Infrastructure for ML Applications
summary
Argues that ML model code is only 5–10% of production and the rest is data pipelines, governance, monitoring and serving; recommends concrete tools: TensorFlow 2.0 for models, Hadoop/Spark or cloud data lakes for large datasets, Aerospike-style key lookups for low-latency feature reads, and serverless or TF Serving for inference. Distinguishes offline lookup tables vs online adaptive models, stresses data versioning so rollbacks and deletion requests work, and shows a PayPal-style fraud pipeline that meets latency and accuracy SLAs.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:12Captcha Challenge
- 10:57Infrastructure
- 18:55Types of Training
- 19:33Dynamic Model
- 20:30Online
- 29:33Recap
related talks
19mSustaining ML in the enterprise
29mDemystifying Machine Learning in Production: Reasoning about a Large-Scale ML Platform
33mBuilding Resilient (ML) Pipelines for MLOps
33mAI with a DevOps mindset - experimentation, sharing ..
24mSmart approach in development and deployment process for ML models
16mOperationalizing AI in Finance: Real-Time Analytics
26mPatterns of ML Models in Production
32mMaking Models Work: How an optimized ML lifecycle can drive better ROI