How to See Where MLflow Stores Runs and Artifacts

An MLflow tracking server keeps your work in two places: a database for experiments, runs, parameters, metrics and tags, and a folder for the files your runs log, such as charts and models. With the default settings, both sit in the folder you started the server from. This guide finds them, follows one run from the UI to its files and database rows, and reads the database safely, with real output from MLflow 3.18.0.

The Layout

A tracking server started with the default settings keeps everything in the folder you start it from, in two places:

The server's folder

mlflow-server/
├── mlflow.db                  experiments, runs, parameters, metrics and tags
└── mlartifacts/               the files that runs log
    └── <experiment ID>/
        └── <run ID>/
            └── artifacts/
                └── <the path you logged>

Place	What it is
mlflow.db	The backend store: a SQLite database with a row for every experiment, run, parameter, metric and tag
mlartifacts/	The artifact store: an ordinary folder with a subfolder for each experiment, then one for each run

The steps below find both, follow one run into each, and look inside the database.

Prerequisites

• A tracking server started with the default storage settings, from a folder of its own. Here the folder is mlflow-server: see How to Start an MLflow Tracking Server on Your Computer.
• A few runs that logged a file. This guide uses the four runs from Part 1 of the MLflow series, random forests that each logged a chart.
• Python 3.12 or newer, for the sqlite3 command line used in Step 4.

Step 1: Find the Two Places in the Server's Folder

Open a terminal in the server's folder, the one you ran mlflow server from, and list what's in it:

macOS / Linux

Terminal (in mlflow-server)

ls -F

Windows (PowerShell)

Terminal (in mlflow-server)

Get-ChildItem -Name

On macOS:

Output

mlartifacts/    mlflow.db       pyproject.toml  uv.lock

• mlflow.db is the database, which MLflow's documentation calls the backend store: "SQLite is the default backend store. When you start MLflow without specifying a backend, it automatically creates and uses sqlite:///mlflow.db in the current directory."
• mlartifacts/ holds the files, the artifact store: "By default, the tracking server stores artifacts in its local filesystem under ./mlartifacts directory."
• pyproject.toml and uv.lock belong to uv, which installed MLflow in this folder.

Because both paths are relative to where the server starts, starting it from another folder gives you a new, empty database there.

Step 2: See the Locations in the UI

The UI doesn't show folder paths. It shows addresses that start with mlflow-artifacts:. Open an experiment and click the ⓘ icon next to Runs:

The experiment's Artifact Location is mlflow-artifacts:/1, after its ID. A run's Artifacts tab shows each file's full address. For the run smiling-moose-157:

mlflow-artifacts:/ means "the server's artifact store". Your code never writes into mlartifacts itself: it sends files to the server, and the server stores them. In the documentation's words, "The tracking server resolves the uri mlflow-artifacts:/ in tracking request from the client to an otherwise explicit object store destination", here the local mlartifacts folder.

Step 3: Follow a Run to Its Files

List every file in the artifact store:

macOS / Linux

Terminal (in mlflow-server)

find mlartifacts -type f

Windows (PowerShell)

Terminal (in mlflow-server)

Get-ChildItem mlartifacts -Recurse -File | Resolve-Path -Relative

On macOS:

Output

mlartifacts/1/09c7c9012ca74eafaf16c8766226d6b7/artifacts/plots/predicted_vs_actual.png
mlartifacts/1/7cb465b94f2e426f9af5a12c55326cbe/artifacts/plots/predicted_vs_actual.png
mlartifacts/1/8dbe08fc91434c5b9b22456dcd740ebf/artifacts/plots/predicted_vs_actual.png
mlartifacts/1/ec7ccb942d66407490b3aee0cb4d9b62/artifacts/plots/predicted_vs_actual.png

PowerShell prints the same paths with ./ in front on macOS; on Windows they start with .\ and use backslashes, as in Microsoft's Resolve-Path examples.

Every file follows the same pattern:

Artifact path

mlartifacts/<experiment ID>/<run ID>/artifacts/<the path you logged>

The third line is the file from Step 2: experiment 1, run 8dbe08fc91434c5b9b22456dcd740ebf (smiling-moose-157), and plots/predicted_vs_actual.png, the path the training code gave mlflow.log_figure(). Replace mlflow-artifacts: with mlartifacts and the UI's address becomes the file's location on disk.

Step 4: Look Inside mlflow.db

Python 3.12 added a command line to its built-in sqlite3 module: python -m sqlite3 <database> "<query>" runs one query and prints each row. It needs nothing extra to install, and the commands below are the same in every shell.

Start with the experiments:

Terminal (in mlflow-server)

uv run python -m sqlite3 mlflow.db "SELECT experiment_id, name, artifact_location FROM experiments;"

Output

(0, 'Default', 'mlflow-artifacts:/0')
(1, 'house-price-prediction', 'mlflow-artifacts:/1')

Default is experiment 0, which MLflow creates when it sets up the database. Next, the runs in experiment 1:

Terminal (in mlflow-server)

uv run python -m sqlite3 mlflow.db "SELECT name, status, artifact_uri FROM runs WHERE experiment_id = 1;"

Output

('smiling-moose-157', 'FINISHED', 'mlflow-artifacts:/1/8dbe08fc91434c5b9b22456dcd740ebf/artifacts')
('50-trees-depth-8', 'FINISHED', 'mlflow-artifacts:/1/ec7ccb942d66407490b3aee0cb4d9b62/artifacts')
('200-trees-depth-16', 'FINISHED', 'mlflow-artifacts:/1/09c7c9012ca74eafaf16c8766226d6b7/artifacts')
('300-trees', 'FINISHED', 'mlflow-artifacts:/1/7cb465b94f2e426f9af5a12c55326cbe/artifacts')

Each run's artifact_uri is its folder in the artifact store. Now one run's parameters, metrics and tags, here 300-trees. Use a run ID of your own, from the query above or the run page:

Terminal (in mlflow-server)

uv run python -m sqlite3 mlflow.db "SELECT key, value FROM params WHERE run_uuid = '7cb465b94f2e426f9af5a12c55326cbe';"
uv run python -m sqlite3 mlflow.db "SELECT key, value, step FROM metrics WHERE run_uuid = '7cb465b94f2e426f9af5a12c55326cbe';"
uv run python -m sqlite3 mlflow.db "SELECT key, value FROM tags WHERE run_uuid = '7cb465b94f2e426f9af5a12c55326cbe';"

Output

('n_estimators', '300')
('max_depth', 'None')
('test_size', '0.2')
('random_state', '42')
('mae', 0.3267739603762923, 0)
('rmse', 0.503821166676639, 0)
('r2', 0.8062927130336361, 0)
('mlflow.user', 'you')
('mlflow.source.name', '/Users/you/source/dewiride-house-price-prediction/.venv/bin/train')
('mlflow.source.type', 'LOCAL')
('mlflow.runName', '300-trees')
('model_type', 'random_forest')
('dataset', 'california_housing')

• Parameters are text: '300', even 'None'. That's why parameters compare as text in searches.
• Metrics are numbers at full precision, each with a step, 0 when none was given. The UI rounds them for display.
• Tags include MLflow's system tags, such as mlflow.user and mlflow.source.name, next to the run's own model_type and dataset.

The metrics table keeps every value ever logged. A second table, latest_metrics, keeps one row per run and metric name:

Terminal (in mlflow-server)

uv run python -m sqlite3 mlflow.db "SELECT key, value, step FROM latest_metrics WHERE run_uuid = '7cb465b94f2e426f9af5a12c55326cbe';"

Output

('mae', 0.3267739603762923, 0)
('rmse', 0.503821166676639, 0)
('r2', 0.8062927130336361, 0)

The same three values here, because each metric was logged once. For a metric logged at many steps, such as a training loss, metrics holds a row per step, and latest_metrics keeps the one with the highest step.

What the Other Tables Hold

The database has many more tables than these:

Terminal (in mlflow-server)

uv run python -m sqlite3 mlflow.db "SELECT count(*) FROM sqlite_master WHERE type = 'table';"

Output

(71,)

MLflow creates all 71 when it sets up a new database, whether you use them or not. The ones for experiment tracking:

Table	What it holds
experiments	One row per experiment: its ID, name and artifact location
experiment_tags	Tags on experiments
runs	One row per run: its ID, name, status, times and artifact URI
params	Run parameters, as text
metrics	Every logged metric value, with its step and time
latest_metrics	The latest value of each metric, one row per run and metric name
tags	Run tags, including MLflow's system tags
datasets, inputs, input_tags	Datasets logged as run inputs
logged_models and the logged_model_… tables	Models logged with MLflow 3, with their own parameters, metrics and tags
registered_models, model_versions and their …_tags and …_aliases tables	The Model Registry
alembic_version	The version of the database layout, which mlflow db upgrade uses

Most of the rest serve MLflow's GenAI features, such as traces (trace_info, spans) and evaluations (assessments, scorers). List every name with SELECT name FROM sqlite_master WHERE type = 'table' ORDER BY name;.

Read the database, never edit it

Every query on this page only reads. Changing rows by hand can leave the database inconsistent with the files in mlartifacts, or with what MLflow expects. Change runs through MLflow: its Python API, its command line or the UI.

Common Mistakes

Mistake	What happens	Do this instead
Starting the server from another folder	A new, empty mlflow.db and mlartifacts appear there, and your runs seem to vanish	Always start the server from the same folder
Logging with no tracking URI set	The run goes to an mlflow.db in your project's folder, not the server's	Set the tracking URI: How to Set the MLflow Tracking URI in Python
Looking for files in mlflow.db	The database only stores the files' location	Look in mlartifacts/<experiment ID>/<run ID>/artifacts
Editing mlflow.db by hand	The data can end up inconsistent	Change runs through MLflow

Troubleshooting

Symptom	Likely cause	Fix
mlartifacts doesn't exist	No run has logged a file yet	Log one; the folder appears with the first file
A run is in the UI but its files aren't in mlartifacts	The run was logged to another server or folder	Compare the run's artifact address with the server you started

FAQ

Is it safe to query mlflow.db while the server is running?

For reading, yes. Every query on this page ran while the server was running. Don't write to it.

Where do runs go when I don't run a server?

With no tracking URI set, MLflow 3.18 logs to an mlflow.db in the folder you run your code from, and puts files in an mlruns folder there: How to Set the MLflow Tracking URI in Python.

What happens to mlflow.db when I upgrade MLflow?

A newer MLflow can need a newer database layout. MLflow's documentation says "mlflow server will fail against a database with an out-of-date schema", and to run mlflow db upgrade after backing the database up.

Conclusion

The tracking server kept everything in the folder it started from: mlflow.db, a SQLite database with experiments, runs, parameters, metrics and tags, and mlartifacts, with each run's files under <experiment ID>/<run ID>/artifacts. The UI's mlflow-artifacts:/1/… addresses pointed at the same files, and python -m sqlite3 read the database without changing it.

Additional Resources

• Backend Stores — the default SQLite database and other databases.
• Tracking Server — proxied artifact access and mlflow-artifacts:/.
• sqlite3 command-line interface — Python's built-in SQLite shell.

Related

• How to Start an MLflow Tracking Server on Your Computer
• How to Create and Set an MLflow Experiment
• How to Log a Matplotlib Figure to MLflow
• MLflow Tutorial for Machine Learning in Python — Part 1: Install MLflow and Track Your First Experiment