Architecture

This section covers the main architectural aspects you need to consider when deploying PostgreSQL in Kubernetes.

Synchronizing the state

PostgreSQL is a database management system and, as such, it needs to be treated as a stateful workload in Kubernetes. While stateless applications mainly use traffic redirection to achieve High Availability (HA) and Disaster Recovery (DR), in the case of a database, state must be replicated in multiple locations, preferably in a continuous and instantaneous way, by adopting either of the following two strategies:

• storage-level replication, normally persistent volumes
• application-level replication, in this specific case PostgreSQL

CloudNativePG relies on application-level replication, for a simple reason: the PostgreSQL database management system comes with robust and reliable built-in physical replication capabilities based on Write Ahead Log (WAL) shipping, which have been used in production by millions of users all over the world for over a decade.

PostgreSQL supports both asynchronous and synchronous streaming replication over the network, as well as asynchronous file-based log shipping (normally used as a fallback option, for example, to store WAL files in an object store). Replicas are usually called standby servers and can also be used for read-only workloads, thanks to the Hot Standby feature.

Kubernetes architecture

Kubernetes natively provides the possibility to span separate physical locations - also known as data centers, failure zones, or more frequently availability zones - connected to each other via redundant, low-latency, private network connectivity.

Being a distributed system, the recommended minimum number of availability zones for a Kubernetes cluster is three (3), in order to make the control plane resilient to the failure of a single zone. For details, please refer to "Running in multiple zones". This means that each data center is active at any time and can run workloads simultaneously.

Multi-availability zone Kubernetes clusters

The multi-availability zone Kubernetes architecture with three (3) or more zones is the one that we recommend for PostgreSQL usage. This scenario is typical of Kubernetes services managed by Cloud Providers.

Such an architecture enables the CloudNativePG operator to control the full lifecycle of a Cluster resource across the zones within a single Kubernetes cluster, by treating all the availability zones as active: this includes, among other operations, scheduling the workloads in a declarative manner (based on affinity rules, tolerations and node selectors), automated failover, self-healing, and updates. All will work seamlessly across the zones in a single Kubernetes cluster.

Please refer to the "PostgreSQL architecture" section below for details on how you can design your PostgreSQL clusters within the same Kubernetes cluster through shared-nothing deployments at the storage, worker node, and availability zone levels.

Additionally, you can leverage Kubernetes clusters to deploy distributed PostgreSQL topologies hosting "passive" PostgreSQL replica clusters in different regions and managing them via declarative configuration. This setup is ideal for disaster recovery (DR), read-only operations, or cross-region availability.

Single availability zone Kubernetes clusters

If your Kubernetes cluster has only one availability zone, CloudNativePG still provides you with a lot of features to improve HA and DR outcomes for your PostgreSQL databases, pushing the single point of failure (SPoF) to the level of the zone as much as possible - i.e. the zone must have an outage before your CloudNativePG clusters suffer a failure.

This scenario is typical of self-managed on-premise Kubernetes clusters, where only one data center is available.

Single availability zone Kubernetes clusters are the only viable option when only two data centers are available within reach of a low-latency connection (typically in the same metropolitan area). Having only two zones prevents the creation of a multi-availability zone Kubernetes cluster, which requires a minimum of three zones. As a result, users must create two separate Kubernetes clusters in an active/passive configuration, with the second cluster primarily used for Disaster Recovery (see the replica cluster feature).

Please refer to the "PostgreSQL architecture" section below for details on how you can design your PostgreSQL clusters within your single availability zone Kubernetes cluster through shared-nothing deployments at the storage and worker node levels only. For HA, in such a scenario it becomes even more important that the PostgreSQL instances be located on different worker nodes and do not share the same storage.

For DR, you can push the SPoF above the single zone, by using additional Kubernetes clusters to define a distributed topology hosting "passive" PostgreSQL replica clusters. As with other Kubernetes workloads in this scenario, promotion of a Kubernetes cluster as primary must be done manually.

Through the replica cluster feature, you can define a distributed PostgreSQL topology and coordinate a controlled switchover between data centers by first demoting the primary cluster and then promoting the replica cluster, without the need to re-clone the former primary. While failover is now fully declarative, automated failover across Kubernetes clusters is not within CloudNativePG's scope, as the operator can only function within a single Kubernetes cluster.

PostgreSQL architecture

CloudNativePG supports clusters based on asynchronous and synchronous streaming replication to manage multiple hot standby replicas within the same Kubernetes cluster, with the following specifications:

• One primary, with optional multiple hot standby replicas for HA

• Available services for applications:

  • -rw: applications connect only to the primary instance of the cluster
  • -ro: applications connect only to hot standby replicas for read-only-workloads (optional)
  • -r: applications connect to any of the instances for read-only workloads (optional)

• Shared-nothing architecture recommended for better resilience of the PostgreSQL cluster:

  • PostgreSQL instances should reside on different Kubernetes worker nodes and share only the network - as a result, instances should not share the storage and preferably use local volumes attached to the node they run on
  • PostgreSQL instances should reside in different availability zones within the same Kubernetes cluster / region

The below diagram provides a simplistic view of the recommended shared-nothing architecture for a PostgreSQL cluster spanning across 3 different availability zones, running on separate nodes, each with dedicated local storage for PostgreSQL data.

CloudNativePG automatically takes care of updating the above services if the topology of the cluster changes. For example, in case of failover, it automatically updates the -rw service to point to the promoted primary, making sure that traffic from the applications is seamlessly redirected.

Read-write workloads

Applications can decide to connect to the PostgreSQL instance elected as current primary by the Kubernetes operator, as depicted in the following diagram:

Applications can use the -rw suffix service.

In case of temporary or permanent unavailability of the primary, for High Availability purposes CloudNativePG will trigger a failover, pointing the -rw service to another instance of the cluster.

Read-only workloads

Applications can access hot standby replicas through the -ro service made available by the operator. This service enables the application to offload read-only queries from the primary node.

The following diagram shows the architecture:

Applications can also access any PostgreSQL instance through the -r service.

Deployments across Kubernetes clusters

In a distributed PostgreSQL cluster there can only be a single PostgreSQL instance acting as a primary at all times. This means that applications can only write inside a single Kubernetes cluster, at any time.

However, for business continuity objectives it is fundamental to:

• reduce global recovery point objectives (RPO) by storing PostgreSQL backup data in multiple locations, regions and possibly using different providers (Disaster Recovery)
• reduce global recovery time objectives (RTO) by taking advantage of PostgreSQL replication beyond the primary Kubernetes cluster (High Availability)

In order to address the above concerns, CloudNativePG introduces the concept of a PostgreSQL Topology that is distributed across different Kubernetes clusters and is made up of a primary PostgreSQL cluster and one or more PostgreSQL replica clusters. This feature is called distributed PostgreSQL topology with replica clusters, and it enables multi-cluster deployments in private, public, hybrid, and multi-cloud contexts.

A replica cluster is a separate Cluster resource that is in continuous recovery, replicating from another source, either via WAL shipping from a WAL archive or via streaming replication from a primary or a standby (cascading).

The diagram below depicts a PostgreSQL cluster spanning over two different Kubernetes clusters, where the primary cluster is in the first Kubernetes cluster and the replica cluster is in the second. The second Kubernetes cluster acts as the company's disaster recovery cluster, ready to be activated in case of disaster and unavailability of the first one.

A replica cluster can have the same architecture as the primary cluster. Instead of a primary instance, a replica cluster has a designated primary instance, which is a standby server with an arbitrary number of cascading standby servers in streaming replication (symmetric architecture).

The designated primary can be promoted at any time, transforming the replica cluster into a primary cluster capable of accepting write connections. This is typically triggered by:

• Human decision: You choose to make the other PostgreSQL cluster (or the entire Kubernetes cluster) the primary. To avoid data loss and ensure that the former primary can follow without being re-cloned (especially with large data sets), you first demote the current primary, then promote the designated primary using the API provided by CloudNativePG.
• Unexpected failure: If the entire Kubernetes cluster fails, you might experience data loss, but you need to failover to the other Kubernetes cluster by promoting the PostgreSQL replica cluster.

The designated primary in the above example is fed via WAL streaming (primary_conninfo), with fallback option for file-based WAL shipping through the restore_command and barman-cloud-wal-restore.

CloudNativePG allows you to define topologies with multiple replica clusters. You can also define replica clusters with a lower number of replicas, and then increase this number when the cluster is promoted to primary.