MLOps at Snapchat: Continuous Machine Learning with Kubeflow & Spinnaker
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Thursday, 2020-11-19· at 16:50· Intrado Virtual Event Platform,·
summary
Snap implements continuous ML for Scan computer vision by containerizing notebook code into modular Docker components, wiring those into Kubeflow Pipelines for data validation feature extraction training and evaluation, and driving deployments with Spinnaker pipelines that submit runs, parse outputs, and trigger model server canary rollouts. CI uses Jenkins/Drone and pushes images to GCR, CD schedules training via cron or Pub/Sub, and monitoring tracks latency, proxy metrics like CTR shifts, and automated rollbacks on degraded model performance.
AI-written summary · about
chapters (10)
- 0:00Start
- 8:22break up your experimental jupiter notebooks into modular code
- 9:04automate your machine learning process
- 12:33set up your favorite ci tool
- 13:51orchestrate the deployment of your ml pipeline
- 15:29deploying our ml pipeline
- 16:08running the ml pipeline stage
- 17:11update your cd pipeline
- 19:25detect deviations in model performance
- 19:47track the results of acting on your predictions
related talks
37mTowards Continuous Computer Vision Model Improvement with Kubeflow
07mMLOps with Spinnaker: through the eyes of a beginner
31mKubeflow 2026 State of the Union: Orchestrating the Full ML Lifecycle
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
29mTeam Teardown: MLOps in the Cloud for IoT
05mCI/CD Pipelines and Machine Learning
46mAccelerating Machine Learning DevOps with Kubeflow
34mManaging Machine Learning in Production with Kubeflow and DevOps