[EPIC] Support Spark 4.3

未关闭
  • 哪里不对？
#7,519 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
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
github-actions, scala, spark
领域
backend, build-system, ci-cd, testing

调研方向

Start with the staged recipe in docs/source/contributor-guide/adding_a_new_spark_version.md and the draft Stage 1 PR, #7518. The issue lists six stages spanning Maven profiles, shims, CI, tests, documentation, and support promotion; begin by checking which stages remain open and their linked work. Done means completing the required stages, including resolving skipped tests and promoting Spark 4.3 to supported.

由索引模型根据 Issue 内容生成。

描述

enhancement EPIC

Summary

Spark 4.3.0 is released (tag v4.3.0). This tracks bringing up Comet support for it, following the staged recipe in the contributor guide, "Adding Support for a New Spark Version" (docs/source/contributor-guide/adding_a_new_spark_version.md).

Spark 4.3.0 publishes the same Java 17, Scala 2.13.18, Parquet 1.17.0, slf4j 2.0.17, Hadoop 3.5.0 and Arrow 19.0.0 versions as 4.2.0, so the build profile is a copy of spark-4.2.

Stages

  • Stage 1: Maven profile, shims, isSpark43Plus, compile-only lint-java CI row. Draft PR: #7518
  • Stage 2: enable Comet's own JVM test suites under -Pspark-4.3, skipping failures with assume() and a linked issue each.
  • Stage 3: Spark SQL tests (dev/diffs/4.3.0.diff, a spark_4_3 job in ci.yml, compute-changes.py and check-ci-config.py) and plan stability golden files.
  • Stage 4: add 4.3 to the experimental table and jar list in installation.md.
  • Stage 5: fix the skipped tests, one issue or cluster per PR.
  • Stage 6: promote from experimental to supported.

Known Spark 4.3 changes to plan for

Found while doing Stage 1:

  • spark.shuffle.manager must be a BlockingShuffleManager (new in 4.3), and shuffleBlockResolver moved onto it. Comet's shuffle managers need the shim from the Stage 1 PR; the Celeborn integration needs a Celeborn build that targets 4.3, so it should be checked at Stage 2.
  • SpecializedGetters gained getTimestampNTZNanos and getTimestampLTZNanos, for new nanosecond timestamp types. Comet currently throws for them; decide per type whether to fall back or support them.
  • BasePythonRunner passes an Option[PythonWorkerHandle] instead of a pid.
  • SparkEnv.shuffleManager is deprecated in favour of shuffleManagerFor(dependency); there is also a new pipelined shuffle manager kind, which Comet's shuffle planning should not mistake for a blocking one.
  • SimpleCostEvaluator counts local sorts (spark.sql.adaptive.costEvaluator.countLocalSort.enabled), which can change AQE plan choices over Comet plans.
  • An opt-in Arrow cache serializer: tracked separately in #7517.

Not yet decided

  • Which Delta and Iceberg versions to pair with the 4.3 profile. Delta has no 4.3 release; the profile reuses 4.3.1 for compilation only, as spark-4.2 does. Iceberg tests reuse the 4.0 runtime.
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。