Leveraging GraphQL for Continual Learning in Real-Time ML Systems
summary
Claims a producible pipeline that turns GraphQL mutations into low-latency online features, uses Kafka to stream session events, Flink to do in-flight feature engineering, and ksqlDB for fast feature storage, with Airflow/MLflow driving triggers, fine-tune experiments and blue‑green deployments. Notes batch prediction wastes work (example: daily batches but only 5% users active => 95% wasted). Architecture supports fault tolerance, rollback and staged online evaluation but calls out added Flink/model latency and risks like catastrophic forgetting.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:42What is ML
- 7:40Why are batch predictions suboptimal
- 9:15Collecting realtime data
- 17:36What is continual learning
- 18:48New data collection
- 20:20Introducing Airflow
- 21:46Introducing MLflow
- 22:58Testing and Deployment
- 24:32BlueGreen Deployment
- 25:57Model Subgraph
- 27:49Benefits
- 29:07Conclusions
- 29:48Next Steps
related talks
24mUnder Pressure: Applying ML in Real Time
33mBuilding Resilient (ML) Pipelines for MLOps
14mPerformance Optimization in High-Volume Transaction Systems
38mReal Time ML Pipelines in Multi-Tenant Environments
31mApplied MLOps to Maintain Model Freshness on Kubernetes
25mContinuous ML Applications in Production
28mEnabling Machine Learning with Apache Flink
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow