Model Deployers

How to deploy your models and serve real-time predictions
Model Deployment is the process of making a machine learning model available to make predictions and decisions on real world data. Getting predictions from trained models can be done in different ways depending on the use-case, a batch prediction is used to generate prediction for a large amount of data at once, while a real-time prediction is used to generate predictions for a single data point at a time.
Model deployers are stack components responsible for serving models on a real-time or batch basis.
Online serving is the process of hosting and loading machine-learning models as part of a managed web service and providing access to the models through an API endpoint like HTTP or GRPC. Once deployed, model inference can be triggered at any time, and you can send inference requests to the model through the web service's API and receive fast, low-latency responses.
Batch inference or offline inference is the process of making a machine learning model make predictions on a batch observations. This is useful for generating predictions for a large amount of data at once. The predictions are usually stored as files or in a database for end users or business applications.

When to use it?

The model deployers are optional components in the ZenML stack. They are used to deploy machine learning models to a target environment either a development (local) or a production (Kubernetes), the model deployers are mainly used to deploy models for real time inference use cases. With the model deployers and other stack components, you can build pipelines that are continuously trained and deployed to a production.

How they experiment trackers slot into the stack

Here is an architecture diagram that shows how model deployers fit into the overall story of a remote stack.
Experiment Tracker

Model Deployers Flavors

ZenML comes with a local MLflow model deployer which is a simple model deployer that deploys models to a local MLflow server. Additional model deployers that can be used to deploy models on production environments are provided by integrations:
Model Deployer
Deploys ML Model locally
Build and Deploy ML models locally or for production grade (Cloud, K8s)
seldon Core
Built on top of Kubernetes to deploy models for production grade environment
Kubernetes based model deployment framework
Extend the Artifact Store abstraction and provide your own implementation
Every model deployer may have different attributes that must be configured in order to interact with the model serving tool, framework or platform (e.g. hostnames, URLs, references to credentials, other client related configuration parameters). The following example shows the configuration of the MLflow and Seldon Core model deployers:
# Configure MLflow model deployer
zenml model-deployer register mlflow --flavor=mlflow
# Configure Seldon Core model deployer
zenml model-deployer register seldon --flavor=seldon \
--kubernetes_context=zenml-eks --kubernetes_namespace=zenml-workloads \

The role Model Deployer plays in a ZenML Stack

  1. 1.
    Holds all the stack related configuration attributes required to interact with the remote model serving tool, service or platform (e.g. hostnames, URLs, references to credentials, other client related configuration parameters). The following are examples of configuring the MLflow and Seldon Core Model Deployers and registering them as a Stack component:
    zenml integration install mlflow
    zenml model-deployer register mlflow --flavor=mlflow
    zenml stack register local_with_mlflow -m default -a default -o default -d mlflow --set
    zenml integration install seldon
    zenml model-deployer register seldon --flavor=seldon \
    --kubernetes_context=zenml-eks --kubernetes_namespace=zenml-workloads \
    zenml stack register seldon_stack -m default -a aws -o default -d seldon
  2. 2.
    Implements the continuous deployment logic necessary to deploy models in a way that updates an existing model server that is already serving a previous version of the same model instead of creating a new model server for every new model version. Every model server that the Model Deployer provisions externally to deploy a model is represented internally as a Service object that may be accessed for visibility and control over a single model deployment. This functionality can be consumed directly from ZenML pipeline steps, but it can also be used outside the pipeline to deploy ad-hoc models. The following code is an example of using the Seldon Core Model Deployer to deploy a model inside a ZenML pipeline step:
    from zenml.artifacts import ModelArtifact
    from zenml.environment import Environment
    from zenml.integrations.seldon.model_deployers import SeldonModelDeployer
    from zenml.integrations.seldon.services.seldon_deployment import (
    from zenml.steps import (
    def seldon_model_deployer_step(
    context: StepContext,
    model: ModelArtifact,
    ) -> SeldonDeploymentService:
    model_deployer = SeldonModelDeployer.get_active_model_deployer()
    # get pipeline name, step name and run id
    step_env = Environment()[STEP_ENVIRONMENT_NAME]
    pipeline_name = step_env.pipeline_name,
    pipeline_run_id = step_env.pipeline_run_id,
    pipeline_step_name = step_env.step_name,
    service = model_deployer.deploy_model(
    service_config, replace=True, timeout=300
    f"Seldon deployment service started and reachable at:\n"
    f" {service.prediction_url}\n"
    return service
  3. 3.
    Acts as a registry for all Services that represent remote model servers. External model deployment servers can be listed and filtered using a variety of criteria, such as the name of the model or the names of the pipeline and step that was used to deploy the model. The Service objects returned by the Model Deployer can be used to interact with the remote model server, e.g. to get the operational status of a model server, the prediction URI that it exposes, or to stop or delete a model server:
    from zenml.integrations.seldon.model_deployers import SeldonModelDeployer
    model_deployer = SeldonModelDeployer.get_active_model_deployer()
    services = model_deployer.find_model_server(
    if services:
    if services[0].is_running:
    f"Seldon deployment service started and reachable at:\n"
    f" {services[0].prediction_url}\n"
    elif services[0].is_failed:
    f"Seldon deployment service is in a failure state. "
    f"The last error message was: {services[0].status.last_error}"
    print(f"Seldon deployment service is not running")
    # start the service
    # delete the service
    model_deployer.delete_service(services[0].uuid, timeout=100, force=False)

Custom pre-processing and post-processing

Pre-processing is the process of transforming the data before it is passed to the machine learning model.
Post-processing is the process of transforming the data after it is returned from the machine learning model, and before it is returned to the user.
Both pre- and post-processing are very essential for the model deployment process. Most models require specific input formats which requires transforming the data before it is passed to the model and after it is returned from the model. ZenML allows you to define your own pre- and post-processing in two ways:
  1. 1.
    At the pipeline level by defining custom steps before and after the predict step in the ZenML pipeline.
  2. 2.
    At the model deployment tool level by defining a custom predict, pre- and post-processing functions that would be wrapped in a Docker container and executed on the model deployment server.
The custom model deployment support is available only for the following integrations:
### How to Interact with model deployer after deployment?
When a Model Deployer is part of the active ZenML Stack, it is also possible to interact with it from the CLI to list, start, stop or delete the model servers that is manages:
$ zenml model-deployer models list
┃ ✅ │ 8cbe671b-9fce-4394-a051-68e001f92765 │ continuous_deployment_pipeline │ seldon_model_deployer_step ┃
$ zenml model-deployer models describe 8cbe671b-9fce-4394-a051-68e001f92765
Properties of Served Model 8cbe671b-9fce-4394-a051-68e001f92765
┃ MODEL_NAME │ mnist ┃
┃ MODEL_URI │ s3://zenprojects/seldon_model_deployer_step/output/884/seldon ┃
┃ PIPELINE_NAME │ continuous_deployment_pipeline ┃
┃ PIPELINE_RUN_ID │ continuous_deployment_pipeline-11_Apr_22-09_39_27_648527 ┃
┃ PIPELINE_STEP_NAME │ seldon_model_deployer_step ┃
┃ PREDICTION_URL │ http://abb84c444c7804aa98fc8c097896479d-377673393.us-east-1.elb.amazonaws.com/seldon/… ┃
┃ SELDON_DEPLOYMENT │ zenml-8cbe671b-9fce-4394-a051-68e001f92765 ┃
┃ STATUS │ ✅ ┃
┃ STATUS_MESSAGE │ Seldon Core deployment 'zenml-8cbe671b-9fce-4394-a051-68e001f92765' is available ┃
┃ UUID │ 8cbe671b-9fce-4394-a051-68e001f92765 ┃
$ zenml model-deployer models get-url 8cbe671b-9fce-4394-a051-68e001f92765
Prediction URL of Served Model 8cbe671b-9fce-4394-a051-68e001f92765 is:
$ zenml model-deployer models delete 8cbe671b-9fce-4394-a051-68e001f92765
In Python, you can alternatively discover the prediction URL of a deployed model by inspecting the metadata of the step that deployed the model:
from zenml.post_execution import get_run
pipeline_run = get_run("<PIPELINE_RUN_NAME>")
deployer_step = pipeline_run.get_step("<NAME_OF_MODEL_DEPLOYER_STEP>")
deployed_model_url = deployer_step.metadata["deployed_model_url"].value
Services can be passed through steps like any other object, and used to interact with the external systems that they represent:
from zenml.steps import step
def my_step(my_service: MyService) -> ...:
if not my_service.is_running:
my_service.start() # starts service
my_service.stop() # stops service
The ZenML integrations that provide Model Deployer stack components also include standard pipeline steps that can directly be inserted into any pipeline to achieve a continuous model deployment workflow. These steps take care of all the aspects of continuously deploying models to an external server and saving the Service configuration into the Artifact Store, where they can be loaded at a later time and re-create the initial conditions used to serve a particular model.