Exact reverse parquet scan counts shared buffers once per batch in its memory reservation

未关闭
  • 哪里不对？
#27,254 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@vinycalves 已经在做这个了。

开始于 2026年11月12日。

  • #27366 来自 @vinycalves —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
45/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
rust
领域
data-engineering, databases

调研方向

Start by reading the exact reverse scan implementation introduced for #27079 and its memory reservation logic, then inspect how batches from one row group share Arrow buffers. Reproduce with the issue's 1M-row, 4,000-distinct-string example and compare pool usage against the plan with the option off. Done means the reservation counts each distinct shared buffer once across held batches and the reproduction no longer fails below its actual memory use.

由索引模型根据 Issue 内容生成。

描述

With datafusion.execution.parquet.enable_exact_reverse_scan (#27079), the scan buffers one decoded row group and charges each batch's get_array_memory_size() to the memory pool. Batches of the same row group can share buffers, such as a page dictionary, so a shared buffer is counted once per batch. The scan cannot spill, so a query can fail with a resources error well below its real memory use.

Example from the #27079 review: 1M rows with 4000 distinct strings (about 25 MB). A 100 MB pool fails in the reverse scan, while the plan with the option off finishes at 30 MB by spilling in the sort.

The reservation should count each distinct buffer once across the batches it holds. This should be fixed before the option is enabled by default.

Related: #27079, #19048

主要语言
Rust
星标
9.6k
派生
2.6k
平均合并
3 天 4 小时
30 天内合并 PR
345

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

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 的其他 Issue

查看 apache/datafusion 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。