Deploying models with MLflow2Seldon
This service is running inside of a Kubernetes cluster with deployed Seldon-core. It synchronizes the MLFlow state to Seldon-core within the Kubernetes cluster by continuously fetching MLFlow-server registered models running as platform jobs via MLFlow Python SDK.
For instance, if a version of a MLFlow-registered model gets assigned to the staging/production stage, the corresponding model binary gets deployed from MLFlow into the K8s cluster as SeldonDeployment (exposing REST/GRPC APIs). If the stage assignment gets removed/updated - the corresponding SeldonDeployment is changed respectively.
Given that, all interaction with the service is done implicitly via the MLFlow server state. There is no need to execute particular commands/workloads directly against this service.
Here's the setup of MLflow and Seldon you need to have before running MLflow2Seldon:
- seldon-core-operator version is at least 1.5.0
kubectltool is authenticated to communicate with a Kubernetes cluster on which Seldon is deployed
To deploy a model with MLflow2Seldon through the CLI, run the following command:
$ make helm_deploy
This command will prompt you to enter some additional information - for example, what is the MLFlow URL, which Neu.ro cluster should be used, etc.
Set the following environment variables to use MLflow2Seldon:
M2S_MLFLOW_STORAGE_ROOT- artifact root path on the platform storage (example: storage:myproject/mlruns).
M2S_SELDON_NEURO_DEF_IMAGE- docker image stored on the platform registry which will be used to deploy the model (example: image:myproject/seldon:v1). You can also configure the service to use another platform image for deployment by tagging the respective registered model (not a model version) with the tag named after the value of the
M2S_MLFLOW_DEPLOY_IMG_TAGchart parameter. For example, with a tag named "deployment-image" and the value "image:myproject/seldon:v2".
M2S_SRC_NEURO_CLUSTER- Neu.ro cluster on which the deployment image, MLflow artifacts, and MLFlow itself are hosted (demo_cluster).
Direct use of a helm chart is possible, but it may be less convenient - all info required by the makefile should be passed as chart values.
When you're done working with MLflow2Seldon and need to clean up the space it was occupying, just run the following command:
$ make helm_delete
This command will delete:
- All resources required for this service created by the helm chart and the service itself
- The Kubernetes namespace in which SeldonDeployments were created (M2S_SELDON_DEPLOYMENT_NS)