23 posts tagged with "MLflow"

The tracking URI tells your code where MLflow records go. Get it wrong and nothing fails: MLflow quietly writes your runs somewhere else, and your UI stays empty. This guide shows the two ways to set it, which one wins, how to check it, and exactly what MLflow 3.18 does when you don't set it, with real output.

The MLflow UI is where your runs become useful: every setting, score and chart in one place, searchable and comparable. MLflow 3.18's UI serves two audiences, AI apps and classic model training, and opens in the AI view by default. This tour switches it to model training and walks through the pages you'll use most: experiments, the runs table, and a run's own page.

Since version 3.2.0, open-source MLflow collects anonymous usage data by default: which features you use, your MLflow and Python versions, your operating system. It doesn't collect your data, parameters or file names, but you may still want it off. One environment variable turns it off for your scripts, the tracking server and the UI. This guide sets it on macOS, Linux and Windows, and proves it worked.

The MLflow tracking server listens on port 5000 unless you tell it otherwise. That port is often taken: on a Mac, the AirPlay Receiver already holds it, and other programs like it too. One option moves MLflow to any free port. This guide changes the port, checks what's using a port, and keeps your code pointed at the right address.

The MLflow tracking server is the place your training runs go: a small web server that stores every run and shows them in the MLflow UI. You don't need a cloud account or a database server to run one; MLflow 3.18 brings everything, including a SQLite database. This guide starts a server on your own computer, explains every line it prints, checks that it's ready, and stops it safely.

MLflow installs like any Python package, but it brings a lot with it: a web server, a database layer and the most common data science libraries. Installing it into a uv project keeps all of that inside the project and records the exact versions. This guide installs MLflow 3.18.0, checks it, and explains which of MLflow's three packages you need.

uv run mlflow server --port 8080 stopped right after it began, with ERROR: [Errno 48] Address already in use and Running the mlflow server failed: The server process exited with code 3. Another MLflow server was already listening on that port. lsof showed which one, and the new server started cleanly on a free port.

An MLflow training run printed seven urllib3 retry warnings over about four minutes, then failed with MlflowException: API request to … failed with exception … Max retries exceeded … [Errno 61] Connection refused. Nothing was listening at the tracking URI. MLflow's client retries for minutes before it gives up, which makes a missing server look like a hang. Starting the server fixed it, and MLFLOW_HTTP_REQUEST_MAX_RETRIES makes the same mistake fail in seconds.

pytest passed all 12 tests of a project that logs to MLflow, then printed a warnings summary: SADeprecationWarning: The ``noload`` loader strategy is deprecated. It comes from MLflow 3.18.0's SQL tracking store meeting SQLAlchemy 2.1, and MLflow fixed it the day after 3.18.0 shipped. Until a release includes the fix, one filterwarnings line hides exactly this warning in your tests.

Pressing Ctrl+C to stop MLflow 3.18.0's tracking server printed Exception ignored on threading shutdown: and a KeyboardInterrupt traceback from Python's own threading.py. Nothing crashed: the server stopped cleanly, and the message comes from its background job runner being interrupted at exit. If you don't use MLflow's server jobs, MLFLOW_SERVER_ENABLE_JOB_EXECUTION=false makes it go away.