Customers can install in multiple ways and need to choose when to use a Trial install or a Production install. So we need to ask:
What is the difference between the Trial and Production Installations?
Can a Trial setup not be used for Production?
Let’s answer these questions.
Trial vs Production Install¶
The trial install is really a scripted installer that installs all components. Many organisations prefer to install each component individually (which we call the Production/Modular Install approach). The reasons are:
Many organisations have existing installs for oidc/identity, elastic and sometimes other tools.
Sometimes access to dockerhub and other public hosting is blocked, meaning each image has to be copied and checked in the org’s hosting.
The trial install is opinionated about versions - sometimes organisations have approved or preferred versions.
The trial install makes assumptions about volume sizes and resource allocations.
The trial install includes some components which are optional.
The trial setup for https uses self-signed certs and many organisations have certs that they prefer to use.
The gitops part of the trial install is optional but if used it assumes public GitHub.
The trial install can certainly be used at scale and is supported. It is just not customised to the organisation or the use case.
Easiest Ways to Get Started¶
The KIND setup is very quick to get running. It does not require a cloud provider account or any git repo access.
The next easiest option is the trial install. We suggest to setup without gitops first and then add gitops after (the trial setup has gitops as a separate script).
If moving to a production installation we suggest to add components progressively. In a minimal setup deploy can be run without auth or gitops and this helps simplify initial configuration so that each step can be verified. Our docs also include verification and troubleshooting steps for each component.
What if I also want an ML training platform?¶
Seldon tools can work with many ways of training models. Some users choose to train models in CI systems or notebooks. There’s also a range of ML training platforms such as kubeflow or mlflow.
We have a special relationship with kubeflow as major contributors. We see many customers interested in kubeflow so we provide an install option that includes kubeflow.
Other platforms can certainly be used. Seldon Deploy can work with any way of building and installing Seldon core models. Seldon core users make use of a wide range of platforms, including cloud provider offerings like sagemaker, as well as open source tools like kubeflow.
How to Size a Cluster?¶
The default install makes sizing assumptions for each of the components. But sizing varies with use.
Some customers will make more requests and put more load on elasticsearch. Others will have more users logged in viewing monitoring, which will mean more prometheus queries.
Sizes can be adjusted with kubernetes. Even with the trial install, the scripts are provided and resource values can be adjusted and reapplied. If cluster limits are reached then with Kubernetes nodes can be added to the cluster while it is running.