[SPARK-20895] Support fast execution based on an optimized plan and parameter placeholders

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Later
• Affects Version/s: 2.1.1
• Component/s: SQL
• Assignee: Unassigned
• Reporter: Takeshi Yamamuro
• Created: 26/May/17 08:14
• Updated: 29/May/17 00:01
• Resolved: 28/May/17 23:30

Description

In database scenarios, users sometimes use parameterized queries for repeated execution (e.g., by using prepared statements).
So, I think this functionality is also useful for Spark users.
What I suggest here seems to be like:
My prototype here: https://github.com/apache/spark/compare/master...maropu:PreparedStmt2

scala> Seq((1, 2), (2, 3)).toDF("col1", "col2").createOrReplaceTempView("t")

// Define a query with a parameter placeholder named `val`
scala> val df = sql("SELECT * FROM t WHERE col1 = $val")
scala> df.explain
== Physical Plan ==
*Project [_1#13 AS col1#16, _2#14 AS col2#17]
+- *Filter (_1#13 = cast(parameterholder(val) as int))
   +- LocalTableScan [_1#13, _2#14]

// Apply optimizer rules and get an optimized logical plan with the parameter placeholder
scala> val preparedDf = df.prepared

// Bind an actual value and do execution
scala> preparedDf.bindParam("val", 1).show()
+----+----+
|col1|col2|
+----+----+
|   1|   2|
+----+----+

To implement this, my prototype adds a new expression leaf node named `ParameterHolder`.
In a binding phase, this node is replaced with `Literal` including an actual value by using `bindParam`.
Currently, Spark sometimes consumes much time to rewrite logical plans in `Optimizer` (e.g. constant propagation desribed in SPARK-19846).
So, I feel this approach is also helpful in that case:

def timer[R](f: => {}): Unit = {
  val count = 9
  val iters = (0 until count).map { i =>
    val t0 = System.nanoTime()
    f
    val t1 = System.nanoTime()
    val elapsed = t1 - t0 + 0.0
    println(s"#$i: ${elapsed / 1000000000.0}")
    elapsed
  }
  println("Avg. Elapsed Time: " + ((iters.sum / count) / 1000000000.0) + "s")
}

import org.apache.spark.sql.Row
import org.apache.spark.sql.types._
val numCols = 50
val df = spark.range(100).selectExpr((0 until numCols).map(i => s"id AS _c$i"): _*)
// Add conditions to take much time in Optimizer
val filter = (0 until 128).foldLeft(lit(false))((e, i) => e.or(df.col(df.columns(i % numCols)) === (rand() * 10).cast("int")))
val df2 = df.filter(filter).sort(df.columns(0))

// Regular path
timer {
  df2.filter(df2.col(df2.columns(0)) === lit(3)).collect
  df2.filter(df2.col(df2.columns(0)) === lit(4)).collect
  df2.filter(df2.col(df2.columns(0)) === lit(5)).collect
  df2.filter(df2.col(df2.columns(0)) === lit(6)).collect
  df2.filter(df2.col(df2.columns(0)) === lit(7)).collect
  df2.filter(df2.col(df2.columns(0)) === lit(8)).collect
}

#0: 24.178487906
#1: 22.619839888
#2: 22.318617035
#3: 22.131305502
#4: 22.532095611
#5: 22.245152778
#6: 22.314114847
#7: 22.284385952
#8: 22.053593855
Avg. Elapsed Time: 22.519732597111112s

// Prepared path
val df3b = df2.filter(df2.col(df2.columns(0)) === param("val")).prepared

timer {
  df3b.bindParam("val", 3).collect
  df3b.bindParam("val", 4).collect
  df3b.bindParam("val", 5).collect
  df3b.bindParam("val", 6).collect
  df3b.bindParam("val", 7).collect
  df3b.bindParam("val", 8).collect
}

#0: 0.744693912
#1: 0.743187129
#2: 0.745100003
#3: 0.721668718
#4: 0.757573342
#5: 0.763240883
#6: 0.731287275
#7: 0.728740601
#8: 0.674275592
Avg. Elapsed Time: 0.7344186061111112s

I'm not sure this approach is acceptable, so welcome any suggestion and advice.