Skip to content

Snowflake MLOps Reference

An Eraneos reference monorepo showing what end-to-end MLOps on Snowflake can look like, and how a broad range of Snowflake-native MLOps primitives fit together in one coherent example.

Why this repo exists

Existing Snowflake ML material covers individual pieces of the workflow. This repo wires those pieces into a single worked example for a realistic, multi-team setup.

It does two things at once:

  • Demonstrates end-to-end MLOps on Snowflake in a multi-team layout. The reference uses a hub-spoke architecture: a central platform hub owns infrastructure and governance; project spokes own their feature views, training pipelines, and inference pipelines. Hub-spoke is one specific way to organize the example; the patterns generalize.
  • Showcases the breadth of Snowflake-native MLOps features and how they interact. Feature engineering with the Snowflake Feature Store, model lifecycle in the Snowflake ML Model Registry, training and batch inference orchestrated through Snowflake Tasks DAGs, native observability via TASK_HISTORY and Account Usage, and the supporting primitives (databases, schemas, warehouses, compute pools, roles, stages) that wire them together. Snowflake-native primitives are preferred over external tooling wherever the native option is viable.

Everything is showcased through a single end-to-end example, drawn from a logistics and parcel-delivery context: predicting capacity utilisation at Pick-Up / Drop-Off points (PUDO), the network of shops, lockers, and service points where parcels are dropped off and collected. All data for the example is synthetic and generated locally (see the mock_data/ component), so the repo runs end-to-end on a fresh checkout without any real upstream data source. See the use case for the business framing.

What you will find here

Use Case

The PUDO (Pick-Up / Drop-Off) capacity prediction business problem that drives the reference implementation.

Tutorials

Step-by-step newcomer path: bootstrap Snowflake, seed data, deploy feature stores, train a model, run inference, simulate daily cycles, and evaluate results.

Concepts

Hub-spoke architecture, Snowflake ML lifecycle stages, feature stores, model registries, task-graph orchestration, and environment promotion.

Guides

Practical how-to pages for local setup, first end-to-end runs, troubleshooting, and working with component Makefiles.

Reference

Concise command reference, component map, and glossary.

Who this is for

  • Platform and ML engineers evaluating Snowflake as an MLOps platform.
  • Data scientists who want to understand how models move from notebooks to production task graphs.
  • Engineering leads designing multi-project ML architectures on Snowflake.

Quick start

If you already have a Snowflake account and the basic tooling installed, jump straight to the first tutorial.

If you are new to the repository, start with Start Here.