decide-on-final-model.adoc

Observing The Performance Of Both Models

Run the following script to send sufficient transactions to the models.

/projects/{{USER_MODEL_REPO_NAME}}/bin/prod-mon-test.sh

Now look at the Grafana {{GRAFANA_URL}}[dashboard^], and see that the Recall score, TP / (TP + FN), for XGBoost is better compared to Logistic Regression.

Logistic Regression

XGBoost

Deploying The Chosen One

Let’s quickly change the SeldonDeployment to use XGBoost, deploy to staging and promote it to Production.

cd /projects/{{USER_MODEL_REPO_NAME}}
git checkout master
GIT_REV=`git rev-parse --short HEAD`
echo "GIT REVISION: $GIT_REV"
. src/seldon/config.sh

cd /projects/{{USER_DEPLOY_REPO_NAME}}
git checkout master
sed -e "s/_USER_/{{USER_ID}}/g" -e "s/_CONTAINER_REGISTRY_/$NEXUS_DOCKER_REGISTRY/g" -e "s/_IMAGE_NAME_/$IMAGE_NAME/g" -e "s/_GIT_REV_/$GIT_REV/g" seldon-model.yaml.tmpl > seldon.yaml
git commit -a -m "Update image tag to $IMAGE_NAME:$GIT_REV"

git checkout stage
git merge master

git checkout prod
git merge stage

git push -u -v origin

Observe that only a single classifier has been deployed.

After the final model has been deployed to OpenShift, you can run some basic tests.

/projects/{{USER_MODEL_REPO_NAME}}/bin/prod-test.sh

Tag It!

Great! The new XGBoost model is working better and we now can tag it as v2.0

cd /projects/{{USER_MODEL_REPO_NAME}}
git checkout stage
git tag -a v2.0 -m "v2.0"
git push -v origin v2.0

The tag has now been pushed to {{GIT_URL}}/{{USER_ID}}/{{USER_MODEL_REPO_NAME}}/src/v2.0[your^] git repository.