Backup
PostgreSQL natively provides first class backup and recovery capabilities based on file system level (physical) copy. These have been successfully used for more than 15 years in mission critical production databases, helping organizations all over the world achieve their disaster recovery goals with Postgres.
Note
There's another way to backup databases in PostgreSQL, through the
pg_dump
utility - which relies on logical backups instead of physical ones.
However, logical backups are not suitable for business continuity use cases
and as such are not covered by CloudNativePG (yet, at least).
If you want to use the pg_dump
utility, let yourself be inspired by the
"Troubleshooting / Emergency backup" section.
In CloudNativePG, the backup infrastructure for each PostgreSQL cluster is made up of the following resources:
- WAL archive: a location containing the WAL files (transactional logs) that are continuously written by Postgres and archived for data durability
- Physical base backups: a copy of all the files that PostgreSQL uses to
store the data in the database (primarily the
PGDATA
and any tablespace)
The WAL archive can only be stored on object stores at the moment.
On the other hand, CloudNativePG supports two ways to store physical base backups:
- on object stores, as tarballs - optionally compressed
- on Kubernetes Volume Snapshots, if supported by the underlying storage class
Important
Before choosing your backup strategy with CloudNativePG, it is important that you take some time to familiarize with some basic concepts, like WAL archive, hot and cold backups.
Important
Please refer to the official Kubernetes documentation for a list of all the supported Container Storage Interface (CSI) drivers that provide snapshotting capabilities.
WAL archive
The WAL archive in PostgreSQL is at the heart of continuous backup, and it is fundamental for the following reasons:
- Hot backups: the possibility to take physical base backups from any instance in the Postgres cluster (either primary or standby) without shutting down the server; they are also known as online backups
- Point in Time recovery (PITR): to possibility to recover at any point in time from the first available base backup in your system
Warning
WAL archive alone is useless. Without a physical base backup, you cannot restore a PostgreSQL cluster.
In general, the presence of a WAL archive enhances the resilience of a PostgreSQL cluster, allowing each instance to fetch any required WAL file from the archive if needed (normally the WAL archive has higher retention periods than any Postgres instance that normally recycles those files).
This use case can also be extended to replica clusters, as they can simply rely on the WAL archive to synchronize across long distances, extending disaster recovery goals across different regions.
When you configure a WAL archive, CloudNativePG provides out-of-the-box an RPO <= 5 minutes for disaster recovery, even across regions.
Important
Our recommendation is to always setup the WAL archive in production. There are known use cases - normally involving staging and development environments - where none of the above benefits are needed and the WAL archive is not necessary. RPO in this case can be any value, such as 24 hours (daily backups) or infinite (no backup at all).
Cold and Hot backups
Hot backups have already been defined in the previous section. They require the presence of a WAL archive and they are the norm in any modern database management system.
Cold backups, also known as offline backups, are instead physical base backups taken when the PostgreSQL instance (standby or primary) is shut down. They are consistent per definition and they represent a snapshot of the database at the time it was shut down.
As a result, PostgreSQL instances can be restarted from a cold backup without the need of a WAL archive, even though they can take advantage of it, if available (with all the benefits on the recovery side highlighted in the previous section).
In those situations with a higher RPO (for example, 1 hour or 24 hours), and shorter retention periods, cold backups represent a viable option to be considered for your disaster recovery plans.
Object stores or volume snapshots: which one to use?
In CloudNativePG, object store based backups:
- always require the WAL archive
- support hot backup only
- don't support incremental copy
- don't support differential copy
VolumeSnapshots instead:
- don't require the WAL archive, although in production it is always recommended
- support incremental copy, depending on the underlying storage classes
- support differential copy, depending on the underlying storage classes
- also support cold backup
Which one to use depends on your specific requirements and environment, including:
- availability of a viable object store solution in your Kubernetes cluster
- availability of a trusted storage class that supports volume snapshots
- size of the database: with object stores, the larger your database, the longer backup and, most importantly, recovery procedures take (the latter impacts RTO); in presence of Very Large Databases (VLDB), the general advice is to rely on Volume Snapshots as, thanks to copy-on-write, they provide faster recovery
- data mobility and possibility to store or relay backup files on a secondary location in a different region, or any subsequent one
- other factors, mostly based on the confidence and familiarity with the underlying storage solutions
The summary table below highlights some of the main differences between the two available methods for storing physical base backups.
Object store | Volume Snapshots | |
---|---|---|
WAL archiving | Required | Recommended (1) |
Cold backup | ✗ | ✓ |
Hot backup | ✓ | ✓ |
Incremental copy | ✗ | ✓ (2) |
Differential copy | ✗ | ✓ (2) |
Backup from a standby | ✓ | ✓ |
Snapshot recovery | ✗ (3) | ✓ |
Point In Time Recovery (PITR) | ✓ | Requires WAL archive |
Underlying technology | Barman Cloud | Kubernetes API |
See the explanation below for the notes in the above table:
- WAL archive must be on an object store at the moment
- If supported by the underlying storage classes of the PostgreSQL volumes
- Snapshot recovery can be emulated using the
bootstrap.recovery.recoveryTarget.targetImmediate
option
Scheduled backups
Scheduled backups are the recommended way to configure your backup strategy in
CloudNativePG. They are managed by the ScheduledBackup
resource.
Info
Please refer to ScheduledBackupSpec
in the API reference for a full list of options.
The schedule
field allows you to define a six-term cron schedule specification,
which includes seconds, as expressed in
the Go cron
package format.
Warning
Beware that this format accepts also the seconds
field, and it is
different from the crontab
format in Unix/Linux systems.
This is an example of a scheduled backup:
apiVersion: postgresql.cnpg.io/v1
kind: ScheduledBackup
metadata:
name: backup-example
spec:
schedule: "0 0 0 * * *"
backupOwnerReference: self
cluster:
name: pg-backup
The above example will schedule a backup every day at midnight because the schedule specifies zero for the second, minute, and hour, while specifying wildcard, meaning all, for day of the month, month, and day of the week.
In Kubernetes CronJobs, the equivalent expression is 0 0 * * *
because seconds
are not included.
Hint
Backup frequency might impact your recovery time object (RTO) after a disaster which requires a full or Point-In-Time recovery operation. Our advice is that you regularly test your backups by recovering them, and then measuring the time it takes to recover from scratch so that you can refine your RTO predictability. Recovery time is influenced by the size of the base backup and the amount of WAL files that need to be fetched from the archive and replayed during recovery (remember that WAL archiving is what enables continuous backup in PostgreSQL!). Based on our experience, a weekly base backup is more than enough for most cases - while it is extremely rare to schedule backups more frequently than once a day.
You can choose whether to schedule a backup on a defined object store or a
volume snapshot via the .spec.method
attribute, by default set to
barmanObjectStore
. If you have properly defined
volume snapshots
in the backup
stanza of the cluster, you can set method: volumeSnapshot
to start scheduling base backups on volume snapshots.
ScheduledBackups can be suspended, if needed, by setting .spec.suspend: true
.
This will stop any new backup from being scheduled until the option is removed
or set back to false
.
In case you want to issue a backup as soon as the ScheduledBackup resource is created
you can set .spec.immediate: true
.
Note
.spec.backupOwnerReference
indicates which ownerReference should be put inside
the created backup resources.
- none: no owner reference for created backup objects (same behavior as before the field was introduced)
- self: sets the Scheduled backup object as owner of the backup
- cluster: set the cluster as owner of the backup
On-demand backups
Info
Please refer to BackupSpec
in the API reference for a full list of options.
To request a new backup, you need to create a new Backup
resource
like the following one:
apiVersion: postgresql.cnpg.io/v1
kind: Backup
metadata:
name: backup-example
spec:
method: barmanObjectStore
cluster:
name: pg-backup
In this case, the operator will start to orchestrate the cluster to take the
required backup on an object store, using barman-cloud-backup
. You can check
the backup status using the plain kubectl describe backup <name>
command:
Name: backup-example
Namespace: default
Labels: <none>
Annotations: API Version: postgresql.cnpg.io/v1
Kind: Backup
Metadata:
Creation Timestamp: 2020-10-26T13:57:40Z
Self Link: /apis/postgresql.cnpg.io/v1/namespaces/default/backups/backup-example
UID: ad5f855c-2ffd-454a-a157-900d5f1f6584
Spec:
Cluster:
Name: pg-backup
Status:
Phase: running
Started At: 2020-10-26T13:57:40Z
Events: <none>
When the backup has been completed, the phase will be completed
like in the following example:
Name: backup-example
Namespace: default
Labels: <none>
Annotations: API Version: postgresql.cnpg.io/v1
Kind: Backup
Metadata:
Creation Timestamp: 2020-10-26T13:57:40Z
Self Link: /apis/postgresql.cnpg.io/v1/namespaces/default/backups/backup-example
UID: ad5f855c-2ffd-454a-a157-900d5f1f6584
Spec:
Cluster:
Name: pg-backup
Status:
Backup Id: 20201026T135740
Destination Path: s3://backups/
Endpoint URL: http://minio:9000
Phase: completed
s3Credentials:
Access Key Id:
Key: ACCESS_KEY_ID
Name: minio
Secret Access Key:
Key: ACCESS_SECRET_KEY
Name: minio
Server Name: pg-backup
Started At: 2020-10-26T13:57:40Z
Stopped At: 2020-10-26T13:57:44Z
Events: <none>
Important
This feature will not backup the secrets for the superuser and the application user. The secrets are supposed to be backed up as part of the standard backup procedures for the Kubernetes cluster.
Backup from a standby
Taking a base backup requires to scrape the whole data content of the PostgreSQL instance on disk, possibly resulting in I/O contention with the actual workload of the database.
For this reason, CloudNativePG allows you to take advantage of a feature which is directly available in PostgreSQL: backup from a standby.
By default, backups will run on the most aligned replica of a Cluster
. If
no replicas are available, backups will run on the primary instance.
Info
Although the standby might not always be up to date with the primary,
in the time continuum from the first available backup to the last
archived WAL this is normally irrelevant. The base backup indeed
represents the starting point from which to begin a recovery operation,
including PITR. Similarly to what happens with
pg_basebackup
,
when backing up from an online standby we do not force a switch of the WAL on the
primary. This might produce unexpected results in the short term (before
archive_timeout
kicks in) in deployments with low write activity.
If you prefer to always run backups on the primary, you can set the backup
target to primary
as outlined in the example below:
apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
[...]
spec:
backup:
target: "primary"
Warning
Beware of setting the target to primary when performing a cold backup with volume snapshots, as this will shut down the primary for the time needed to take the snapshot, impacting write operations. This also applies to taking a cold backup in a single-instance cluster, even if you did not explicitly set the primary as the target.
When the backup target is set to prefer-standby
, such policy will ensure
backups are run on the most up-to-date available secondary instance, or if no
other instance is available, on the primary instance.
By default, when not otherwise specified, target is automatically set to take backups from a standby.
The backup target specified in the Cluster
can be overridden in the Backup
and ScheduledBackup
types, like in the following example:
apiVersion: postgresql.cnpg.io/v1
kind: Backup
metadata:
[...]
spec:
cluster:
name: [...]
target: "primary"
In the previous example, CloudNativePG will invariably choose the primary
instance even if the Cluster
is set to prefer replicas.