.png%3Falt%3Dmedia%26token%3D3f1e92da-c9f8-4b18-a9ca-5f0c5faa81ad)
.png%3Falt%3Dmedia%26token%3D3f1e92da-c9f8-4b18-a9ca-5f0c5faa81ad)
🏗
Set up your MLOps platform
Birds-eye view on the necessities of your MLOps platform.
In order to establish your own MLOps platform with ZenML, you need to deploy a few ingredients:
System architecture with a deployment of ZenML along with a cloud stack
The ZenML server exposes a RESTful API to the client and other consumers.
The database will act as the central metadata store that tracks all pipeline runs and all the configurations of the stacks.
For a production setting, you should also set up a secret manager as a backend for all the secrets that will be used to configure stacks
The Compute infrastructure (e.g. Kubernetes or serverless alternatives)
- This will be used to run the pipeline code runs in production using the orchestrator and step operator stack components
- Optionally, the same infrastructure can be used for the deployment of models using the model deployer stack component
The orchestrator will need to have access to all the other stack components, and an egress path to post to the ZenML Server
The orchestrator/step operator will use this as the artifact store where step outputs are persisted. You can also use the same infrastructure to host your data as data sources/ data sinks
This is where the docker images for all pipeline code is pushed. The orchestrator will consume docker images from here
Deployments of all the other tools that you need (such as experiment trackers, model registries, and feature stores). Learn more about the options in our Component Guide.
There are some additional tools that should be considered in the setup of the MLOps Platform
Optionally, you can configure an image builder service to build the docker images centrally, rather than doing so on the client machine.
Code Repositories can be configured within the ZenML deployment. With a code repository configured the pipeline code no longer needs to be baked into the docker images. Instead, the code is loaded at runtime within the orchestrator.
Coming Soon: True Client-Server Architecture
We are hard at work implementing a true client-server architecture where the client no longer needs to have any access to the individual stack components. In the meantime, we recommend using your favorite CI/CD tool to enable using ZenML without the need to configure direct access to stack components from the client machines.
Last modified 8d ago