KServe
Deploying models to Kubernetes with KServe.
KServe
When to use it?
KServe encapsulates the complexity of autoscaling, networking, health checking, and server configuration to bring cutting-edge serving features like GPU Autoscaling, Scale to Zero, and Canary Rollouts to your ML deployments. It enables a simple, pluggable, and complete story for Production ML Serving including prediction, pre-processing, post-processing, and explainability. KServe is being used across various organizations.
You should use the KServe Model Deployer:
If you are looking to deploy your model with an advanced Model Inference Platform with Kubernetes, built for highly scalable use cases.
If you want to handle the lifecycle of the deployed model with no downtime, with the possibility of scaling to zero on GPUs.
Looking for out-of-the-box model serving runtimes that are easy to use and easy to deploy model from the majority of frameworks.
If you want more advanced deployment strategies like A/B testing, canary deployments, ensembles, and transformers.
How to deploy it?
ZenML provides a KServe flavor build on top of the KServe Integration to allow you to deploy and use your models in a production-grade environment. In order to use the integration you need to install it on your local machine to be able to register the KServe Model deployer with ZenML and add it to your stack:
To deploy and make use of the KServe integration we need to have the following prerequisites:
access to a Kubernetes cluster. The example accepts a
--kubernetes-context
command line argument. This Kubernetes context needs to point to the Kubernetes cluster where KServe model servers will be deployed. If the context is not explicitly supplied to the example, it defaults to using the locally active context.
Since the KServe Model Deployer is interacting with the KServe model serving Platform deployed on a Kubernetes cluster, you need to provide a set of configuration parameters. These parameters are:
kubernetes_context: the Kubernetes context to use to contact the remote KServe installation. If not specified, the current configuration is used. Depending on where the KServe model deployer is being used
kubernetes_namespace: the Kubernetes namespace where the KServe deployment servers are provisioned and managed by ZenML. If not specified, the namespace set in the current configuration is used.
base_url: the base URL of the Kubernetes ingress used to expose the KServe deployment servers.
Infrastructure Deployment
The KServe Model Deployer can be deployed directly from the ZenML CLI:
You can pass other configurations specific to the stack components as key-value arguments. If you don't provide a name, a random one is generated for you. For more information about how to work use the CLI for this, please refer to the dedicated documentation section.
Managing KServe Authentication
The KServe Model Deployer requires access to the persistent storage where models are located. In most cases, you will use the KServe model deployer to serve models that are trained through ZenML pipelines and stored in the ZenML Artifact Store, which implies that the KServe model deployer needs to access the Artifact Store.
When explicit credentials are configured in the Artifact Store, the KServe Model Deployer doesn't need any additional configuration and will use those credentials automatically to authenticate to the same persistent storage service used by the Artifact Store. If the Artifact Store doesn't have explicit credentials configured, then KServe will default to using whatever implicit authentication method is available in the Kubernetes cluster where it is running. For example, in AWS this means using the IAM role attached to the EC2 or EKS worker nodes, and in GCP this means using the service account attached to the GKE worker nodes.
If the Artifact Store used in combination with the KServe Model Deployer in the same ZenML stack does not have explicit credentials configured, then the KServe Model Deployer might not be able to authenticate to the Artifact Store which will cause the deployed model servers to fail.
To avoid this, we recommend that you use Artifact Stores with explicit credentials in the same stack as the KServe Model Deployer. Alternatively, if you're running KServe in one of the cloud providers, you should configure implicit authentication for the Kubernetes nodes.
If you want to use a custom persistent storage with KServe, or if you prefer to manually manage the authentication credentials attached to the KServe model servers, you can use the approach described in the next section.
Advanced: Configuring a Custom KServe Secret
This method is not recommended, because it limits the KServe model deployer to a single persistent storage service, whereas using the Artifact Store credentials gives you more flexibility in combining the KServe model deployer with any Artifact Store in the same ZenML stack.
How do you use it?
For registering the model deployer, we need the URL of the Istio Ingress Gateway deployed on the Kubernetes cluster. We can get this URL by running the following command (assuming that the service name is istio-ingressgateway
, deployed in the istio-system
namespace):
Now register the model deployer:
We can now use the model deployer in our stack.
As the packaging and preparation of the model artifacts to the right format can be a bit of a challenge, ZenML's KServe Integration comes with a built-in model deployment step that can be used to deploy your models with the minimum of effort.
This step will:
Verify if the model is already deployed in the KServe cluster. If not, it will deploy the model.
Prepare the model artifacts to the right format for the TF, MLServer runtimes servers.
Package, verify, and prepare the model artifact for the PyTorch runtime server since it requires additional files.
Upload the model artifacts to the Artifact Store.
An example of how to use the model deployment step is shown below.
Within the KServeDeploymentConfig
you can configure:
model_name
: the name of the model in the KServe cluster and in ZenML.replicas
: the number of replicas with which to deploy the modelpredictor
: the type of predictor to use for the model. The predictor type can be one of the following:tensorflow
,pytorch
,sklearn
,xgboost
,custom
.resources
: This can be configured by passing a dictionary with therequests
andlimits
keys. The values for these keys can be a dictionary with thecpu
andmemory
keys. The values for these keys can be a string with the amount of CPU and memory to be allocated to the model.
Custom Model Deployment
While KServe is a good fit for most use cases with the built-in model servers, it is not always the best fit for your custom model deployment use case. For that reason KServe allows you to create your own model server using the KServe ModelServer
API where you can customize the predict, the pre- and post-processing functions. With ZenML's KServe Integration, you can create your own custom model deployment code by creating a custom predict function that will be passed to a custom deployment step responsible for preparing a Docker image for the model server.
This custom_predict
function should be getting the model and the input data as arguments and returns the output data. ZenML will take care of loading the model into memory, starting the KServe ModelServer
that will be responsible for serving the model, and running the predict function.
Then this custom_predict
function path
can be passed to the custom deployment parameters.
Advanced Custom Code Deployment with KServe Integration
The built-in KServe custom deployment step is a good starting point for deploying your custom models. However, if you want to deploy more than the trained model, you can create your own Custom Model Class and a custom step to achieve this.
Last updated