Configuration of Model Management
In order to use the Model Management feature, a running reachable MLflow instance is required. Also, Group Administrator must configure Group Setting with MLflow-related information.
According to a binding MLflow instance, Group Administrators could choose either one of ways:
installed MLflow server by PrimeHub Apps
integrate with an external MLflow server
There are some key considerations when setting up a MLflow instance to make Model Management and Deployment working together, especially deploy a model from the
MLflow Model Registry.
Installed MLflow server by PrimeHub Apps
The easiest way to set up a MLflow instance is to install a MLflow instance from the PrimeHub Apps.
Requirements (under same group-context):
- enable the Group Volume for saving the artifacts.
- install a MLflow App instance
In the default settings, MLflow App uses the path
DEFAULT_ARTIFACT_ROOT env where it stores artifacts.
When deploying a registered model from MLflow Model Registry according to its modelUri
models:/<model-name>/<version-name>, PrimeHub needs to copy artifacts from
$(PRIMEHUB_APP_ROOT)/mlruns. Which requires this group with enabled Group Volume and a running installed MLflow App.
If want to share
MLflow Model Registryamong other groups which have installed MLflow server by Apps either. Please prepare and configure proper Artifact Stores. It is very common to use Amazon S3 and S3-compatible storage to save artifacts.
Integrate with the external MLflow server
Alternatively, it is possible to bind a self-hosted MLflow instance outside of PrimeHub.
Since you have a running external MLflow server, it should be with an prepared artifact store.
The only one consideration is to make sure that
DEFAULT_ARTIFACT_ROOT, the path to the artifact store, is available to your MLflow client.
Same as the use-case of sharing
MLflow Model Registryto other groups, it is a good idea to use Amazon S3 and S3-compatible storage.
So far, you supposedly already have a running MLFlow instance, either installed by PrimeHub Apps or externally-hosted. We have to bind the service by configuring Group Setting.
If using installed MLflow server by Apps, see Group Setting - MLflow section.
Otherwise, in MLflow tab of Group Setting, configure these two settings with the information of externally-hosted MLflow:
MLflow Tracking URIwhere the MLflow instance serves tracking clients. PrimeHub uses it as a corresponding
MLFLOW_TRACKING_URIenvironment variable in system. You could use this env in Notebooks and Jobs.
MLflow UI URIis a URL to the MLflow web server.
Two groups of environmental variables ,
Tracking Environment Variables and
Artifact Store Environment Variables:
Tracking Environment Variables
Usually, it is for externally-hosted MLflow server.
It is used to set the authentication configuration to your tracking server.
MLFLOW_TRACKING_PASSWORD- username and password to use with HTTP Basic authentication. To use Basic authentication, you must set both environment variables.
MLFLOW_TRACKING_TOKEN- token to use with HTTP Bearer authentication. Basic authentication takes precedence if set.
Please refer to MLflow document to find more information.
Artifact Store Environment Variables
Usually, it is for externally-hosted MLflow server or for external artifact stores.
It is used to instruct your client how to connect to the artifact storage. We take
S3-compatible storage as an example:
AWS_SECRET_ACCESS_KEYare used by the S3 client.
MLFLOW_S3_ENDPOINT_URLis used to tell s3 client to connect your own S3-compatible storage rather than the AWS S3 server.
Please refer to MLflow Artifact Storage to find more information.