Reading a Variant column from Parquet files with different shredding layouts null-fills typed_value; every shredded field then reads as NULL, with no error

未关闭
  • 哪里不对？
#27,089 10 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

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
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
rust
领域
backend, databases

调研方向

Start with the linked schema inference and rewrite functions in datafusion/datasource-parquet/src/file_format.rs and datafusion/physical-expr-adapter/src/schema_rewriter.rs, then read #26282 and the Variant shredding specification. Reproduce with the two mixed-layout Parquet cases in the issue and trace how the adapter handles the missing typed_value children. Done means mixed layouts preserve the Variant fields across files, with coverage for both cases.

由索引模型根据 Issue 内容生成。

描述

Part of #26517.

Is your feature request related to a problem or challenge?

I have a table of Parquet files with one Variant column, event. Each row is a small JSON object like {"service":"api","status":200,"latency_ms":12}. The files were not all written the same way:

  • a.parquet is shredded on service and status. Those two fields are stored as typed columns under typed_value; the rest stays in value.
  • b.parquet is not shredded. The whole object is in value, there is no typed_value.

Both files hold the same 3 rows. Spark reads a table like this fine. Each file also reads fine on its own in DataFusion.

Query (with datafusion-variant registered):

SELECT variant_get(event, 'service', 'VARCHAR') AS service FROM events

What I get:

+---------+
| service |
+---------+
| api     |   <- a.parquet
| api     |
| db      |
|         |   <- b.parquet
|         |
|         |
+---------+

What I expect: api, api, db for both files, so 6 non-null rows. That is what each file returns on its own.

What happens: the merged table schema has typed_value, and b.parquet does not. The schema adapter fills typed_value with NULL for that file. For a plain struct that is right. For Variant it is not: the shredding spec says a NULL typed_value means "this value is not an object", so the reader never looks in value for service. The data is still there, it is just skipped. No error, no warning.

The same thing happens when two files are shredded on different fields. Before #26195 that case failed at open; on branch-56 it null-fills too.

Describe the solution you'd like

Let an extension type decide how a file's layout maps to the table's layout, instead of the default struct null-fill. The Variant version would live in datafusion-variant. Per the discussion below I am starting with a PhysicalExprAdapter there and will report back on what, if anything, core needs.

Additional context

Repro: a 90-line Rust program that writes the two files with arrow-rs shred_variant and runs the query. It is in the edit history of this issue; I will replace it with a SQL-only version once datafusion-contrib/datafusion-variant#145 (COPY ... OPTIONS ('variant.shredding' ...)) lands.

DataFusion branch-56 at 1635ed9, arrow/parquet 60.1, datafusion-variant 642248a.

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。