[ English | Indonesia | 한국어 (대한민국) | Deutsch | English (United Kingdom) ]

Cleaning the Deployment

Removing Helm Charts

To delete an installed helm chart, use the following command:

helm delete ${RELEASE_NAME} --purge

This will delete all Kubernetes resources generated when the chart was instantiated. However for OpenStack charts, by default, this will not delete the database and database users that were created when the chart was installed. All OpenStack projects can be configured such that upon deletion, their database will also be removed. To delete the database when the chart is deleted the database drop job must be enabled before installing the chart. There are two ways to enable the job, set the job_db_drop value to true in the chart’s values.yaml file, or override the value using the helm install command as follows:

helm install ${RELEASE_NAME} --set manifests.job_db_drop=true

Environment tear-down

To tear-down, the development environment charts should be removed first from the ‘openstack’ namespace and then the ‘ceph’ namespace using the commands from the Removing Helm Charts section. Additionally charts should be removed from the ‘nfs’ and ‘libvirt’ namespaces if deploying with NFS backing or bare metal development support. You can run the following commands to loop through and delete the charts, then stop the kubelet systemd unit and remove all the containers before removing the directories used on the host by pods.

for NS in openstack ceph nfs libvirt; do
   helm ls --namespace $NS --short | xargs -r -L1 -P2 helm delete --purge
done

sudo systemctl stop kubelet
sudo systemctl disable kubelet

sudo docker ps -aq | xargs -r -L1 -P16 sudo docker rm -f

sudo rm -rf /var/lib/openstack-helm/*

# NOTE(portdirect): These directories are used by nova and libvirt
sudo rm -rf /var/lib/nova/*
sudo rm -rf /var/lib/libvirt/*
sudo rm -rf /etc/libvirt/qemu/*
#NOTE(chinasubbareddy) cleanup LVM volume groups in case of disk backed ceph osd deployments
for VG in `vgs|grep -v VG|grep -i ceph|awk '{print $1}'`; do
  echo $VG
  vgremove -y $VG
done
# lets delete loopback devices setup for ceph, if the device names are different in your case,
# please update them here as environmental variables as shown below.
: "${CEPH_OSD_DATA_DEVICE:=/dev/loop0}"
: "${CEPH_OSD_DB_WAL_DEVICE:=/dev/loop1}"
if [ ! -z "$CEPH_OSD_DATA_DEVICE" ]; then
  ceph_osd_disk_name=`basename "$CEPH_OSD_DATA_DEVICE"`
  if losetup -a|grep $ceph_osd_disk_name; then
     losetup -d "$CEPH_OSD_DATA_DEVICE"
  fi
fi
if [ ! -z "$CEPH_OSD_DB_WAL_DEVICE" ]; then
  ceph_db_wal_disk_name=`basename "$CEPH_OSD_DB_WAL_DEVICE"`
  if losetup -a|grep $ceph_db_wal_disk_name; then
     losetup -d "$CEPH_OSD_DB_WAL_DEVICE"
  fi
fi
echo "let's disable the service"
sudo systemctl disable loops-setup
echo "let's remove the service to setup loopback devices"
if [ -f "/etc/systemd/system/loops-setup.service" ]; then
  rm /etc/systemd/system/loops-setup.service
fi

# NOTE(portdirect): Clean up mounts left behind by kubernetes pods
sudo findmnt --raw | awk '/^\/var\/lib\/kubelet\/pods/ { print $1 }' | xargs -r -L1 -P16 sudo umount -f -l

These commands will restore the environment back to a clean Kubernetes deployment, that can either be manually removed or over-written by restarting the deployment process. It is recommended to restart the host before doing so to ensure any residual state, eg. Network interfaces are removed.