Skip to content

Design overview

The design of the Operator is tighly bound to the Percona Server for MongoDB replica set or sharded cluster. Replica set cluster is briefly described in the following diagram.

image

A replica set consists of one primary server and several secondary ones (two in the picture), and the client application accesses the servers via a driver.

In the case of a sharded cluster, each shard is a replica set which contains a subset of data stored in the database, and the mongos query router acts as an entry point for client applications. You can find out more details about sharding on a dedicated documentation page, and a simplified diagram is as follows:

image

To provide high availability the Operator uses node affinity to run MongoDB instances on separate worker nodes if possible, and the database cluster is deployed as a single Replica Set with at least three nodes. If a node fails, the pod with the mongod process is automatically re-created on another node. If the failed node was hosting the primary server, the replica set initiates elections to select a new primary. If the failed node was running the Operator, Kubernetes will restart the Operator on another node, so normal operation will not be interrupted.

Client applications should use a mongo+srv URI for the connection. This allows the drivers (4.2 and up) to retrieve the list of replica set members from DNS SRV entries without having to list hostnames for the dynamically assigned nodes.

Note

The Operator uses security settings which are more secure than the default Percona Server for MongoDB setup. The initial configuration contains default passwords for all needed user accounts, which should be changed in the production environment, as stated in the installation instructions.

image

To provide data storage for stateful applications, Kubernetes uses Persistent Volumes. A PersistentVolumeClaim (PVC) is used to implement the automatic storage provisioning to pods. If a failure occurs, the Container Storage Interface (CSI) should be able to re-mount storage on a different node. The PVC StorageClass must support this feature (Kubernetes and OpenShift support this in versions 1.9 and 3.9 respectively).

The Operator functionality extends the Kubernetes API with PerconaServerMongoDB object, and it is implemented as a golang application. Each PerconaServerMongoDB object maps to one separate Percona Server for MongoDB setup. The Operator listens to all events on the created objects. When a new PerconaServerMongoDB object is created, or an existing one undergoes some changes or deletion, the operator automatically creates/changes/deletes all needed Kubernetes objects with the appropriate settings to provide a properly operating replica set.


Last update: 2024-09-15