Favorita: MLflow Tracking Utility

The actual tracking helper: a small experiment-tracking utility that every model script imports to log params, metrics, and artifacts to MLflow the same way. One module, consistent run history across XGBoost, random forest, and time-series models.

Primary Outcome

A single shared utility that gives every model consistent MLflow logging without per-script boilerplate.

Solution

A single shared tracking utility that standardizes what every model logs, so runs are consistent, comparable, and reproducible.

Deliverables

  • experiment_tracking.py — the shared MLflow logging utility

  • A consistent run-logging interface used by every model script

  • Param, metric, and artifact logging in one place

  • A test that documents the expected tracking behavior

Strategic Context

Tracking only helps if it's consistent. Five scripts each logging slightly different things produce runs you can't actually compare. A shared utility enforces the consistency that makes experiment history trustworthy — the unglamorous detail that decides whether tracking is useful or noise.

Technical Architecture

Tracking only helps if it's consistent. Five scripts each logging slightly different things produce runs you can't actually compare. A shared utility enforces the consistency that makes experiment history trustworthy — the unglamorous detail that decides whether tracking is useful or noise.

Problem Statement

When each model script logs to MLflow its own way, runs aren't comparable and tracking becomes noise instead of signal.

Links

What's Included

The vertex/utils/experiment_tracking.py module and its test in vertex/tests, providing the shared MLflow logging interface the model scripts call.

FAQs

Is this tied to these models?

No — it's a thin wrapper over MLflow. Any Python model can import it and log consistently.

Local or hosted MLflow?

Either. The utility points at whatever tracking backend you configure.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4

Primary Outcome

A single shared utility that gives every model consistent MLflow logging without per-script boilerplate.

Problem Statement

When each model script logs to MLflow its own way, runs aren't comparable and tracking becomes noise instead of signal.

Solution

A single shared tracking utility that standardizes what every model logs, so runs are consistent, comparable, and reproducible.

Links

Deliverables

  • experiment_tracking.py — the shared MLflow logging utility

  • A consistent run-logging interface used by every model script

  • Param, metric, and artifact logging in one place

  • A test that documents the expected tracking behavior

What's Included

The vertex/utils/experiment_tracking.py module and its test in vertex/tests, providing the shared MLflow logging interface the model scripts call.

Strategic Context

Tracking only helps if it's consistent. Five scripts each logging slightly different things produce runs you can't actually compare. A shared utility enforces the consistency that makes experiment history trustworthy — the unglamorous detail that decides whether tracking is useful or noise.

FAQs

Is this tied to these models?

No — it's a thin wrapper over MLflow. Any Python model can import it and log consistently.

Local or hosted MLflow?

Either. The utility points at whatever tracking backend you configure.

Technical Architecture

Tracking only helps if it's consistent. Five scripts each logging slightly different things produce runs you can't actually compare. A shared utility enforces the consistency that makes experiment history trustworthy — the unglamorous detail that decides whether tracking is useful or noise.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4