影响，为了保证样本时间的随机性，通常还需要对​​月份 + y标签​​或者​​日期 + y标签​​进行双层的分层抽样。

抽取固定数量

WITH RankedData AS (
    SELECT *,
           row_number() OVER (PARTITION BY 分层字段 ORDER BY rand(2077)) as rn
    FROM your_table
)
SELECT *
FROM RankedData
WHERE rn <= 100  -- 每层抽取100条数据

抽取固定比例

对单字段分层

通常用在对时间的随机性要求不严格的场景，如在二分类任务中，可以将分桶字段设置为​​y​​列，那么就可以保证最终抽样出来的样本​​y​​均值和总体​​y​​均值相等：

WITH RankedData AS (
    SELECT *,
           row_number() OVER (PARTITION BY 分层字段 ORDER BY rand()) as rn,
           count(*) OVER (PARTITION BY 分层字段) as total_count
    FROM your_table
)
SELECT *
FROM RankedData
WHERE rn <= total_count * 0.1  -- 每层抽取10%的数据

对双字段分层

通常用在对时间的随机性也有严格要求的场景，这时可以将分层字段1和分层字段2分别设置为y列和时间列，那么就可以保证样本逐时间的y分布和整体随时间的y分布是近似的：

WITH RankedData AS (
    SELECT *,
           row_number() OVER (PARTITION BY 分层字段1, 分层字段2 ORDER BY rand()) as rn,
           count(*) OVER (PARTITION BY 分层字段1, 分层字段2) as total_count
    FROM your_table
)
SELECT *
FROM RankedData
WHERE rn <= total_count * 0.1  -- 每层抽取10%的数据

附 相关源码：

  /**
   * Add a [[Sample]] to a logical plan.
   *
   * This currently supports the following sampling methods:
   * - TABLESAMPLE(x ROWS): Sample the table down to the given number of rows.
   * - TABLESAMPLE(x PERCENT) [REPEATABLE (y)]: Sample the table down to the given percentage with
   * seed 'y'. Note that percentages are defined as a number between 0 and 100.
   * - TABLESAMPLE(BUCKET x OUT OF y) [REPEATABLE (z)]: Sample the table down to a 'x' divided by
   * 'y' fraction with seed 'z'.
   */
  private def withSample(ctx: SampleContext, query: LogicalPlan): LogicalPlan = withOrigin(ctx) {
    // Create a sampled plan if we need one.
    def sample(fraction: Double, seed: Long): Sample = {
      // The range of fraction accepted by Sample is [0, 1]. Because Hive's block sampling
      // function takes X PERCENT as the input and the range of X is [0, 100], we need to
      // adjust the fraction.
      val eps = RandomSampler.roundingEpsilon
      validate(fraction >= 0.0 - eps && fraction <= 1.0 + eps,
        s"Sampling fraction ($fraction) must be on interval [0, 1]",
        ctx)
      Sample(0.0, fraction, withReplacement = false, seed, query)
    }
 
    if (ctx.sampleMethod() == null) {
      throw QueryParsingErrors.emptyInputForTableSampleError(ctx)
    }
 
    val seed = if (ctx.seed != null) {
      ctx.seed.getText.toLong
    } else {
      (math.random() * 1000).toLong
    }
 
    ctx.sampleMethod() match {
      case ctx: SampleByRowsContext =>
        Limit(expression(ctx.expression), query)
 
      case ctx: SampleByPercentileContext =>
        val fraction = ctx.percentage.getText.toDouble
        val sign = if (ctx.negativeSign == null) 1 else -1
        sample(sign * fraction / 100.0d, seed)
 
      case ctx: SampleByBytesContext =>
        val bytesStr = ctx.bytes.getText
        if (bytesStr.matches("[0-9]+[bBkKmMgG]")) {
          throw QueryParsingErrors.tableSampleByBytesUnsupportedError("byteLengthLiteral", ctx)
        } else {
          throw QueryParsingErrors.invalidByteLengthLiteralError(bytesStr, ctx)
        }
 
      case ctx: SampleByBucketContext if ctx.ON() != null =>
        if (ctx.identifier != null) {
          throw QueryParsingErrors.tableSampleByBytesUnsupportedError(
            "BUCKET x OUT OF y ON colname", ctx)
        } else {
          throw QueryParsingErrors.tableSampleByBytesUnsupportedError(
            "BUCKET x OUT OF y ON function", ctx)
        }
 
      case ctx: SampleByBucketContext =>
        sample(ctx.numerator.getText.toDouble / ctx.denominator.getText.toDouble, seed)
    }
  }