Abstract
The unique nature of machine learning (ML) software systems, characterized by a high level of uncertainty and a crucial dependency on data, has led to challenges for traditional DevOps practices. As a result, a new domain entitled MLOps emerged, which considered these specifics. The evolution of MLOps is aligned with its specialization on certain quality attributes, such as security (SecMLOps) or reliability (SafeMLOps). However, due to their focus on one exclusive quality characteristic, such frameworks have limited applicability for production aimed at achieving multiple quality objectives at once (e.g., high reliability with the least resources consumed). Explicitly managing the trade-offs between different, potentially competing, quality objectives can help organizations by enhancing the flexibility and predictability of the MLOps workflow. This vision paper presents a vision around the novel notion of MLTradeOps, focused on explicitly managing trade-offs during the MLOps workflow. It brings together the expertise of existing and emerging DevOps branches focused on specific quality attributes, and also ongoing monitoring and addressing of other general quality characteristics of in-production software systems. We envision a framework that makes trade-off management a core part of the decision-making process and contains a high-level cycle to make conscious trade-offs for the ML-enabled system, which are then reflected in lower-level decisions during the MLOps lifecycle. We supplement our vision with a roadmap for the potential formation of this framework.
