MetaOps: Metadata Operations For End-To-End Data & Machine Learning Platforms
summary
Argues that artifact stores and ad hoc schemas fail at production scale where thousands of models and many-to-many deployments exist, shows Seldon Core experiences with models deployed across environments and pipelines replicated 40 times across 12 teams, and points to an implemented interoperability stack: the V2 protocol (NVIDIA, Triton, KServe, Seldon) plus a proposed Open Inference Schema. Claims artifact stores don’t capture runtime service metadata, recommends serving systems expose model metadata for lineage and audit, and reports the protocol is already adopted in production but full ecosystem standards remain unfinished.
AI-written summary · about
chapters (16)
- 0:00Start
- 2:48Converging Into Cannonical Stack(s)
- 4:23Motivation for End To End Lineage
- 6:15Centralised to Decentralised Heterogeneus
- 8:43Challenges of Metadata Mgmt at Web Scale
- 18:04The Limits of Model Artifact Stores
- 20:16Deployed Model Relationships
- 20:51Production MLOps Interoperability
- 22:41Bridging Training & Inference Data
- 24:05From Model-Centric to Data-Centric
- 25:30Inference Data Heterogeneity
- 26:16The Open Inference Protocol (aka V2)
- 27:37End to End Metadata Interoperability
- 28:58Extrapolating to Complex Pipelines
- 29:45Leveraging the Open Inference Schema
- 31:04Call to action
related talks
29mMetadata Systems for End-to-End Data & Machine Learning
10mData Management and Exchange between a Meta-Orchestration Platform and Data Spaces
46mJourney of metadata
52mA MLOps Platform for All Data Personas
28mthe semantic layer your AI needs
37mMetadata Automation in the Data Mesh
45mTechnical Challenges in Open Metadata
33mDiscover How to Create Your Own Metadata-Driven ML Platform from Scratch