PostgreSQL® - Concepts - High availability and failure scenarios

This page provides the main concepts and mechanisms to achieve high availability for our PostgreSQL® Cloud Databases. We then describe multiple failure scenarios and consequences.

High availability

Cloud Databases for PostgreSQL® are available on three service plans, offering different levels of high availability. The selected plan defines the features available. A summary is provided in the table below:

About primary and replica nodes

A node is a virtual machine with dedicated resources (CPU, RAM, storage).

The primary node, also known as the master node, acts as the leader of your cluster. It can read and write to your database. There is only one primary node per PostgreSQL® cluster.

Replica nodes are optional and maintain up-to-date copies of the primary node data. Replica nodes do not allow write operations, only reads.

Primary and replica nodes are interchangeable: When the primary node encounters an issue, the system triggers an automatic failover. A replica node is promoted and becomes the new leader of the cluster. From that point on, it carries out the duties of the primary node (e.g., writes, data replication source). The Service URI remains the same, and the internal IP address changes to point to the new primary node.

Having at least one replica node in a PostgreSQL® cluster brings a lot of benefits:

• Data resilience: your data gets replicated to another physical resource, carrying "hot data".
• High availability: in case of failure of the primary node, the time to restore is drastically reduced. Thanks to the automatic failover to a replica node, your cluster and data are up and ready again in a few minutes.
• Performance: for read-only operations, you can connect to replica nodes, allowing you to reduce the load on the primary server.

Having at least two replica nodes is important so that even during recovery from (single-node) failures, there are always two copies of the data on two different nodes. If another failure strikes after a failover when there is just a single master node running, we again risk losing some of the latest changes written to the database. It takes time to rebuild a new standby node and get it in sync node after a failover when there is a lot of data in the database, and it often makes sense to protect the data over this period by having another replica. This is especially important when the database size is large, and recreating a replacement node for the faulted one can take hours.

We can represent a cluster like this:

About backups and restores

OVHcloud uses the popular and open-source backup daemon PGHoard for backups and restores. It makes real-time copies of PostgreSQL® Write Ahead Logs (WAL) files and encrypts and compresses the files. Backup files are stored in OVHcloud Object storage containers.

You can find more information on their official page.

Failure scenarios

We distinguish between two categories of failures:

• Minor failures, such as service process crashes or temporary loss of network access, are handled automatically by OVHcloud in all plans without any major changes to the service deployment.
• Severe failures, such as losing a node entirely in case of hardware or critical software problems. OVHcloud monitoring infrastructure detects a failed node automatically when the node starts reporting issues in self-diagnostics or when it stops communicating. In such cases, the monitoring infrastructure automatically schedules the creation of a new node to replace the faulty one.

The impact of different failure scenarios depends on the selected service plan since the number of nodes is related to it. Browse the right section according to your offer.

Comparative table

The table below summarizes the scenarios detailed in the following paragraphs.

RPO is the Recovery Point Objective, meaning how far back before the incident you can recover the data. RTO is the Recovery Time Objective, meaning how long it takes to go back to a normal situation.

Scenarios for Essential service plans with a single-node

Essential plans provide a single node: there is no replica.

Node failure

OVHcloud detects when the node is lost. It immediately starts creating a new replacement node. Once the new node is up, it restores its state to the latest available backup and then resumes its operations.

Since there is only one node, during the whole recovery process, your databases are unavailable. You cannot read them or write on them. Also, some write operations data may have been lost since we restored from the last backup and the latest backed-up PostgreSQL® Write Ahead Logs (WAL) files. Typically, this time window is limited to either five minutes of time or one WAL file.

The duration of the recovery process depends on the amount of data stored, the region, and the network bandwidth. OVHcloud cannot guarantee a maximum recovery time, and you can expect a few hours.

No intervention is required from the customer.

Datacenter failure

OVHcloud also detects when a whole data center is lost. As a user, your cluster will still be visible in the API and Control Panel.

Automatic backups are first performed on-site (i.e., in the same region as the service), then replicated off-site to another data center.

In case of a data center failure, the recovery strategy involves creating a new service in another region, either duplicating it from an automatic backup (creating a fork) that was replicated off-site or restoring from backups performed manually (e.g., pg_restore a previously performed pg_dump). You will then need to configure the newly created service (e.g., adding users, IP restrictions, etc.) and reconfigure applications consuming the database service to use the new service's URIs and credentials. In both cases, the Recovery Time Objective will vary, depending on multiple parameters such as the amount of data to restore, network congestion, and latency or node performance. It can take from a few hours to a few days.

Scenarios for Highly available Business and Enterprise service plans

The Business plan provides a primary and a replica node; the Enterprise plan starts with a primary and two replica nodes.

Replica node failure

When the failed node is a PostgreSQL® replica node, the primary node keeps running normally and provides a normal service level to client applications. However, applications that read data directly from the defective replica node may experience degraded service.

OVHcloud automatically detects the loss of a replica node and waits 300 seconds before creating a new replica node. This timeout allows OVHcloud to make sure that the node is persistently faulty (and that a network failure is not involved, for example). Once the new replacement replica node is ready and synchronized with the primary node, it starts replicating the primary node in real time as the situation reverts to normal.

If both PostgreSQL® replica nodes have a failure, the same process happens, but this time twice.

Although OVHcloud cannot commit to a set duration of the whole process, you can expect zero downtime since the primary is up and a few hours to get the new replacement nodes up and ready (allowing time to restore the data).

No intervention is required from the customer.

Primary node failure

When the failed node is a PostgreSQL® primary node, the combined information from the OVHcloud monitoring infrastructure and the replica node is used to decide on an automatic failover.

After a 60-second timeout, the replica node is then promoted to the primary role and immediately starts serving client applications. A new replacement node is automatically scheduled and becomes the new replica node.

All-node failure

If all the primary and replica nodes fail at the same time, the creation of new nodes to become the new primary and replicas is automatically scheduled. The primary node is restored from the latest available backup. This could involve some data loss because any write operation since the backup of the latest WAL file is lost. Typically, this time window is limited to either five minutes of time or one WAL file.

The duration of the recovery process depends on the amount of data stored, the region, and the network bandwidth. OVHcloud cannot guarantee a maximum recovery time, but you can expect a few hours. During this period, your service remains unavailable.

No intervention is required from the customer.

Datacenter failure

OVHcloud also detects when a whole data center is lost. As a user, your cluster will still be visible in the API and Control Panel.

Automatic backups are first performed on-site (i.e., in the same region as the service), then replicated off-site to another data center.

In case of a data center failure, the recovery strategy involves creating a new service in another region, either duplicating it from an automatic backup (creating a fork) that was replicated off-site or restoring from backups performed manually (e.g., pg_restore a previously performed pg_dump). You will then need to configure the newly created service (e.g., adding users, IP restrictions, etc.) and reconfigure applications consuming the database service to use the new service's URIs and credentials. In both cases, the Recovery Time Objective will vary, depending on multiple parameters such as the amount of data to restore, network congestion, and latency or node performance. It can take from a few hours to a few days.

Go further

For more information and tutorials, please see our other Managed Databases & Analytics or Platform as a Service guides. You can also explore the guides for other OVHcloud products and services.

OVHcloud Managed Databases and Analytics:

• Grafana® is a registered trademark of Grafana Labs and is used with the permission of Grafana Labs. OVH SAS and its subsidiaries are not affiliated with or endorsed by Grafana Labs.
• Kafka® is a registered trademark of The Apache Software Foundation and has been licensed for use by OVHcloud, who has no affiliation with and is not endorsed by The Apache Software Foundation.
• MongoDB® is a registered trademark of MongoDB, Inc.
• MySQL® is a registered trademark of Oracle and/or its affiliates.
• PostgreSQL® is a registered trademark of the PostgreSQL Community Association of Canada.