MLflow UI Telemetry

Overview

Telemetry in MLflow's UI is based on a SharedWorker, which is a type of web worker that can be accessed by multiple tabs. This allows us to effectively consolidate and batch logs.

Architecture

Client

TelemetryClient.ts:

• Client API for logging events. A singleton is exported from this file, and is imported / used in app.tsx (the top-level MLflow frontend component).
• We hook into the built-in DesignSystemEventProvider, which generates view and click events for every interactive component.
• When logEvent is called, the client forwards the log to the SharedWorker via postMessage

SharedWorker

TelemetryLogger.worker.ts:

• Main worker class that handles communication with the /ui-telemetry server endpoint.
• Please keep external dependencies here minimal, as it is bundled separately from the main app and we'd ideally keep the generated bundle relatively light (see the telemetry-worker entrypoint in craco.config.js)

LogQueue.ts:

• Simple class that batches logs and uploads them to the server every 15s.