Our requirement was an SQL server for modest loads but with high availability because it backs our login system and public web.
We want automatic failover, not manual, and a single configuration to maintain, not leaders and followers.
Initially we tried Postgres, but clustering solutions like pgpool-II felt a bit outdated, and patroni etc was overly complex for our needs.
Galera is a good fit for us with Kubernetes because it allows nodes to share the same configuration through wsrep_cluster_address="gcomm://host-0,host-1... where access to only 1 of those instances lets a new instance join the cluster.
As of MariaDB 10.1 the official docker image comes with "wsrep" support. Using official images direcly mean less maintenance for us.
Kubernetes recently added support for "stateful applications", the StatefulSet. There's a semi-official example using MySQL. It's more an example of what you can do with StatefulSet than a production setup. It uses the init container concept, which looks heavily alpha, to try to automate installation.
Using a semi-manual bootstrap process and a container with galera support built in, we were able to simplify the setup and thus the maintenance.
Testing this is a TODO.
The database should remain available during cluster upgrades and node replacement, i.e. when something like 1 out of 3 VMs is gone. A Kubernetes upgrade on GKE means 1 at a time is gone, with only a short period of time between one upgrade completes and the next starts.
Failure on 2 VMs concurrently probably mean our cluster is going down badly, and that SQL downtime isn't our major headache. In this case we expect the DB to recover automatically once 2 out of 3 nodes are live.
With 3 nodes we aim for the following characteristics, prioritizing consistency over performance:
- Consistent reads if 1 or even 2 nodes are gone.
- Block writes if 2 nodes are gone.
- Ideally accept writes if 1 nodes are gone.
Unless your Kubernetes setup has volume provisioning for StatefulSet (GKE has) you need to make sure the Persistent Volumes exist first.
Then:
- Create namespace.
- Create
10pvc.ymlif you created PVs manually. - Create configmap (see
40configmap.sh) and secret (see41secret.sh). - Create StatefulSet's "headless" service
20mariadb-service.yml. - Create the service that other applications depend on
30mysql-service.yml.
After that start bootstrapping.
First get a single instance with --wsrep-new-cluster up and running:
kubectl create -f ./
kubectl logs -f mariadb-0
You should see something like
...[Note] WSREP: Quorum results:
version = 3,
component = PRIMARY,
conf_id = 0,
members = 1/1 (joined/total),
act_id = 4,
last_appl. = -1,
protocols = 0/7/3 (gcs/repl/appl),
Now keep that pod running, but change StatefulSet to create normal replicas.
./70unbootstrap.sh
This scales to three nodes. You can kubectl logs -f mariadb-1 to see something like:
[Note] WSREP: Quorum results:
version = 3,
component = PRIMARY,
conf_id = 4,
members = 2/3 (joined/total),
act_id = 4,
last_appl. = 0,
protocols = 0/7/3 (gcs/repl/appl),
Now you can kubectl delete mariadb-0 and it'll be re-created without the --wsrep-new-cluster argument. Logs will confirm that the new mariadb-0 joins the cluster.
Keep at least 1 node running at all times - which is what you want anyway, and the manual "unbootstrap" step isn't a big deal.
Carefully consider the security implications before you create this. Note that it uses a non-official image.
kubectl create -f myadmin/
PhpMyAdmin has a login page where you need a mysql user. The default for root is localhost-only access (the merits of this in a micorservices context can be discussed). To allow root login from phpMyAdmin:
# enter pod
kubectl exec -ti mariadb-0 -- /bin/bash
# inside pod
mysql --password=$MYSQL_ROOT_PASSWORD -e "GRANT ALL PRIVILEGES ON *.* TO 'root'@'%' IDENTIFIED BY '$MYSQL_ROOT_PASSWORD' WITH GRANT OPTION;