Applied MLOps to Maintain Model Freshness on Kubernetes
summary
A working Kubernetes pipeline that ingests Twitter via an Apache Flink consumer, naively counts hashtags in Redis, classifies movie summaries with a zero-shot classifier (Hugging Face NLI), indexes into Elasticsearch, and serves models with KFServing is presented; models and artifacts are versioned with DVC, judgments collected with Cupid, and code is on GitHub. Claim: ML freshness needs full-cycle automation because 53–87% of AI projects fail to reach production; the demo runs end-to-end and is reproducible.
AI-written summary · about
chapters (29)
- 0:00About Us
- 1:18Introduction
- 2:07What is MLOps?
- 2:59Easy to deploy, hard to maintain
- 6:14ML Technical Debt
- 8:01Back to MLOps
- 9:17What We Are Building
- 10:05Example Search Results
- 10:59Components
- 12:08Architecture
- 13:01The Data Set
- 13:31Twitter Stream Consumer
- 14:07How to classify movie overviews based on hashtags?
- 14:59Natural Language Inference (NLD)
- 15:59NLI Training Data
- 16:36NLI Model as a Zero-shot Learning Classifier
- 17:25Classifying Movie Overviews
- 18:15Using the Model
- 18:43Model Training
- 19:18Model Versioning
- 20:05Evaluating the Models
- 21:19Evaluation Example
- 22:11Deploying a Model
- 23:46Serving Model with KFServing
- 24:16Deploying with KFServing
- 24:33Monitoring
- 25:35Code and Running
- 26:26Why not use more of Kubeflow?
- 27:10Summary
related talks
13mMLOps: From AI Research to Production-Ready Systems
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
28mMastering MLOps: Building Future Ready ML Systems
26mKubernetes and MLOps for Scalable and Reproducible Generative AI
13mMLOps at Scale
34mThe State of AI and Machine Learning Operations in 2019
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
31mKubernetes and MLOps for Scalable and Reproducible Generative AI