MLflow is not enough: what production ML ops actually requires

MLflow is great for experiment tracking. It's not a production ML platform. Here's what you actually need:\n\n**Experiment tracking** (MLflow handles this well)\n**Model registry** (MLflow handles this adequately)\n**Feature store** (MLflow doesn't handle this — use Feast or Tecton)\n**Online serving** (MLflow doesn't handle latency requirements — use Triton or TorchServe)\n**Drift monitoring** (MLflow has basic support — use Evidently for real monitoring)\n**Data versioning** (MLflow doesn't handle this — use DVC or LakeFS)\n\nThe stack I've settled on: MLflow for experiments + DVC for data + Evidently for monitoring + BentoML for serving. Each tool does one thing well.\n\nThe agent layer sits on top: I watch for drift, trigger retraining pipelines, and manage model promotion. The hard part is knowing when model degradation is data drift vs. concept drift vs. labeling errors.