The MLflow Tracking component is an API and UI for logging parameters, code versions, metrics, and output files when running your machine learning code and for later visualizing the results. MLflow Tracking lets you log and query experiments using Python, REST, R API, and Java API APIs.
Table of Contents
- Where Runs Are Recorded
- Logging Data to Runs
- Organizing Runs in Experiments
- Tracking UI
- Querying Runs Programmatically
- MLflow Tracking Servers
MLflow Tracking is organized around the concept of runs, which are executions of some piece of data science code. Each run records the following information:
- Code Version
- Git commit hash used for the run, if it was run from an MLflow Project.
- Start & End Time
- Start and end time of the run
- Name of the file to launch the run, or the project name and entry point for the run if run from an MLflow Project.
- Key-value input parameters of your choice. Both keys and values are strings.
- Key-value metrics where the value is numeric. Each metric can be updated throughout the course of the run (for example, to track how your model’s loss function is converging), and MLflow records and lets you visualize the metric’s full history.
- Output files in any format. For example, you can record images (for example, PNGs), models (for example, a pickled scikit-learn model), or even data files (for example, a Parquet file) as artifacts.
You can record runs using MLflow Python, R, Java, and REST APIs from anywhere you run your code. For example, you can record them in a standalone program, on a remote cloud machine, or in an interactive notebook. If you record runs in an MLflow Project, MLflow remembers the project URI and source version.
You can optionally organize runs into experiments, which group together runs for a
specific task. You can create an experiment using the
mlflow experiments CLI, with
mlflow.create_experiment(), or using the corresponding REST parameters. The MLflow API and
UI let you create and search for experiments.
Once your runs have been recorded, you can query them using the Adding Tags to Runs or the MLflow API.
MLflow runs can be recorded to local files, to a SQLAlchemy compatible database, or remotely
to a tracking server.
By default, the MLflow Python API logs runs locally to files in an
mlruns directory wherever you
ran your program. You can then run
mlflow ui to see the logged runs.
To log runs remotely, set the
MLFLOW_TRACKING_URI environment variable to a tracking server’s URI or
There are different kinds of remote tracking URIs:
- Local file path (specified as
file:/my/local/dir), where data is just directly stored locally.
- Database encoded as a connection string (specified as
- HTTP server (specified as
https://my-server:5000), which is a server hosting an MLFlow tracking server.
- Databricks workspace (specified as
databricks://<profileName>, a Databricks CLI profile.
You can log data to runs using the MLflow Python, R, Java, or REST API. This section shows the Python API.
mlflow.set_tracking_uri() connects to a tracking URI. You can also set the
MLFLOW_TRACKING_URI environment variable to have MLflow find a URI from there. In both cases,
the URI can either be a HTTP/HTTPS URI for a remote server, a database connection string, or a
local path to log data to a directory. The URI defaults to
mlflow.tracking.get_tracking_uri() returns the current tracking URI.
mlflow.create_experiment() creates a new experiment and returns its ID. Runs can be
launched under the experiment by passing the experiment ID to
mlflow.set_experiment() sets an experiment as active. If the experiment does not exist,
creates a new experiment. If you do not specify an experiment in
runs are launched under this experiment.
mlflow.start_run() returns the currently active run (if one exists), or starts a new run
and returns a
mlflow.ActiveRun object usable as a context manager for the
current run. You do not need to call
start_run explicitly: calling one of the logging functions
with no active run automatically starts a new one.
mlflow.end_run() ends the currently active run, if any, taking an optional run status.
mlflow.log_param() logs a single key-value param in the currently active run. The key and
value are both strings. Use
mlflow.log_params() to log multiple params at once.
mlflow.log_metric() logs a single key-value metric. The value must always be a number.
MLflow remembers the history of values for each metric. Use
mlflow.log_metrics() to log
multiple metrics at once.
mlflow.set_tag() sets a single key-value tag in the currently active run. The key and
value are both strings. Use
mlflow.set_tags() to set multiple tags at once.
mlflow.log_artifact() logs a local file as an artifact, optionally taking an
artifact_path to place it in within the run’s artifact URI. Run artifacts can be organized into
directories, so you can place the artifact in a directory this way.
mlflow.log_artifacts() logs all the files in a given directory as artifacts, again taking
mlflow.get_artifact_uri() returns the URI that artifacts from the current run should be
Sometimes you want to launch multiple MLflow runs in the same program: for example, maybe you are
performing a hyperparameter search locally or your experiments are just very fast to run. This is
easy to do because the
ActiveRun object returned by
mlflow.start_run() is a Python
context manager. You can “scope” each run to
just one block of code as follows:
with mlflow.start_run(): mlflow.log_param("x", 1) mlflow.log_metric("y", 2) ...
The run remains open throughout the
with statement, and is automatically closed when the
statement exits, even if it exits due to an exception.
MLflow allows you to group runs under experiments, which can be useful for comparing runs intended
to tackle a particular task. You can create experiments using the Command-Line Interface (
mlflow experiments) or
mlflow.create_experiment() Python API. You can pass the experiment ID for a individual run
using the CLI (for example,
mlflow run ... --experiment-id [ID]) or the
# Prints "created an experiment with ID <id> mlflow experiments create fraud-detection # Set the ID via environment variables export MLFLOW_EXPERIMENT_ID=<id>
# Launch a run. The experiment ID is inferred from the MLFLOW_EXPERIMENT_ID environment # variable, or from the --experiment-id parameter passed to the MLflow CLI (the latter # taking precedence) with mlflow.start_run(): mlflow.log_param("a", 1) mlflow.log_metric("b", 2)
MLflow provides a more detailed Tracking Service API for managing experiments and runs directly,
which is available through client SDK in the
This makes it possible to query data about past runs, log additional information about them, create experiments,
add tags to a run, and more.
from mlflow.tracking import MlflowClient client = MlflowClient() experiments = client.list_experiments() # returns a list of mlflow.entities.Experiment run = client.create_run(experiments.experiment_id) # returns mlflow.entities.Run client.log_param(run.info.run_uuid, "hello", "world") client.set_terminated(run.info.run_uuid)
The Tracking UI lets you visualize, search and compare runs, as well as download run artifacts or
metadata for analysis in other tools. If you have been logging runs to a local
mlflow ui in the directory above it, and it loads the corresponding runs.
Alternatively, the MLflow tracking server serves the same UI and enables remote storage of run artifacts.
The UI contains the following key features:
- Experiment-based run listing and comparison
- Searching for runs by parameter or metric value
- Visualizing run metrics
- Downloading run results
All of the functions in the Tracking UI can be accessed programmatically. This makes it easy to do several common tasks:
- Query and compare runs using any data analysis tool of your choice, for example, pandas.
- Determine the artifact URI for a run to feed some of its artifacts into a new run when executing a workflow. For an example of querying runs and constructing a multistep workflow, see the MLflow Multistep Workflow Example project.
- Load artifacts from past runs as MLflow Models. For an example of training, exporting, and loading a model, and predicting using the model, see the MLFlow TensorFlow example.
- Run automated parameter search algorithms, where you query the metrics from various runs to submit new ones. For an example of running automated parameter search algorithms, see the MLflow Hyperparameter Tuning Example project.
You run an MLflow tracking server using
mlflow server. An example configuration for a server is:
mlflow server \ --backend-store-uri /mnt/persistent-disk \ --default-artifact-root s3://my-mlflow-bucket/ \ --host 0.0.0.0
An MLflow tracking server has two components for storage: a backend store and an artifact store.
The backend store is where MLflow Tracking Server stores experiment and run metadata as well as params, metrics, and tags for runs. MLflow supports two types of backend stores: file store and database-backed store.
--backend-store-uri to configure type of backend store. This can be a local path file
store specified as
file:/path_to_store, or a SQL connection string
for a database-backed store. For the latter, the argument must be a SQL connection string
database types are
--backend-store-uri is set to the local
./mlruns directory (the same as when
mlflow run locally), but when running a server, make sure that this points to a
persistent (that is, non-ephemeral) file system location.
For backwards compatibility,
--file-store is an alias for this option.
The artifact store is a location suitable for large data (such as an S3 bucket or shared NFS
file system) and is where clients log their artifact output (for example, models).
artifact_location is a property recorded on
default location to store artifacts for all runs in this experiment. Additional,
is a property on
mlflow.entities.RunInfo to indicate location where all artifacts for
this run are stored.
--default-artifact-root (defaults to local
./mlruns directory) to configure default
location to server’s artifact store. This will be used as artifact location for newly-created
experiments that do not specify one. Once you create an experiment,
is no longer relevant to that experiment.
To allow the server and clients to access the artifact location, you should configure your cloud
provider credentials as normal. For example, for S3, you can set the
AWS_SECRET_ACCESS_KEY environment variables, use an IAM role, or configure a default
See Set up AWS Credentials and Region for Development for more info.
If you do not specify a
--default-artifact-root or an artifact URI when creating the experiment
mlflow experiments create --artifact-location s3://<my-bucket>), the artifact root
is a path inside the file store. Typically this is not an appropriate location, as the client and
server probably refer to different physical locations (that is, the same path on different disks).
In addition to local file paths, MLflow supports the following storage systems as artifact stores: Amazon S3, Azure Blob Storage, Google Cloud Storage, SFTP server, and NFS.
To store artifacts in S3, specify a URI of the form
s3://<bucket>/<path>. MLflow obtains
credentials to access S3 from your machine’s IAM role, a profile in
the environment variables
AWS_SECRET_ACCESS_KEY depending on which of
these are available. For more information on how to set credentials, see
Set up AWS Credentials and Region for Development.
To store artifacts in a custom endpoint, set the
MLFLOW_S3_ENDPOINT_URL to your endpoint’s URL.
For example, if you have a Minio server at 188.8.131.52 on port 9000:
To store artifacts in Azure Blob Storage, specify a URI of the form
MLflow expects Azure Storage access credentials in the
AZURE_STORAGE_ACCESS_KEY environment variables (preferring
a connection string if one is set), so you must set one of these variables on both your client
application and your MLflow tracking server. Finally, you must run
pip install azure-storage
separately (on both your client and the server) to access Azure Blob Storage; MLflow does not declare
a dependency on this package by default.
To store artifacts in Google Cloud Storage, specify a URI of the form
You should configure credentials for accessing the GCS container on the client and server as described
in the GCS documentation.
Finally, you must run
pip install google-cloud-storage (on both your client and the server)
to access Google Cloud Storage; MLflow does not declare a dependency on this package by default.
Specify a URI of the form ftp://user@host/path/to/directory to store artifacts in a FTP server.
The URI may optionally include a password for logging into the server, e.g.
To store artifacts in an SFTP server, specify a URI of the form
You should configure the client to be able to log in to the SFTP server without a password over SSH (e.g. public key, identity file in ssh_config, etc.).
sftp://user:pass@host/ is supported for logging in. However, for safety reasons this is not recommended.
When using this store,
pysftp must be installed on both the server and the client. Run
pip install pysftp to install the required package.
--host option exposes the service on all interfaces. If running a server in production, we
would recommend not exposing the built-in server broadly (as it is unauthenticated and unencrypted),
and instead putting it behind a reverse proxy like NGINX or Apache httpd, or connecting over VPN.
Additionally, you should ensure that the
--backend-store-uri (which defaults to the
./mlruns directory) points to a persistent (non-ephemeral) disk or database connection.
To log to a tracking server, set the
MLFLOW_TRACKING_URI environment variable to the server’s URI,
along with its scheme and port (for example,
http://10.0.0.1:5000) or call
import mlflow with mlflow.start_run(): mlflow.log_param("a", 1) mlflow.log_metric("b", 2)