當機器學習（ML）工作流程日益複雜，橫跨異質環境、工具鏈與執行脈絡時，輸入 、設定與輸出之間缺乏可靠的關聯性，已成為一項根本性挑戰。若無系統化的可追蹤性 ，工作流程容易出現不一致性，造成除錯困難、協作效率低落，以及維護成本上升。 本論文聚焦於一個模組化編排框架的架構設計與實作，旨在於整個機器學習工作流 程生命週期中，系統性地實現可追溯性。 該系統透過以下五個整合面向來增強追溯能力： 1.版本控制的資源管理：涵蓋資料集、程式碼、執行環境與模型； 2.基於 API 的任務抽象進行自動化工作流程整合； 3.工作流程結果紀錄，輸入對應與中繼資料追蹤； 4.集中式的任務層級日誌紀錄； 5.統一的執行時元件日誌紀錄。 這些機制共同支援工作流程執行脈絡的系統化記錄、關聯與檢視，從而在異質環境 中，於機器學習工作流程的各階段實現一致且可驗證的可追溯性。 ;As machine learning (ML) workflows grow in complexity—spanning heterogeneous environments, toolchains, and execution contexts—the lack of reliable associations between inputs, configurations, and outputs has emerged as a fundamental challenge. Without systematic traceability, workflows become prone to inconsistencies, hindered debugging, collaboration inefficiencies, and increased maintenance overhead. This thesis focuses on the architectural design and implementation of a modular orchestration framework to systematically enable traceability throughout the ML workflow lifecycle. The system enhances traceability through five integrated dimensions: (1) version-controlled resource management for datasets, code, environments, and models; (2) automated workflow integration via API based task abstraction; (3) structured input-output association and metadata tracking; (4) centralized task-level logging; and (5) unified runtime component logging. These mechanisms collectively support the systematic recording, correlation, and inspection of workflow execution contexts—thereby enabling consistent and verifiable traceability across all stages of ML workflows in heterogeneous environments.
