Hacktoberfest 2026：維護者為十月標記出來的 issue，仍然開放、適合新手。 瀏覽 Hacktoberfest issue

Decode Python UDFs opaquely so a scheduler needs no Python interpreter

未關閉
  • 哪裡不對？
#1,705 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
35/100
Issue 類型
功能
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
python, rust
領域
backend, distributed-systems

研究方向

先閱讀 crates/core/src/codec.rs 以及 #1678 中的 extension codec 工作，以了解 DFPYUDF wire format 與 codec 組合方式。圍繞現有 payload 設計 scheduler 端的 ScalarUDFImpl 和 codec，在重新編碼時保留 cloudpickle 位元組，同時公開可復原的 metadata，並讓 invoke 回傳錯誤。檢查 #1703 中的相依性脈絡。

由索引模型根據 Issue 內容生成。

描述

enhancement rust

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

In a distributed setup the scheduler plans a query and hands stages to executors; only the executors ever call a Python UDF. Decoding an inlined Python UDF unpickles the function, which requires a Python interpreter and every module the function closes over to be importable. Doing that on the scheduler costs work nobody needs and forces the scheduler image to carry Python and the full dependency set of user code it will never run. Raised in https://github.com/apache/datafusion-python/pull/1678#pullrequestreview-5100366976.

Describe the solution you'd like

An opaque decode path: a ScalarUDFImpl that holds the still-pickled blob rather than a live Python object, and a codec that produces it. A scheduler installs that codec, decodes a plan into something it can inspect, route, and re-encode, and never touches cloudpickle. The executor installs the ordinary codec and unpickles as it does now.

The wire format already allows this. An inlined UDF payload is DFPYUDF followed by a version byte and the cloudpickle blob (crates/core/src/codec.rs), so an opaque holder can carry those bytes verbatim and no format change is needed.

The part that needs design is re-encoding. A scheduler that forwards a stage has to emit the blob byte-identically, so the executor sees exactly what the client wrote. That also raises what such a UDF should report for the things DataFusion asks of a ScalarUDFImpl during planning — name, signature, and return type are all recoverable from the payload without unpickling, since they are stored alongside the function, but invoke has to be an error rather than a surprise.

Describe alternatives you've considered

Encoding Python UDFs by name only and registering them on every node. Already supported and appropriate when the function is available everywhere; it does not cover the case inlining exists for, which is a function the receiving process does not have.

Having the scheduler unpickle and immediately drop the object. Keeps the code simple, and still requires Python plus all user dependencies on the scheduler, which is the actual cost being avoided.

Additional context

Follow-up from #1678, which made extension codecs compose so a setup like this can install a scheduler-side codec alongside others. Likely also depends on #1703, gating pyo3/extension-module, if the consumer is a Rust crate rather than a Python process.

主要語言
Python
星號
606
分支
176
平均合併
1 天 7 小時
30 天內合併 PR
14

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 有 Pull Request 範本
  • 沒有貢獻指南

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

apache/datafusion-python 的其他 Issue

查看 apache/datafusion-python 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。