# Iceberg, MCP, and MLOps: Bridging the Gaps for Enterprise

Caleb Baechtold, Snowflake & Hamza Tahir, ZenML & Simba Khadder, Featureform | MLOps Mini Summit 2025 | Episode 11 | 1:02:19
Hosted by Ben Epstein

Source: https://www.youtube.com/watch?v=gicCH6FC-a4
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/iceberg-mcp-and-mlops-bridging-the-gaps-for-enterprise
Published: 2025-05-05
Tags: governance, mcp, orchestration, platform-teams

## TL;DR
- Caleb Baechtold argues that enterprise MLOps depends as much on governance, access control, lineage, and organizational processes as it does on technical infrastructure.
- Hamza Tahir presents ZenML as a workflow framework that lets platform teams govern infrastructure while allowing data scientists to use different orchestration, compute, and tracking systems.
- Simba Khadder argues that feature platforms should expose governed structured data to agents through MCP, while Iceberg provides a shared table format across data systems.

## Summary
This Mini Summit connects three parts of enterprise MLOps. Caleb Baechtold describes Snowflake's approach to bringing feature work, model development, deployment, monitoring, access control, lineage, and cost management into one governed environment. Hamza Tahir focuses on the organizational problem created by varied data, infrastructure, teams, and legacy tools. ZenML provides an abstraction for defining workflows and infrastructure stacks while preserving room for different providers and orchestration systems. Simba Khadder then extends the feature platform toward agentic systems. He argues that agents need governed access to structured, user-specific data, not only retrieved documents. Featureform uses Iceberg-backed data pipelines and an MCP interface to make those features discoverable and authorized. The discussion closes with Iceberg's adoption, catalog design, and the tradeoffs between Iceberg and Delta. The speakers are candid about the operational work that remains around permissions, catalogs, low-latency serving, and Python tooling.

## Key ideas
### Enterprise MLOps is partly an organizational governance problem
[01:50](https://www.youtube.com/watch?v=gicCH6FC-a4&t=110s)
Caleb Baechtold says MLOps is only about half a technology problem. Enterprises also have to manage governance, compliance, auditability, and processes across many teams. When data and models move through separate services, those responsibilities become harder to track. Snowflake's approach is to extend existing data governance to machine learning, including role-based access control for feature definitions, model objects, deployments, and inference results. The goal is a consistent view from source data through model outputs, with lineage and auditability across the lifecycle.

### Snowflake separates experimentation from production without losing control
[11:14](https://www.youtube.com/watch?v=gicCH6FC-a4&t=674s)
Snowflake supports sandbox, test, and production environments through separate accounts or through database and schema boundaries within one account. Data scientists can experiment with flexible permissions in a sandbox, while CI/CD promotes approved objects into production schemas. The same separation applies to features, notebooks, registries, inference results, monitors, and compute. Caleb also describes cost controls, including resource monitors, budgets, and different pricing tiers for development and production environments.

### ZenML treats workflow infrastructure as a governed set of interchangeable stacks
[25:35](https://www.youtube.com/watch?v=gicCH6FC-a4&t=1535s)
Hamza Tahir describes ZenML as a framework for standardizing how workflows run on infrastructure. A stack can describe an on-premises Kubernetes cluster, SageMaker, an experiment tracker, a model registry, or other components. Platform engineers provision and govern these stacks, while data scientists write workflows that consume them. ZenML handles tasks such as building images, compiling environments, deploying workflows, and creating experiments in underlying tools. The framework supports different stacks so teams can work across clouds, on-premises systems, and varied data environments.

### Enterprise platform teams need a governance envelope rather than a single enforced stack
[21:57](https://www.youtube.com/watch?v=gicCH6FC-a4&t=1317s)
Tahir says enterprises face heterogeneous compute, distributed data, mixed team expertise, legacy tools, compliance requirements, and concerns about vendor lock-in. Enforcing one stack can fail because a real-time edge workload and a fraud model have different needs. Allowing every team to choose everything creates its own problems. His preferred approach is a governance envelope that controls access to data, compute, and tools while leaving teams enough room to work. ZenML applies this through permissions, service connectors, shared stacks, and expiring access.

### Feature platforms can turn local feature experiments into production pipelines
[40:40](https://www.youtube.com/watch?v=gicCH6FC-a4&t=2440s)
Simba Khadder explains Featureform's approach to defining feature transformations as Python, SQL, or data-frame code. A data scientist can experiment locally in a notebook, then place feature definitions in a repository and deploy them through CI/CD. Featureform runs on existing infrastructure such as Snowflake, Dynamo, and other data systems. Khadder says Featureform uses Iceberg tables and handles materialization so updated feature data can move into serving systems. The aim is to reduce the gap between local experimentation and production feature pipelines.

### Agents need structured, personalized data alongside retrieved documents
[47:00](https://www.youtube.com/watch?v=gicCH6FC-a4&t=2820s)
Khadder uses a delayed DoorDash order to explain the limitation of a generic retrieval workflow. A support bot that searches a help center may describe common causes of delays, while a useful agent should inspect the user's order, restaurant status, delivery location, and estimated arrival time. Those details already exist as structured data used by operational models. Featureform's MCP interface lets an agent discover relevant feature views and use authorized attributes for the current intent.

### Production MCP needs catalogs, permissions, serving, and audit controls
[49:56](https://www.youtube.com/watch?v=gicCH6FC-a4&t=2996s)
Khadder says an MCP server alone does not solve the enterprise problem. An agent needs a semantic catalog that explains what attributes mean, low-latency serving, user propagation, real-time updates, audit logs, and access controls. Featureform's design prevents an agent from querying data it should not see. He also describes MCP Engine, an open-source project for building stateless MCP servers with package management and AWS Lambda support.

### Iceberg's adoption raises a separate catalog and tooling problem
[55:36](https://www.youtube.com/watch?v=gicCH6FC-a4&t=3336s)
The panel describes Iceberg as an open table format that lets storage live in object stores while different compute systems operate over it. Caleb says Snowflake aims for users to see the same experience whether they use native Snowflake tables or Iceberg tables, while Polaris addresses open catalog interoperability. Simba attributes Iceberg's adoption to both technical features and its distance from a single vendor. The speakers also discuss gaps in Python support, predicate pushdown, and DuckDB integration.

## Notable quotes
- Caleb Baechtold: "MLOps is only maybe half of the technology problem and a lot of it boils down to organizational governance processes and things like that." (02:11)
- Hamza Tahir: "The goal is that ZenML handles all the rest: building the Docker images, compiling the environments, deploying these things, creating the experiments in these underlying tools." (28:21)
- Simba Khadder: "Featureform's MCP interface enables agents to discover and use relevant and authorized structured data to solve for an intent." (48:49)
- Simba Khadder: "We don't trust the agent to not look at data it shouldn't look at." (50:52)
- Caleb Baechtold: "Iceberg is becoming the standard." (56:15)

## Tools & references mentioned
- Snowflake
- ZenML
- Featureform
- MLOps Community
- MLflow
- Kubernetes
- SageMaker
- Airflow
- Terraform
- Snowflake ML
- Iceberg
- MCP Engine
- Model Context Protocol
- Polaris
- AWS Lambda
- DoorDash
- Delta

## Who should watch
- You are building an enterprise ML platform and need to balance team flexibility with access control, compliance, cost tracking, and production governance.
- Your teams use several clouds, orchestration engines, registries, and data systems, and you want a common workflow layer without forcing everyone onto one stack.
- You are extending a feature platform toward agents and need structured data access with semantic discovery, user permissions, low-latency serving, and audit logs.

## Editor's note

Hamza Tahir says enterprises need a governance envelope because heterogeneous compute, distributed data, legacy tools, and compliance make one enforced stack impractical. ZenML lets platform engineers provision and govern interchangeable stacks while data scientists write workflows against them, so teams can use different infrastructure without giving up centrally managed access to data, compute, and tools.

Written by the MLOps Talks editors (the ZenML team), not by the speaker.

## Related talks

- [Enterprise Security and Governance MLOps](https://mlopstalks.com/talks/enterprise-security-and-governance-mlops) (Diego Oppenheimer, Algorithmia, 53:24)
- [Maturing Machine Learning in Enterprise](https://mlopstalks.com/talks/maturing-machine-learning-in-enterprise) (Kyle Gallatin, Etsy, 47:09)
- [MLOps - The Blind Men and the Elephant](https://mlopstalks.com/talks/mlops-the-blind-men-and-the-elephant) (Saurav Chakravorty, Brillo, 55:02)
- [Creating MLOps Standards](https://mlopstalks.com/talks/creating-mlops-standards) (Alex Chung, Social Good Technologies & Srivathsan Canchi, Intuit, 47:45)
- [Designing ML Infra for ML & LLM Use Cases](https://mlopstalks.com/talks/designing-ml-infra-for-ml-llm-use-cases) (Amritha Arun Babu, Klaviyo & Abhik Choudhury, IBM, 1:00:18)