`CometBenchmarkBase.runExpressionBenchmark` checks the initial AQE plan, not the executed plan

未关闭
  • 哪里不对？
#7,481 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 2 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
55/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
scala, spark
领域
performance, testing

调研方向

Read spark/src/test/scala/org/apache/spark/sql/benchmark/CometBenchmarkBase.scala and spark/src/test/scala/org/apache/spark/sql/comet/CometPlanChecker.scala. Start by checking how the benchmark executes its DataFrame and how the plan checker traverses AQE query stages. Done means the check inspects the final executed plan, recognizes the AQE wrapper nodes, and reports any non-native operator in that plan.

由索引模型根据 Issue 内容生成。

描述

bug priority:low
Describe the bug

Before timing a case, runExpressionBenchmark in
spark/src/test/scala/org/apache/spark/sql/benchmark/CometBenchmarkBase.scala checks that the
Comet plan is fully native:

// Check that the plan is fully Comet native before running the benchmark. Unlike the check
// above this one does execute the query, because AQE can re-plan a join after the first
// stage completes and the Comet operators are what we are checking for.
withSQLConf(cometConfigs: _*) {
  val df = spark.sql(query)
  df.noop()
  val plan = stripAQEPlan(df.queryExecution.executedPlan)
  findFirstNonCometOperator(plan).foreach { op => ...

df.noop() runs a separate write query. It does not execute df's own plan. After it,
df.queryExecution.executedPlan is still AdaptiveSparkPlan isFinalPlan=false, and
stripAQEPlan returns the initial plan. The check therefore never sees what AQE changes during
execution, which is what the comment says it is checking for. The warning it prints shows the
initial plan, not the plan that ran.

Steps to reproduce

Run on main (0f7f8e453) with the default profile (Spark 4.1.3), in a CometBenchmarkBase
subclass, over a Parquet table of 4096 rows made with
spark.range(4096).selectExpr("id % 10 AS k", "id AS v"), with
spark.comet.exec.shuffle.enabled=false. The query is SELECT k, sum(v) FROM t GROUP BY k.

  1. With the three lines above, after df.noop(), df.queryExecution.executedPlan is
    AdaptiveSparkPlan isFinalPlan=false. The plan that the check reads, and that the warning
    prints, is
    HashAggregate > Exchange > CometHashAggregate > CometNativeScan.
    It has no CometColumnarToRow, AQEShuffleRead or ShuffleQueryStage.
  2. For a second DataFrame of the same query, after collect(),
    executedPlan is AdaptiveSparkPlan isFinalPlan=true, and the plan is
    ResultQueryStage > HashAggregate > AQEShuffleRead coalesced > ShuffleQueryStage >
    Exchange > CometColumnarToRow > CometHashAggregate > CometNativeScan.
Expected behavior

The check reads the plan that ran, after AQE has finished, and the warning prints that plan.

Additional context

Executing the DataFrame's own plan, for example with
df.queryExecution.executedPlan.execute().foreach(_ => ()), makes the plan final. The final plan
contains query stages. findFirstNonCometOperator in
spark/src/test/scala/org/apache/spark/sql/comet/CometPlanChecker.scala walks the plan with
foreach, which does not descend into a QueryStageExec (a leaf node), and its allow-list
contains neither QueryStageExec nor AQEShuffleReadExec. So a fix also needs the walk to go
into the query stages, for example through AdaptiveSparkPlanHelper, and to allow the AQE
wrapper nodes.

🤖 Generated with Claude Code

主要语言
Scala
星标
1.3k
派生
414
平均合并
3 天 5 小时
30 天内合并 PR
210

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

apache/datafusion-comet 的其他 Issue

查看 apache/datafusion-comet 的全部 Issue

相似的 Issue

更多 Scala Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。