Operator capability levels

These capabilities were implemented by CloudNativePG, classified using the Operator SDK definition of Capability Levels framework.

Each capability level is associated with a certain set of management features the operator offers:

1. Basic install
2. Seamless upgrades
3. Full lifecycle
4. Deep insights
5. Auto pilot

Level 1: Basic install

Capability level 1 involves installing and configuring the operator. This category includes usability and user experience enhancements, such as improvements in how you interact with the operator and a PostgreSQL cluster configuration.

Operator deployment via declarative configuration

The operator is installed in a declarative way using a Kubernetes manifest that defines four major CustomResourceDefinition objects: Cluster, Pooler, Backup, and ScheduledBackup.

PostgreSQL cluster deployment via declarative configuration

You define a PostgreSQL cluster (operand) using the Cluster custom resource in a fully declarative way. The PostgreSQL version is determined by the operand container image defined in the CR, which is automatically fetched from the requested registry. When deploying an operand, the operator also creates the following resources: Pod, Service, Secret, ConfigMap,PersistentVolumeClaim, PodDisruptionBudget, ServiceAccount, RoleBinding, and Role.

Override of operand images through the CRD

The operator is designed to support any operand container image with PostgreSQL inside. By default, the operator uses the latest available minor version of the latest stable major version supported by the PostgreSQL community and published on ghcr.io. You can use any compatible image of PostgreSQL supporting the primary/standby architecture directly by setting the imageName attribute in the CR. The operator also supports imagePullSecrets to access private container registries, and it supports digests and tags for finer control of container image immutability.

Labels and annotations

You can configure the operator to support inheriting labels and annotations that are defined in a cluster's metadata. The goal is to improve the organization of the CloudNativePG deployment in your Kubernetes infrastructure.

Self-contained instance manager

Instead of relying on an external tool to coordinate PostgreSQL instances in the Kubernetes cluster pods, such as Patroni or Stolon, the operator injects the operator executable inside each pod, in a file named /controller/manager. The application is used to control the underlying PostgreSQL instance and to reconcile the pod status with the instance based on the PostgreSQL cluster topology. The instance manager also starts a web server that's invoked by the kubelet for probes. Unix signals invoked by the kubelet are filtered by the instance manager. Where appropriate, they're forwarded to the postgres process for fast and controlled reactions to external events. The instance manager is written in Go and has no external dependencies.

Storage configuration

Storage is a critical component in a database workload. Taking advantage of the Kubernetes native capabilities and resources in terms of storage, the operator gives you enough flexibility to choose the right storage for your workload requirements, based on what the underlying Kubernetes environment can offer. This implies choosing a particular storage class in a public cloud environment or fine-tuning the generated PVC through a PVC template in the CR's storage parameter.

For better performance and finer control, you can also choose to host your cluster's write-ahead log (WAL, also known as pg_wal) on a separate volume, preferably on different storage. The "Benchmarking" section of the documentation provides detailed instructions on benchmarking both storage and the database before production. It relies on the cnpg plugin to ensure optimal performance and reliability.

Replica configuration

The operator detects replicas in a cluster through a single parameter, called instances. If set to 1, the cluster comprises a single primary PostgreSQL instance with no replica. If higher than 1, the operator manages instances -1 replicas, including high availability (HA) through automated failover and rolling updates through switchover operations.

CloudNativePG manages replication slots for all the replicas in the HA cluster. The implementation is inspired by the previously proposed patch for PostgreSQL, called failover slots.

Database configuration

The operator is designed to manage a PostgreSQL cluster with a single database. The operator transparently manages access to the database through three Kubernetes services provisioned and managed for read-write, read, and read-only workloads. Using the convention-over-configuration approach, the operator creates a database called app, by default owned by a regular Postgres user with the same name. You can specify both the database name and the user name, if required.

Although no configuration is required to run the cluster, you can customize both PostgreSQL runtime configuration and PostgreSQL host-based authentication rules in the postgresql section of the CR.

Configuration of Postgres roles, users, and groups

CloudNativePG supports management of PostgreSQL roles, users, and groups through declarative configuration using the .spec.managed.roles stanza.

Pod security policies

For InfoSec requirements, the operator doesn't require privileged mode for any container. It enforces a read-only root filesystem to guarantee containers immutability for both the operator and the operand pods. It also explicitly sets the required security contexts.

Affinity

The cluster's affinity section enables fine-tuning of how pods and related resources, such as persistent volumes, are scheduled across the nodes of a Kubernetes cluster. In particular, the operator supports:

• Pod affinity and anti-affinity
• Node selector
• Taints and tolerations

Topology spread constraints

The cluster's topologySpreadConstraints section enables additional control of scheduling pods across topologies, enhancing what affinity and anti-affinity can offer.

Command-line interface

CloudNativePG doesn't have its own command-line interface. It relies on the best command-line interface for Kubernetes, kubectl, by providing a plugin called cnpg. This plugin enhances and simplifies your PostgreSQL cluster management experience.

Current status of the cluster

The operator continuously updates the status section of the CR with the observed status of the cluster. The entire PostgreSQL cluster status is continuously monitored by the instance manager running in each pod. The instance manager is responsible for applying the required changes to the controlled PostgreSQL instance to converge to the required status of the cluster. (For example, if the cluster status reports that pod -1 is the primary, pod -1 needs to promote itself while the other pods need to follow pod -1.) The same status is used by the cnpg plugin for kubectl to provide details.

Operator's certification authority

The operator creates a certification authority for itself. It creates and signs with the operator certification authority a leaf certificate for the webhook server to use. This certificate ensures safe communication between the Kubernetes API server and the operator.

Cluster's certification authority

The operator creates a certification authority for every PostgreSQL cluster. This certification authority is used to issue and renew TLS certificates for clients' authentication, including streaming replication standby servers (instead of passwords). Support for a custom certification authority for client certificates is available through secrets, which also includes integration with cert-manager. Certificates can be issued with the cnpg plugin for kubectl.

TLS connections

The operator transparently and natively supports TLS/SSL connections to encrypt client/server communications for increased security using the cluster's certification authority. Support for custom server certificates is available through secrets, which also includes integration with cert-manager.

Certificate authentication for streaming replication

To authorize streaming replication connections from the standby servers, the operator relies on TLS client certificate authentication. This method is used instead of relying on a password (and therefore a secret).

Continuous configuration management

The operator enables you to apply changes to the Cluster resource YAML section of the PostgreSQL configuration. Depending on the configuration option, it also makes sure that all instances are properly reloaded or restarted.

Import of existing PostgreSQL databases

The operator provides a declarative way to import existing Postgres databases in a new CloudNativePG cluster in Kubernetes, using offline migrations. The same feature also covers offline major upgrades of PostgreSQL databases. Offline means that applications must stop their write operations at the source until the database is imported. The feature extends the initdb bootstrap method to create a new PostgreSQL cluster using a logical snapshot of the data available in another PostgreSQL database. This data can be accessed by way of the network through a superuser connection. Import is from any supported version of Postgres. It relies on pg_dump and pg_restore being executed from the new cluster primary for all databases that are part of the operation and, if requested, for roles.

PostGIS clusters

CloudNativePG supports the installation of clusters with the PostGIS open source extension for geographical databases. This extension is one of the most popular extensions for PostgreSQL.

Basic LDAP authentication for PostgreSQL

The operator allows you to configure LDAP authentication for your PostgreSQL clients, using either the simple bind or search+bind mode, as described in the LDAP authentication section of the PostgreSQL documentation.

Multiple installation methods

The operator can be installed through a Kubernetes manifest by way of kubectl apply, to be used in a traditional Kubernetes installation in public and private cloud environments. CloudNativePG also supports installation by way of a Helm chart or OLM bundle from OperatorHub.io.

Convention over configuration

The operator supports the convention-over-configuration paradigm, deciding standard default values while allowing you to override them and customize them. You can specify a deployment of a PostgreSQL cluster using the Cluster CRD in a couple of lines of YAML code.

Level 2: Seamless upgrades

Capability level 2 is about enabling updates of the operator and the actual workload, in this case PostgreSQL servers. This includes PostgreSQL minor release updates (security and bug fixes normally) as well as major online upgrades.

Upgrade of the operator

You can upgrade the operator seamlessly as a new deployment. Because of the instance manager's injection, a change in the operator doesn't require a change in the operand. The operator can manage older versions of the operand.

CloudNativePG also supports in-place updates of the instance manager following an upgrade of the operator. In-place updates don't require a rolling update (and subsequent switchover) of the cluster.

Upgrade of the managed workload

The operand can be upgraded using a declarative configuration approach as part of changing the CR and, in particular, the imageName parameter. The operator prevents major upgrades of PostgreSQL while making it possible to go in both directions in terms of minor PostgreSQL releases within a major version, enabling updates and rollbacks.

In the presence of standby servers, the operator performs rolling updates starting from the replicas. It does this by dropping the existing pod and creating a new one with the new requested operand image that reuses the underlying storage. Depending on the value of the primaryUpdateStrategy, the operator proceeds with a switchover before updating the former primary (unsupervised). Or, it waits for the user to manually issue the switchover procedure (supervised) by way of the cnpg plugin for kubectl. The setting to use depends on the business requirements, as the operation might generate some downtime for the applications. This downtime can range from a few seconds to minutes, based on the actual database workload.

Display cluster availability status during upgrade

At any time, convey the cluster's high availability status, for example, Setting up primary, Creating a new replica, Cluster in healthy state, Switchover in progress, Failing over, and Upgrading cluster.

Level 3: Full lifecycle

Capability level 3 requires the operator to manage aspects of business continuity and scalability.

Disaster recovery is a business continuity component that requires that both backup and recovery of a database work correctly. While as a starting point, the goal is to achieve RPO < 5 minutes, the long-term goal is to implement RPO=0 backup solutions. High availability is the other important component of business continuity. Through PostgreSQL native physical replication and hot standby replicas, it allows the operator to perform failover and switchover operations. This area includes enhancements in:

• Control of PostgreSQL physical replication, such as synchronous replication, (cascading) replication clusters, and so on
• Connection pooling, to improve performance and control through a connection pooling layer with pgBouncer

PostgreSQL WAL archive

The operator supports PostgreSQL continuous archiving of WAL files to an object store (AWS S3 and S3-compatible, Azure Blob Storage, Google Cloud Storage, and gateways like MinIO).

WAL archiving is defined at the cluster level, declaratively, through the backup parameter in the cluster definition. This is done by specifying an S3 protocol destination URL (for example, to point to a specific folder in an AWS S3 bucket) and, optionally, a generic endpoint URL.

WAL archiving, a prerequisite for continuous backup, doesn't require any further user action. The operator transparently sets the archive_command to rely on barman-cloud-wal-archive to ship WAL files to the defined endpoint. You can decide the compression algorithm, as well as the number of parallel jobs to concurrently upload WAL files in the archive. In addition, Instance Manager checks the correctness of the archive destination by performing the barman-cloud-check-wal-archive command before beginning to ship the first set of WAL files.

PostgreSQL backups

The operator was designed to provide application-level backups using PostgreSQL’s native continuous hot backup technology based on physical base backups and continuous WAL archiving. Base backups can be saved on:

• Kubernetes volume snapshots
• Object stores (AWS S3 and S3-compatible, Azure Blob Storage, Google Cloud Storage, and gateways like MinIO)

Base backups are defined at the cluster level, declaratively, through the backup parameter in the cluster definition.

You can define base backups in two ways:

• On-demand, through the Backup custom resource definition
• Scheduled, through the ScheduledBackupcustom resource definition, using a cron-like syntax

Volume snapshots rely directly on the Kubernetes API, which delegates this capability to the underlying storage classes and CSI drivers. Volume snapshot backups are suitable for very large database (VLDB) contexts.

Object store backups rely on barman-cloud-backup for the job (distributed as part of the application container image) to relay backups in the same endpoint, alongside WAL files.

Both barman-cloud-wal-restore and barman-cloud-backup are distributed in the application container image under GNU GPL 3 terms.

Object store backups and volume snapshot backups are taken while PostgreSQL is up and running (hot backups). Volume snapshots also support taking consistent database snapshots with cold backups.

Backups from a standby

The operator supports offloading base backups onto a standby without impacting the RPO of the database. This allows resources to be preserved on the primary, in particular I/O, for standard database operations.

Full restore from a backup

The operator enables you to bootstrap a new cluster (with its settings) starting from an existing and accessible backup, either on a volume snapshot or in an object store.

Once the bootstrap process is completed, the operator initiates the instance in recovery mode. It replays all available WAL files from the specified archive, exiting recovery and starting as a primary. Subsequently, the operator clones the requested number of standby instances from the primary. CloudNativePG supports parallel WAL fetching from the archive.

Point-in-time recovery (PITR) from a backup

The operator enables you to create a new PostgreSQL cluster by recovering an existing backup to a specific point in time, defined with a timestamp, a label, or a transaction ID. This capability is built on top of the full restore one and supports all the options available in PostgreSQL for PITR.

Zero-data-loss clusters through synchronous replication

Achieve zero data loss (RPO=0) in your local high-availability CloudNativePG cluster through quorum-based synchronous replication support. The operator provides two configuration options that control the minimum and maximum number of expected synchronous standby replicas available at any time. The operator reacts accordingly, based on the number of available and ready PostgreSQL instances in the cluster. It uses the following formula for the quorum (q):

1 <= minSyncReplicas <= q <= maxSyncReplicas <= readyReplicas

Replica clusters

Establish a robust cross-Kubernetes cluster topology for PostgreSQL clusters, harnessing the power of native streaming and cascading replication. With the replica option, you can configure an autonomous cluster to consistently replicate data from another PostgreSQL source of the same major version. This source can be located anywhere, provided you have access to a WAL archive for fetching WAL files or a direct streaming connection via TLS between the two endpoints.

Notably, the source PostgreSQL instance can exist outside the Kubernetes environment, whether in a physical or virtual setting.

Replica clusters can be instantiated through various methods, including volume snapshots, a recovery object store (utilizing the Barman Cloud backup format), or streaming using pg_basebackup. Both WAL file shipping and WAL streaming are supported. The deployment of replica clusters significantly elevates the business continuity posture of PostgreSQL databases within Kubernetes, extending across multiple data centers and facilitating hybrid and multi-cloud setups. (While anticipating Kubernetes federation native capabilities, manual switchover across data centers remains necessary.)

Additionally, the flexibility extends to creating delayed replica clusters intentionally lagging behind the primary cluster. This intentional lag aims to minimize the Recovery Time Objective (RTO) in the event of unintended errors, such as incorrect DELETE or UPDATE SQL operations.

Liveness and readiness probes

The operator defines liveness and readiness probes for the Postgres containers that are then invoked by the kubelet. They're mapped respectively to the /healthz and /readyz endpoints of the web server managed directly by the instance manager.

The liveness probe is based on the pg_isready executable, and the pod is considered healthy with exit codes 0 (server accepting connections normally) and 1 (server is rejecting connections, for example, during startup). The readiness probe issues a simple query (;) to verify that the server is ready to accept connections.

Rolling deployments

The operator supports rolling deployments to minimize the downtime. If a PostgreSQL cluster is exposed publicly, the service load-balances the read-only traffic only to available pods during the initialization or the update.

Scale up and down of replicas

The operator allows you to scale up and down the number of instances in a PostgreSQL cluster. New replicas are started up from the primary server and participate in the cluster's HA infrastructure. The CRD declares a "scale" subresource that allows you to use the kubectl scale command.

Maintenance window and PodDisruptionBudget for Kubernetes nodes

The operator creates a PodDisruptionBudget resource to limit the number of concurrent disruptions to one primary instance. This configuration prevents the maintenance operation from deleting all the pods in a cluster, allowing the specified number of instances to be created. The PodDisruptionBudget is applied during the node-draining operation, preventing any disruption of the cluster service.

While this strategy is correct for Kubernetes clusters where storage is shared among all the worker nodes, it might not be the best solution for clusters using local storage or for clusters installed in a private cloud. The operator allows you to specify a maintenance window and configure the reaction to any underlying node eviction. The ReusePVC option in the maintenance window section enables to specify the strategy to use. Allocate new storage in a different PVC for the evicted instance, or wait for the underlying node to be available again.

Fencing

Fencing is the process of protecting the data in one, more, or even all instances of a PostgreSQL cluster when they appear to be malfunctioning. When an instance is fenced, the PostgreSQL server process is guaranteed to be shut down, while the pod is kept running. This ensures that, until the fence is lifted, data on the pod isn't modified by PostgreSQL and that you can investigate file system for debugging and troubleshooting purposes.

Hibernation (declarative)

CloudNativePG supports hibernation of a running PostgreSQL cluster in a declarative manner, through the cnpg.io/hibernation annotation. Hibernation enables saving CPU power by removing the database pods while keeping the database PVCs. This feature simulates scaling to 0 instances.

Hibernation (imperative)

CloudNativePG supports hibernation of a running PostgreSQL cluster by way of the cnpg plugin. Hibernation shuts down all Postgres instances in the high-availability cluster and keeps a static copy of the PVC group of the primary. The copy contains PGDATA and WALs. The plugin enables you to exit the hibernation phase by resuming the primary and then recreating all the replicas, if they exist.

Reuse of persistent volumes storage in pods

When the operator needs to create a pod that was deleted by the user or was evicted by a Kubernetes maintenance operation, it reuses the PersistentVolumeClaim, if available. This ability avoids the need to clone the data from the primary again.

CPU and memory requests and limits

The operator allows administrators to control and manage resource usage by the cluster's pods in the resources section of the manifest. In particular, you can set requests and limits values for both CPU and RAM.

Connection pooling with PgBouncer

CloudNativePG provides native support for connection pooling with PgBouncer, one of the most popular open source connection poolers for PostgreSQL. From an architectural point of view, the native implementation of a PgBouncer connection pooler introduces a new layer to access the database. This optimizes the query flow toward the instances and makes the use of the underlying PostgreSQL resources more efficient. Instead of connecting directly to a PostgreSQL service, applications can now connect to the PgBouncer service and start reusing any existing connection.

Level 4: Deep insights

Capability level 4 is about observability: monitoring, alerting, trending, and log processing. This might involve the use of external tools, such as Prometheus, Grafana, and Fluent Bit, as well as extensions in the PostgreSQL engine for the output of error logs directly in JSON format.

CloudNativePG was designed to provide everything needed to easily integrate with industry-standard and community-accepted tools for flexible monitoring and logging.

Prometheus exporter with configurable queries

The instance manager provides a pluggable framework. By way of its own web server listening on the metrics port (9187), it exposes an endpoint to export metrics for the Prometheus monitoring and alerting tool. The operator supports custom monitoring queries defined as ConfigMap or Secret objects using a syntax that's compatible with postgres_exporter for Prometheus. CloudNativePG provides a set of basic monitoring queries for PostgreSQL that can be integrated and adapted to your context.

Grafana dashboard

CloudNativePG comes with a Grafana dashboard that you can use as a base to monitor all critical aspects of a PostgreSQL cluster, and customize.

Standard output logging of PostgreSQL error messages in JSON format

Every log message is delivered to standard output in JSON format. The first level is the definition of the timestamp, the log level, and the type of log entry, such as postgres for the canonical PostgreSQL error message channel. As a result, every pod managed by CloudNativePG can be easily and directly integrated with any downstream log processing stack that supports JSON as source data type.

Real-time query monitoring

CloudNativePG transparently and natively supports:

• The essential pg_stat_statements extension, which enables tracking of planning and execution statistics of all SQL statements executed by a PostgreSQL server
• The auto_explain extension, which provides a means for logging execution plans of slow statements automatically, without having to manually run EXPLAIN (helpful for tracking down un-optimized queries)
• The pg_failover_slots extension, which makes logical replication slots usable across a physical failover, ensuring resilience in change data capture (CDC) contexts based on PostgreSQL's native logical replication

Audit

CloudNativePG allows database and security administrators, auditors, and operators to track and analyze database activities using PGAudit for PostgreSQL. Such activities flow directly in the JSON log and can be properly routed to the correct downstream target using common log brokers like Fluentd.

Kubernetes events

Record major events as expected by the Kubernetes API, such as creating resources, removing nodes, and upgrading. Events can be displayed by using the kubectl describe and kubectl get events commands.

Level 5: Auto pilot

Capability level 5 is focused on automated scaling, healing, and tuning through the discovery of anomalies and insights that emerged from the observability layer.

Automated failover for self-healing

In case of detected failure on the primary, the operator changes the status of the cluster by setting the most aligned replica as the new target primary. As a consequence, the instance manager in each alive pod initiates the required procedures to align itself with the requested status of the cluster. It does this by either becoming the new primary or by following it. In case the former primary comes back up, the same mechanism avoids a split-brain by preventing applications from reaching it, running pg_rewind on the server and restarting it as a standby.

Automated recreation of a standby

If the pod hosting a standby is removed, the operator initiates the procedure to re-create a standby server.