A toy sample that sets up a geo-distributed Wordpress across three Kubernetes clusters in three GCE regions
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This repository sets up a geo-distributed Wordpress installation running across three Google Cloud regions. It is intended to showcase the feasibilty of using NewSQL databases to build fault-tolerant systems.

NOTE: the project is intended as an experiment, and is nowhere near ready for production.

Terraform is used to spawn three Kubernetes clusters in three separate Google Cloud regions.

After this, Kubernetes manifests for a geo-wordpress stack are rendered and deployed to the clusters. A separate geo-wordpress stack is run on each cluster and it consists of:

  • TiDB: a MySQL-compatible distributed database offering high availability and strong (ACID) consistency. TiDB operates in a distributed manner and uses the Raft protocol to agree on and order transactions. One instance is run in each region, meaning that the database can tolerate the loss of one region and still be able to form a quorum and, hence, remain available.
  • nfs-provisioner: A dynamic provisioner of NFS volumes for use by Wordpress pods.
  • Wordpress: Uses TiDB as a drop-in replacement for MySQL and mounts an NFS volume created by the nfs-provisioner.
  • SyncThing: forms a "file synchronization group" with the SyncThing processes on the other clusters and listens for filesystem notifications (inotify) on the Wordpress volume. Whenever a change is detected, the change is propagated to its peers. Think of it as a bidirectional rsync. If a peer gets disconnected from its group it should be able to catch up when it gets back, given that the system clocks on the hosts are fairly well synchronized.

The whole setup is fronted by a global cloud load-balancer set up to spread traffic across all regions. The load-balancer uses a health check to detect failed nodes and, if detected, take that target out of rotation.

The setup is illustrated by the following image:


Bring up Kubernetes clusters

  1. Prepare a credentials file for GCE. You can fill out the gce-secrets.var. The parameters can be found in infra/terraform/main.tf.

  2. Add selected regions (and any additional variables) to clusters.var.

  3. Bring up Kubernetes clusters in AWS, GCE, and Azure. Refer to infra/terraform/main.tf for all the variables.

     terraform init infra/terraform
     terraform apply --var-file gce-secrets.var --var-file clusters.var \

When terraform completes, note the ip address of the cloud load-balancer.

Install the geo-wordpress stack onto the Kubernetes clusters

If Terraform finished successfully, just run:


The script will output a kubeconfig file which can be used to communicate with the clusters using kubectl. To use it, set

export KUBECONFIG=$PWD/kubeconfig

You can now wait for all pods to enter the Running state. In separate terminal windows, issue the following commands:

KUBECONFIG=/tmp/cluster0.config watch kubectl get pods -n wp
KUBECONFIG=/tmp/cluster1.config watch kubectl get pods -n wp
KUBECONFIG=/tmp/cluster2.config watch kubectl get pods -n wp

When everything is Running, navigate your web browser to http://$(terraform output loadbalancer_ip). This should take you to the Wordpress installation page and your ready to go.


You can verify that files are properly synchronized by, for instance, uploading media and checking, for each cluster, that they have the same content in their upload folder. For each cluster, run something like (replace pod name):

watch kubectl exec -n wp wordpress-654b4dd45b-7j9fp -- ls -al /var/www/html/wp-content/uploads/2018/11

Tear down clusters

Start by stopping instances and deleting the geokube-* created routes in Google Cloud, since terraform will otherwise fail:

gcloud compute instances list --filter="name~geokube" --format=json | jq -r .[].selfLink | xargs gcloud compute instances stop

gcloud compute routes list --filter="name~geokube" --format=json | jq -r '.[].name' | xargs gcloud compute routes delete --quiet

Then run:

terraform destroy --force --var-file gce-secrets.var --var-file clusters.var infra/terraform