XGBoost Operator

⚠️ kubeflow/xgboost-operator is not maintained

This operator has been merged into Kubeflow Training Operator. This repository is not maintained and has been archived.

Overview

Incubating project for XGBoost operator. The XGBoost operator makes it easy to run distributed XGBoost job training and batch prediction on Kubernetes cluster.

The overall design can be found here.

This repository contains the specification and implementation of XGBoostJob custom resource definition. Using this custom resource, users can create and manage XGBoost jobs like other built-in resources in Kubernetes.

Prerequisites

• Kubernetes >= 1.8
• kubectl

Install XGBoost Operator

You can deploy the operator with default settings by running the following commands using kustomize:

cd manifests
kubectl create namespace kubeflow
kustomize build base | kubectl apply -f -

Note that since Kubernetes v1.14, kustomize became a subcommand in kubectl so you can also run the following command instead:

kubectl kustomize base | kubectl apply -f -

Build XGBoost Operator

XGBoost Operator is developed based on Kubebuilder and Kubeflow Common.

You can follow the installation guide of Kubebuilder to install XGBoost operator into the Kubernetes cluster.

You can check whether the XGBoostJob custom resource has been installed via:

kubectl get crd

The output should include xgboostjobs.kubeflow.org like the following:

NAME                                                CREATED AT
xgboostjobs.xgboostjob.kubeflow.org                 2021-03-24T22:03:07Z

If it is not included you can add it as follows:

## setup the build enviroment
export GOPATH=$HOME/go
export PATH=$PATH:$GOPATH/bin
export GO111MODULE=on
cd $GOPATH
mkdir src/github.com/kubeflow
cd src/github.com/kubeflow

## clone the code
git clone [email protected]:kubeflow/xgboost-operator.git
cd xgboost-operator

## build and install xgboost operator
make
make install
make run

If the XGBoost Job operator can be installed into cluster, you can view the logs likes this

Logs

{"level":"info","ts":1589406873.090652,"logger":"entrypoint","msg":"setting up client for manager"}
{"level":"info","ts":1589406873.0991302,"logger":"entrypoint","msg":"setting up manager"}
{"level":"info","ts":1589406874.2192929,"logger":"entrypoint","msg":"Registering Components."}
{"level":"info","ts":1589406874.219318,"logger":"entrypoint","msg":"setting up scheme"}
{"level":"info","ts":1589406874.219448,"logger":"entrypoint","msg":"Setting up controller"}
{"level":"info","ts":1589406874.2194738,"logger":"controller","msg":"Running controller in local mode, using kubeconfig file"}
{"level":"info","ts":1589406874.224564,"logger":"controller","msg":"gang scheduling is set: ","gangscheduling":false}
{"level":"info","ts":1589406874.2247412,"logger":"kubebuilder.controller","msg":"Starting EventSource","controller":"xgboostjob-controller","source":"kind source: /, Kind="}
{"level":"info","ts":1589406874.224958,"logger":"kubebuilder.controller","msg":"Starting EventSource","controller":"xgboostjob-controller","source":"kind source: /, Kind="}
{"level":"info","ts":1589406874.2251048,"logger":"kubebuilder.controller","msg":"Starting EventSource","controller":"xgboostjob-controller","source":"kind source: /, Kind="}
{"level":"info","ts":1589406874.225237,"logger":"entrypoint","msg":"setting up webhooks"}
{"level":"info","ts":1589406874.225247,"logger":"entrypoint","msg":"Starting the Cmd."}
{"level":"info","ts":1589406874.32791,"logger":"kubebuilder.controller","msg":"Starting Controller","controller":"xgboostjob-controller"}
{"level":"info","ts":1589406874.430336,"logger":"kubebuilder.controller","msg":"Starting workers","controller":"xgboostjob-controller","worker count":1}

Creating a XGBoost Training/Prediction Job

You can create a XGBoost training or prediction (batch oriented) job by modifying the XGBoostJob config file. See the distributed XGBoost Job training and prediction example. You can change the config file and related python file (i.e., train.py or predict.py) based on your requirement.

Following the job configuration guild in the example, you can deploy a XGBoost Job to start training or prediction like:

## For training job
cat config/samples/xgboost-dist/xgboostjob_v1_iris_train.yaml
kubectl create -f  config/samples/xgboost-dist/xgboostjob_v1_iris_train.yaml

## For batch prediction job
cat config/samples/xgboost-dist/xgboostjob_v1_iris_predict.yaml
kubectl create -f  config/samples/xgboost-dist/xgboostjob_v1_iris_predict.yaml

Monitor a distributed XGBoost Job

Once the XGBoost job is created, you should be able to watch how the related pod and service working. Distributed XGBoost job is trained by synchronizing different worker status via tne Rabit of XGBoost. You can also monitor the job status.

 kubectl get -o yaml XGBoostJob/xgboost-dist-iris-test-train

Here is the sample output when training job is finished.

XGBoost Job Details

apiVersion: xgboostjob.kubeflow.org/v1
kind: XGBoostJob
metadata:
  annotations:
    kubectl.kubernetes.io/last-applied-configuration: |
      {"apiVersion":"xgboostjob.kubeflow.org/v1","kind":"XGBoostJob","metadata":{"annotations":{},"name":"xgboost-