Skip to content

Latest commit

 

History

History
141 lines (96 loc) · 7.24 KB

developer-guide.md

File metadata and controls

141 lines (96 loc) · 7.24 KB
Raw

Developer Guide

This developer guide is for people who want to contribute to the Katib project. If you're interesting in using Katib in your machine learning project, see the following user guides:

Requirements

Build from source code

Check source code as follows:

make build REGISTRY=<image-registry> TAG=<image-tag>

To use your custom images for the Katib components, modify Kustomization file and Katib Config

You can deploy Katib v1beta1 manifests into a Kubernetes cluster as follows:

make deploy

You can undeploy Katib v1beta1 manifests from a Kubernetes cluster as follows:

make undeploy

Modify controller APIs

If you want to modify Katib controller APIs, you have to generate deepcopy, clientset, listers, informers, open-api and Python SDK with the changed APIs. You can update the necessary files as follows:

make generate

Controller Flags

Below is a list of command-line flags accepted by Katib controller:

Name Type Default Description
enable-grpc-probe-in-suggestion bool true Enable grpc probe in suggestions
experiment-suggestion-name string "default" The implementation of suggestion interface in experiment controller
metrics-addr string ":8080" The address that the metrics endpoint binds to
healthz-addr string ":18080" The address that the healthz endpoint binds to
trial-resources []schema.GroupVersionKind null The list of resources that can be used as trial template, in the form: Kind.version.group (e.g. TFJob.v1.kubeflow.org)
webhook-inject-securitycontext bool false Inject the securityContext of container[0] in the sidecar
webhook-port int 8443 The port number to be used for admission webhook server
enable-leader-election bool false Enable leader election for katib-controller. Enabling this will ensure there is only one active katib-controller.
leader-election-id string "3fbc96e9.katib.kubeflow.org" The ID for leader election.

DB Manager Flags

Below is a list of command-line flags accepted by Katib DB Manager:

Name Type Default Description
connect-timeout time.Duration 60s Timeout before calling error during database connection

Workflow design

Please see workflow-design.md.

Katib admission webhooks

Katib uses three Kubernetes admission webhooks.

  1. validator.experiment.katib.kubeflow.org - Validating admission webhook to validate the Katib Experiment before the creation.

  2. defaulter.experiment.katib.kubeflow.org - Mutating admission webhook to set the default values in the Katib Experiment before the creation.

  3. mutator.pod.katib.kubeflow.org - Mutating admission webhook to inject the metrics collector sidecar container to the training pod. Learn more about the Katib's metrics collector in the Kubeflow documentation.

You can find the YAMLs for the Katib webhooks here.

Note: If you are using a private Kubernetes cluster, you have to allow traffic via TCP:8443 by specifying the firewall rule and you have to update the master plane CIDR source range to use the Katib webhooks

Katib cert generator

Katib uses the custom cert-generator Kubernetes Job to generate certificates for the webhooks.

Once Katib is deployed in the Kubernetes cluster, the cert-generator Job follows these steps:

  • Generate the self-signed CA certificate and private key.

  • Generate public certificate and private key signed with the key generated in the previous step.

  • Create a Kubernetes Secret with the signed certificate. Secret has the katib-webhook-cert name and cert-generator Job's ownerReference to clean-up resources once Katib is uninstalled.

    Once Secret is created, the Katib controller Deployment spawns the Pod, since the controller has the katib-webhook-cert Secret volume.

  • Patch the webhooks with the CABundle.

You can find the cert-generator source code here.

Implement a new algorithm and use it in Katib

Please see new-algorithm-service.md.

Katib UI documentation

Please see Katib UI README.

Design proposals

Please see proposals.