MLOps 2025: A Journey into the Past and the Future
summary
History traces platforms like Michelangelo, SageMaker, TFX and runtimes like Flink, Ray, Metaflow while orchestration and experiment tools such as Kubeflow, Prefect and Weights & Biases proliferate; feature stores and multimodel serving scheduler prototypes run as Kubernetes CRDs but real deployment is messy: a 2024 survey finds 60% custom Python serving and 50% without ML monitoring. Built artifacts include explainers and drift detectors; predictions peg personal agents by 2027, unified MLOps by 2030 and commoditized self‑healing systems by 2035.
AI-written summary · about
chapters (21)
- 0:00MLOps journey intro
- 1:20Speaker background
- 2:26Why MLOps matters
- 4:54ML vs traditional software
- 8:22MLOps genesis 2015-2017
- 10:30Uber’s Michelangelo
- 13:19ML flaws and security
- 14:45MLOps innovation 2018-2021
- 16:54Real-time ML serving
- 19:43ML monitoring challenges
- 22:06COVID-19 MLOps boost
- 23:53MLOps solidified
- 24:35GenAI and LLMs emerge
- 25:17LLM ops era
- 27:03LLM ops security
- 29:10Agentic stack emerges
- 30:13MLOps in 2025
- 31:38MLOps predictions 2027-2035
- 36:39Reflections and takeaways
- 38:02MLOps engineering funnel
- 40:51Reducing MLOps complexity
related talks
13mMLOps: From AI Research to Production-Ready Systems
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
13mMLOps at Scale
28mMastering MLOps: Building Future Ready ML Systems
45mThe Culture, Architecture and Technology of MLOps
47mMachine Learning in Production: An Intro to MLOps.
34mThe State of AI and Machine Learning Operations in 2019
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE