Hi,
I'm working through an architectural decision for our team's production machine learning platform and would really value the perspective who have been through this in practice.
The core question is whether Fabric Data Science alone is sufficient for a production-grade MLOps practice, or whether pairing it with Azure Machine Learning using Fabric for data preparation, exploration, and experiment tracking in OneLake, and Azure ML for deployment, model monitoring, and CI/CD is genuinely worth the added complexity.
From the official documentation, a few things stand out as potential friction points for a Fabric-only path: real-time online endpoints are still in preview, there doesn't appear to be a native production model-monitoring or drift-detection story comparable to what Azure ML offers, and the CI/CD approach for model pipelines seems notebook/item-oriented rather than a full ML-pipeline framework. Compute is another dimension. Fabric runs everything through capacity-based Spark pools, which means ML training workloads share the same capacity as your data engineering, warehousing, and BI workloads whereas Azure ML gives you dedicated compute tiers (instances, clusters, serverless, and Kubernetes) with GPU options and fine-grained per-job control. In theory, that shared capacity model could create contention during heavy training runs or lead to unpredictable performance at scale, but I'm genuinely curious whether anyone have hit that in practice or whether the Spark engine handles it well enough for most ML workloads.
What I'd love to hear who have actually shipped something: did you find Fabric Data Science sufficient end-to-end for a production scenario, and if so, what did that look like? If you went hybrid, what was the tipping point, and were there integration headaches the documentation doesn't warn you about? And if you started one way and migrated to the other, what would you do differently?
Any gotchas, lessons learned, or honest "it depends on X" answers would be hugely helpful especially around monitoring, deployment reliability, compute performance under load, and keeping CI/CD manageable. Thanks in advance.
