Embracing Multi-Tenancy While Scaling MLOps
summary
Demo shows Flyte running on Kubernetes with two projects (demo data science and demo mlops), executing workflows via pyflyte run -P and producing per-workflow namespaces and pods visible in kubectl. Claims: full multi‑tenant isolation plus resource sharing, enforced RBAC, declarative pod templates, and explicit GPU/CPU requests. Implementation detail: tasks are containerized Python functions, DAGs are visualized, and one project references a model-producing task from another for deployment, proving Flyte supports per-task Kubernetes namespaces and GPU assignment with task sharing across projects.
AI-written summary · about
chapters (13)
- 0:00MLOps and multi-tenancy intro
- 1:28Traditional ML challenges
- 3:56Why multi-tenancy matters
- 5:00Multi-tenant MLOps features
- 7:48Multi-tenancy analogy
- 9:14Orchestrators in MLOps
- 12:21Flyte for ML orchestration
- 14:29Flyte architecture
- 17:15Flyte projects and domains
- 19:02Resource isolation in Flyte
- 21:34Flyte multi-tenancy demo
- 28:36GPU allocation control
- 29:38Summary and Q&A
related talks
26mEmbracing Multi-Tenancy While Scaling MLOps
25mMLOps with Flyte
38mFlyte: Cloud Native Machine Learning & Data Processing Platform
39mFlyte: Production-Grade Data and Machine Learning Orchestration
28mRun Your Data and Machine Learning Workflows on Kubernetes with Flyte
01h 24mRobust and End-to-End Cloud Native Machine Learning & Data Processing Platform
23mManage Multi-tenant ML Workloads Using Istio
20mML Ops design patterns with Kubeflow Pipelines