Storage is the most critical component in a database workload. Storage should be always available, scale, perform well, and guarantee consistency and durability. The same expectations and requirements that apply to traditional environments, such as virtual machines and bare metal, are also valid in container contexts managed by Kubernetes.
Kubernetes has its own specificities, when it comes to dynamically provisioned storage. These include storage classes, persistent volumes, and persistent volume claims. You need to own these concepts, on top of all the valuable knowledge you have built over the years in terms of storage for database workloads on VMs and physical servers.
There are two primary methods of access to storage:
- network: either directly or indirectly (think of an NFS volume locally mounted on a host running Kubernetes)
- local: directly attached to the node where a Pod is running (this also includes directly attached disks on bare metal installations of Kubernetes)
Network storage, which is the most common usage pattern in Kubernetes, presents the same issues of throughput and latency that you can experience in a traditional environment. These can be accentuated in a shared environment, where I/O contention with several applications increases the variability of performance results.
Local storage enables shared-nothing architectures, which is more suitable for high transactional and Very Large DataBase (VLDB) workloads, as it guarantees higher and more predictable performance.
EDB Postgres for Kubernetes does not use
StatefulSets for managing data persistence.
Rather, it manages persistent volume claims (PVCs) directly. If you want
to know more, please read the
"Custom Pod Controller" document.
Since EDB Postgres for Kubernetes supports volume snapshots for both backup and recovery, we recommend that you also consider this aspect when you choose your storage solution, especially if you manage very large databases.
Please refer to the official Kubernetes documentation for a list of all the supported Container Storage Interface (CSI) drivers that provide snapshotting capabilities.
We recommend that you benchmark EDB Postgres for Kubernetes in a controlled Kubernetes environment, before deploying the database in production, by following the guidelines in the "Benchmarking" section.
Briefly, our advice is to operate at two levels:
- measuring the performance of the underlying storage using
fio, with relevant metrics for database workloads such as throughput for sequential reads, sequential writes, random reads and random writes
- measuring the performance of the database using the default benchmarking tool
distributed along with PostgreSQL:
Measuring both the storage and database performance is an activity that must be done before the database goes in production. However, such results are extremely valuable not only in the planning phase (e.g., capacity planning), but also in the production lifecycle, especially in emergency situations (when we don't have the luxury anymore to run this kind of tests). Databases indeed change and evolve over time, so does the distribution of data, potentially affecting performance: knowing the theoretical maximum throughput of sequential reads or writes will turn out to be extremely useful in those situations. Especially in shared-nothing contexts, where results do not vary due to the influence of external workloads. Know your system, benchmark it.
Encryption at rest is possible with EDB Postgres for Kubernetes. The operator delegates that to the underlying storage class. Please refer to the storage class for information about this important security feature.
The operator creates a persistent volume claim (PVC) for each PostgreSQL
instance, with the goal to store the
PGDATA, and then mounts it into each Pod.
Additionally, it supports the creation of clusters with a separate PVC on which to store PostgreSQL Write-Ahead Log (WAL), as explained in the "Volume for WAL" section below.
In EDB Postgres for Kubernetes, the volumes attached to a single PostgreSQL instance are defined as PVC group.
EDB Postgres for Kubernetes has been designed to be storage class agnostic. As usual, our recommendation is to properly benchmark the storage class in a controlled environment, before deploying to production.
The easier way to configure the storage for a PostgreSQL class is to just request storage of a certain size, like in the following example:
Using the previous configuration, the generated PVCs will be satisfied by the default storage class. If the target Kubernetes cluster has no default storage class, or even if you need your PVCs to be satisfied by a known storage class, you can set it into the custom resource:
To further customize the generated PVCs, you can provide a PVC template inside the Custom Resource, like in the following example:
By default, PostgreSQL stores all its data in the so-called
PGDATA (a directory).
One of the core directories inside
pg_xlog in PostgreSQL), which contains the log of transactional
changes occurred in the database, in the form of segment files.
Normally, each segment is 16 MB in size, but the size can be configured
walSegmentSize option, applied at cluster initialization time, as
described in "Bootstrap an empty cluster".
While in most cases, having
pg_wal on the same volume where
resides is fine, there are a few benefits from having WALs stored in a separate
I/O performance: by storing WAL files on different storage than
PGDATA, PostgreSQL can exploit parallel I/O for WAL operations (normally sequential writes) and for data files (tables and indexes for example), thus improving vertical scalability
more reliability: by reserving dedicated disk space to WAL files, you can always be sure that exhaustion of space on the
PGDATAvolume will never interfere with WAL writing, ensuring that your PostgreSQL primary is correctly shut down.
finer control: you can define the amount of space dedicated to both
pg_wal, fine tune WAL configuration and checkpoints, even use a different storage class for cost optimization
better I/O monitoring: you can constantly monitor the load and disk usage on both
pg_wal, and set proper alerts that notify you in case, for example,
Write-Ahead Log (WAL)
Please refer to the "Reliability and the Write-Ahead Log" page from the official PostgreSQL documentation for more information.
You can add a separate volume for WAL through the
which follows the same rules described for the
storage field and provisions a
dedicated PVC. For example:
walStorage is not supported: once added, a separate volume for
WALs cannot be removed from an existing Postgres cluster.
Kubernetes exposes an API allowing expanding PVCs
that is enabled by default but needs to be supported by the underlying
To check if a certain
StorageClass supports volume expansion, you can read the
field for your storage class:
Given the storage class supports volume expansion, you can change the size requirement
Cluster, and the operator will apply the change to every PVC.
StorageClass supports online volume resizing
the change is immediately applied to the Pods. If the underlying Storage Class doesn't support
that, you will need to delete the Pod to trigger the resize.
The best way to proceed is to delete one Pod at a time, starting from replicas and waiting for each Pod to be back up.
At the moment, Azure is not able to resize the PVC's volume without restarting the pod.
EDB Postgres for Kubernetes has overcome this limitation through the
ENABLE_AZURE_PVC_UPDATES environment variable in the
When set to
'true', EDB Postgres for Kubernetes triggers a rolling update of the
Alternatively, you can follow the workaround below to manually resize the volume in AKS.
You can manually resize a PVC on AKS by following these procedures. As an example, let's suppose you have a cluster with 3 replicas:
An Azure disk can only be expanded while in "unattached" state, as described in the
This means, that to resize a disk used by a PostgreSQL cluster, you will need to perform a manual rollout, first cordoning the node that hosts the Pod using the PVC bound to the disk. This will prevent the Operator to recreate the Pod and immediately reattach it to its PVC before the background disk resizing has been completed.
First step is to edit the cluster definition applying the new size, let's say "2Gi", as follows:
cluster-example-1 Pod is the cluster's primary, we can proceed with the replicas first.
For example start with cordoning the kubernetes node that hosts the
Then delete the
Run the following command:
Wait until you see the following output:
Then, you can uncordon the node:
Wait for the Pod to be recreated correctly and get in Running and Ready state:
Now verify the PVC expansion by running the following command, which should return "2Gi" as configured:
So, you can repeat these steps for the remaining Pods.
Please leave the resizing of the disk associated with the primary instance as last disk,
after promoting through a switchover a new resized Pod, using
kubectl cnp promote
kubectl cnp promote cluster-example 3 to promote
cluster-example-3 to primary).
If the storage class does not support volume expansion, you can still regenerate your cluster on different PVCs, by allocating new PVCs with increased storage and then move the database there. This operation is feasible only when the cluster contains more than one node.
While you do that, you need to prevent the operator from changing the existing PVC
by disabling the
resizeInUseVolumes flag, like in the following example:
In order to move the entire cluster to a different storage area, you need to recreate all the PVCs and all the Pods. Let's suppose you have a cluster with three replicas like in the following example:
To recreate the cluster using different PVCs, you can edit the cluster definition to disable
resizeInUseVolumes, and then recreate every instance in a different PVC.
As an example, to recreate the storage for
cluster-example-3 you can:
In case you have created a dedicated WAL volume, both PVCs will have to be deleted during this process.
Additionally, the same procedure applies in case you want to regenerate the WAL volume PVC, which can be done
resizeInUseVolumes also for the
For example (in case a PVC dedicated to WAL storage is present):
Having done that, the operator will orchestrate the creation of another replica with a resized PVC:
EDB Postgres for Kubernetes has been designed to work with dynamic volume provisioning, which allows storage volumes to be automatically created on-demand when requested by users, via storage classes and persistent volume claim templates as described above.
However, in some cases, Kubernetes administrators prefer to manually create new
storage volumes, and then create the related
PersistentVolume objects for
their representation inside the Kubernetes cluster. This is also known as
pre-provisioning of volumes.
Our recommendation is to avoid pre-provisioning of volumes as it impacts on the high availability and self-healing capabilities of the operator, and breaks the fully declarative model on which EDB Postgres for Kubernetes has been built.
You can use a pre-provisioned volume in EDB Postgres for Kubernetes by following these steps:
- Manually create the volume outside Kubernetes
- Create the
PersistentVolumeobject to match the above volume using the correct parameters as required by actual CSI driver (i.e.
storageClassName, and so on)
- Create the Postgres
Clusterusing, for each storage section, a coherent
pvcTemplatesection that can help Kubernetes match the above
PersistentVolumeand enable EDB Postgres for Kubernetes to create the needed
With static provisioning, it is your responsibility to ensure that, based
on the affinity rules of your cluster, Postgres pods can be correctly scheduled
by Kubernetes where a pre-provisioned volume exists. Make sure you check
for any pods stuck in
Pending after you have deployed the cluster, and
if the condition persists investigate why this is happening.
Most block storage solutions in Kubernetes suggest to have multiple 'replicas' of a volume
to improve resiliency. This works well for workloads that don't have resiliency built into the
application. However, EDB Postgres for Kubernetes has this resiliency built directly into the Postgres
through the number of instances and the persistent volumes that are attached to them.
In these cases it makes sense to define the storage class used by the Postgres clusters to be defined as 1 replica. By having additional replicas defined in the storage solution like Longhorn and Ceph you might incur in the issue known as write amplification, unnecessarily increasing disk I/O and space used.