alibaba/GraphScope

[BUG] Loading from large dataframe/large numpy requires holding all chunks in coordinator

Open

#2,342 建立於 2022年12月23日

在 GitHub 查看
 (0 留言) (1 反應) (0 負責人)HTML (2,401 star) (301 fork)batch import
bugcomponent:coordinatorgood first issue

描述

Describe the bug

It looks strange that we need to accumulate all chunks in the request stream into a list in coordinator before sending to analytical engine, that would requires large available memory for the coordinator pod.

https://github.com/alibaba/GraphScope/blob/b80a35599424580325a750e734f8a3b2dead2a5b/coordinator/gscoordinator/dag_manager.py#L77-L107

貢獻者指南

技術棧
pythonnumpypandas
領域
backend
議題類型
bug
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
3
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
half day
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
PythonNumPyGraphScope coordinator architecture
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
30
研究方向
Investigate the coordinator code in dag manager.py lines 77-107. The issue is that all chunks are accumulated in a list before being sent to the analytical engine, causing high memory usage. Consider modifying the code to stream chunks directly to the engine or send them in smaller batches. Look at the AnalyticalEngine interface to see if it supports streaming input. No comments or PRs exist, so this is an open opportunity for optimization.