Comet Native Executors: report each stage's Comet block metrics to Ballista

未关闭
  • 哪里不对？
#7,465 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
一周以上
新手友好度
22/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
rust, scala, spark
领域
data-engineering, distributed-systems

调研方向

The block is CometFragmentExec, which reports no metrics; its metrics() and the plan it builds inside execute are where the change starts. Ballista's stage metric aggregation, which sums elapsed_compute over the stage's operators and counts input rows from plan leaves, is the other side. Done means a two-stage query shows non-zero compute time and correct input rows for the range stage in the scheduler event log.

由索引模型根据 Issue 内容生成。

描述

enhancement requires-triage
What is the problem the feature request solves?

Ballista takes a task's metrics from the operators of its stage plan. On a Comet Native Executor cluster, a stage's plan is a shuffle writer over a CometFragmentExec, whose children are the stage's partition descriptor and its shuffle readers. CometFragmentExec reports no metrics, and the Comet operators run in a plan it builds inside execute, which Ballista never sees. So in the scheduler's event logs (COMET_SCHEDULER_EVENT_LOG_DIR, #7420) and stage summaries:

  • A task's compute time (elapsed_compute_nanos) covers only its shuffle readers' poll time, and is 0 for a stage with no shuffle input.
  • Input rows count only the rows a stage reads from other stages, so a stage that reads files or makes its own rows logs 0.
  • The plan shows each block only as CometFragmentExec: block=N bytes.

In a two-stage query, a range stage hash-shuffled into a stage that reads it, the range stage's task logged 0 input rows and 0 compute time, and the reading stage's task logged 1000 input rows and 3.3 ms, the shuffle reader's poll time. Output rows and task times are right.

Describe the potential solution

CometFragmentExec could keep the plan it builds for each partition and report the block's metrics from metrics(): its operators' elapsed_compute, and the output rows of its file scans and other leaves. Ballista sums elapsed_compute over every operator of the stage, so the block's figure shouldn't count the time its inputs spend in the shuffle readers, which report their own. Ballista counts a stage's input rows from the plan's leaves, and CometFragmentExec isn't one, so rows read from files need another way in, whether a leaf that reports them or a change to how Ballista counts input rows.

#7373 needs the block's per-operator metrics too, to fill in the Spark UI.

Additional context

Found in the review of #7420: https://github.com/apache/datafusion-comet/pull/7420#discussion_r4460104688

Part of #7288.

Willingness to contribute

Yes. I can contribute this feature independently.

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。