Replication

Physical replication is one of the strengths of PostgreSQL and one of the reasons why some of the largest organizations in the world have chosen it for the management of their data in business continuity contexts. Primarily used to achieve high availability, physical replication also allows scale-out of read-only workloads and offloading of some work from the primary.

Application-level replication

Having contributed throughout the years to the replication feature in PostgreSQL, we have decided to build high availability in CloudNativePG on top of the native physical replication technology, and integrate it directly in the Kubernetes API.

In Kubernetes terms, this is referred to as application-level replication, in contrast with storage-level replication.

A very mature technology

PostgreSQL has a very robust and mature native framework for replicating data from the primary instance to one or more replicas, built around the concept of transactional changes continuously stored in the WAL (Write Ahead Log).

Started as the evolution of crash recovery and point in time recovery technologies, physical replication was first introduced in PostgreSQL 8.2 (2006) through WAL shipping from the primary to a warm standby in continuous recovery.

PostgreSQL 9.0 (2010) enhanced it with WAL streaming and read-only replicas via hot standby, while 9.1 (2011) introduced synchronous replication at the transaction level (for RPO=0 clusters). Cascading replication was released with PostgreSQL 9.2 (2012). The foundations of logical replication were laid in PostgreSQL 9.4, while version 10 (2017) introduced native support for the publisher/subscriber pattern to replicate data from an origin to a destination.

Streaming replication support

At the moment, CloudNativePG natively and transparently manages physical streaming replicas within a cluster in a declarative way, based on the number of provided instances in the spec:

replicas = instances - 1 (where  instances > 0)

Immediately after the initialization of a cluster, the operator creates a user called streaming_replica as follows:

CREATE USER streaming_replica WITH REPLICATION;
   -- NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOBYPASSRLS

Out of the box, the operator automatically sets up streaming replication within the cluster over an encrypted channel and enforces TLS client certificate authentication for the streaming_replica user - as highlighted by the following excerpt taken from pg_hba.conf:

# Require client certificate authentication for the streaming_replica user
hostssl postgres streaming_replica all cert
hostssl replication streaming_replica all cert

If configured, the operator manages replication slots for all the replicas in the HA cluster, ensuring that WAL files required by each standby are retained on the primary's storage, even after a failover or switchover.

Continuous backup integration

In case continuous backup is configured in the cluster, CloudNativePG transparently configures replicas to take advantage of restore_command when in continuous recovery. As a result, PostgreSQL can use the WAL archive as a fallback option whenever pulling WALs via streaming replication fails.

Synchronous replication

CloudNativePG supports the configuration of quorum-based synchronous streaming replication via two configuration options called minSyncReplicas and maxSyncReplicas, which are the minimum and the maximum number of expected synchronous standby replicas available at any time. For self-healing purposes, the operator always compares these two values with the number of available replicas to determine the quorum.

Synchronous replication is disabled by default (minSyncReplicas and maxSyncReplicas are not defined). In case both minSyncReplicas and maxSyncReplicas are set, CloudNativePG automatically updates the synchronous_standby_names option in PostgreSQL to the following value:

ANY q (pod1, pod2, ...)

Where:

• q is an integer automatically calculated by the operator to be:
  1 <= minSyncReplicas <= q <= maxSyncReplicas <= readyReplicas
• pod1, pod2, ... is the list of all PostgreSQL pods in the cluster

As stated in the PostgreSQL documentation, the method ANY specifies a quorum-based synchronous replication and makes transaction commits wait until their WAL records are replicated to at least the requested number of synchronous standbys in the list.

Select nodes for synchronous replication

CloudNativePG enables you to select which PostgreSQL instances are eligible to participate in a quorum-based synchronous replication set through anti-affinity rules based on the node labels where the PVC holding the PGDATA and the Postgres pod are.

As an example use-case for this feature: in a cluster with a single sync replica, we would be able to ensure the sync replica will be in a different availability zone from the primary instance, usually identified by the topology.kubernetes.io/zone label on a node. This would increase the robustness of the cluster in case of an outage in a single availability zone, especially in terms of recovery point objective (RPO).

The idea of anti-affinity is to ensure that sync replicas that participate in the quorum are chosen from pods running on nodes that have different values for the selected labels (in this case, the availability zone label) then the node where the primary is currently in execution. If no node matches such criteria, the replicas are eligible for synchronous replication.

The example below shows how this can be done through the syncReplicaElectionConstraint section within .spec.postgresql. nodeLabelsAntiAffinity allows you to specify those node labels that need to be evaluated to make sure that synchronous replication will be dynamically configured by the operator between the current primary and the replicas which are located on nodes having a value of the availability zone label different from that of the node where the primary is:

spec:
  instances: 3
  postgresql:
    syncReplicaElectionConstraint:
      enabled: true
      nodeLabelsAntiAffinity:
      - topology.kubernetes.io/zone

As you can imagine, the availability zone is just an example, but you could customize this behavior based on other labels that describe the node, such as storage, CPU, or memory.

Replication slots

Replication slots are a native PostgreSQL feature introduced in 9.4 that provides an automated way to ensure that the primary does not remove WAL segments until all the attached streaming replication clients have received them, and that the primary does not remove rows which could cause a recovery conflict even when the standby is (temporarily) disconnected.

A replication slot exists solely on the instance that created it, and PostgreSQL does not replicate it on the standby servers. As a result, after a failover or a switchover, the new primary does not contain the replication slot from the old primary. This can create problems for the streaming replication clients that were connected to the old primary and have lost their slot.

CloudNativePG provides a turn-key solution to synchronize the content of physical replication slots from the primary to each standby, addressing two use cases:

• the replication slots automatically created for the High Availability of the Postgres cluster (see "Replication slots for High Availability" below for details)
• user-defined replication slots created on the primary

Replication slots for High Availability

CloudNativePG fills this gap by introducing the concept of cluster-managed replication slots, starting with high availability clusters. This feature automatically manages physical replication slots for each hot standby replica in the High Availability cluster, both in the primary and the standby.

In CloudNativePG, we use the terms:

• Primary HA slot: a physical replication slot whose lifecycle is entirely managed by the current primary of the cluster and whose purpose is to map to a specific standby in streaming replication. Such a slot lives on the primary only.
• Standby HA slot: a physical replication slot for a standby whose lifecycle is entirely managed by another standby in the cluster, based on the content of the pg_replication_slots view in the primary, and updated at regular intervals using pg_replication_slot_advance().

This feature is enabled by default and can be disabled via configuration. For details, please refer to the "replicationSlots" section in the API reference. Here follows a brief description of the main options:

.spec.replicationSlots.highAvailability.enabled

if true, the feature is enabled (true is the default since 1.21)

.spec.replicationSlots.highAvailability.slotPrefix

the prefix that identifies replication slots managed by the operator for this feature (default: _cnpg_)

.spec.replicationSlots.updateInterval

how often the standby synchronizes the position of the local copy of the replication slots with the position on the current primary, expressed in seconds (default: 30)

Although it is not recommended, if you desire a different behavior, you can customize the above options.

For example, the following manifest will create a cluster with replication slots disabled.

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: cluster-example
spec:
  instances: 3
  # Disable replication slots for HA in the cluster
  replicationSlots:
    highAvailability:
      enabled: false

  storage:
    size: 1Gi

User-Defined Replication slots

Although CloudNativePG doesn't support a way to declaratively define physical replication slots, you can still create your own slots via SQL.

CloudNativePG can manage the synchronization of any user managed physical replication slots between the primary and standbys, similarly to what it does for the HA replication slots explained above (the only difference is that you need to create the replication slot).

This feature is enabled by default (meaning that any replication slot is synchronized), but you can disable it or further customize its behavior (for example by excluding some slots using regular expressions) through the synchronizeReplicas stanza. For example:

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: cluster-example
spec:
  instances: 3
  replicationSlots:
    synchronizeReplicas:
      enabled: true
      excludePatterns:
      - "^foo"

For details, please refer to the "replicationSlots" section in the API reference. Here follows a brief description of the main options:

.spec.replicationSlots.synchronizeReplicas.enabled

When true or not specified, every user-defined replication slot on the primary is synchronized on each standby. If changed to false, the operator will remove any replication slot previously created by itself on each standby.

.spec.replicationSlots.synchronizeReplicas.excludePatterns

A list of regular expression patterns to match the names of user-defined replication slots to be excluded from synchronization. This can be useful to exclude specific slots based on naming conventions.

Synchronization frequency

You can also control the frequency with which a standby queries the pg_replication_slots view on the primary, and updates its local copy of the replication slots, like in this example:

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: cluster-example
spec:
  instances: 3
  # Reduce the frequency of standby HA slots updates to once every 5 minutes
  replicationSlots:
    updateInterval: 300

  storage:
    size: 1Gi

Capping the WAL size retained for replication slots

When replication slots is enabled, you might end up running out of disk space due to PostgreSQL trying to retain WAL files requested by a replication slot. This might happen due to a standby that is (temporarily?) down, or lagging, or simply an orphan replication slot.

Starting with PostgreSQL 13, you can take advantage of the max_slot_wal_keep_size configuration option controlling the maximum size of WAL files that replication slots are allowed to retain in the pg_wal directory at checkpoint time. By default, in PostgreSQL max_slot_wal_keep_size is set to -1, meaning that replication slots may retain an unlimited amount of WAL files. As a result, our recommendation is to explicitly set max_slot_wal_keep_size when replication slots support is enabled. For example:

  # ...
  postgresql:
    parameters:
      max_slot_wal_keep_size: "10GB"
  # ...

Monitoring replication slots

Replication slots must be carefully monitored in your infrastructure. By default, we provide the pg_replication_slots metric in our Prometheus exporter with key information such as the name of the slot, the type, whether it is active, the lag from the primary.