Custom Pod Controller v1
Kubernetes uses the Controller pattern to align the current cluster state with the desired one.
Stateful applications are usually managed with the
controller, which creates and reconciles a set of Pods built from the same
specification, and assigns them a sticky identity.
EDB Postgres for Kubernetes implements its own custom controller to manage PostgreSQL
instances, instead of relying on the
While bringing more complexity to the implementation, this design choice
provides the operator with more flexibility on how we manage the cluster,
while being transparent on the topology of PostgreSQL clusters.
Like many choices in the design realm, different ones lead to other compromises. The following sections discuss a few points where we believe this design choice has made the implementation of EDB Postgres for Kubernetes more reliable, and easier to understand.
This is a well known limitation of
StatefulSet: it does not support resizing
PVCs. This is inconvenient for a database. Resizing volumes requires
In contrast, EDB Postgres for Kubernetes leverages the configured storage class to manage the underlying PVCs directly, and can handle PVC resizing if the storage class supports it.
StatefulSet controller is designed to create a set of Pods
from just one template. Given that we use one
Pod per PostgreSQL instance,
we have two kinds of Pods:
- primary instance (only one)
- replicas (multiple, optional)
This difference is relevant when deciding the correct deployment strategy to execute for a given operation.
Some operations should be performed on the replicas first,
and then on the primary, but only after an updated replica is promoted
as the new primary.
For example, when you want to apply a different PostgreSQL image version,
or when you increase configuration parameters like
max_connections (which are
treated specially by PostgreSQL because EDB Postgres for Kubernetes uses hot standby
While doing that, EDB Postgres for Kubernetes considers the PostgreSQL instance's role - and not just its serial number.
Sometimes the operator needs to follow the opposite process: work on the
primary first and then on the replicas. For example, when you
max_connections. In that case, EDB Postgres for Kubernetes will:
- apply the new setting to the primary instance
- restart it
- apply the new setting on the replicas
StatefulSet controller, being application-independent, can't
incorporate this behavior, which is specific to PostgreSQL's native
PostgreSQL instances can be configured to work with multiple PVCs: this is how
WAL storage can be separated from
The two data stores need to be coherent from the PostgreSQL point of view,
as they're used simultaneously. If you delete the PVC corresponding to
the WAL storage of an instance, the PVC where
PGDATA is stored will not be
This behavior is specific to PostgreSQL and is not implemented in the
StatefulSet controller - the latter not being application specific.
After the user dropped a PVC, a
StatefulSet would just recreate it, leading
to a corrupted PostgreSQL instance.
EDB Postgres for Kubernetes would instead classify the remaining PVC as unusable, and start creating a new pair of PVCs for another instance to join the cluster correctly.
Sometimes you need to take down a Kubernetes node to do an upgrade. After the upgrade, depending on your upgrade strategy, the updated node could go up again, or a new node could replace it.
Supposing the unavailable node was hosting a PostgreSQL instance, depending on your database size and your cloud infrastructure, you may prefer to choose one of the following actions:
drop the PVC and the Pod residing on the downed node; create a new PVC cloning the data from another PVC; after that, schedule a Pod for it
drop the Pod, schedule the Pod in a different node, and mount the PVC from there
leave the Pod and the PVC as they are, and wait for the node to be back up.
The first solution is practical when your database size permits, allowing you to immediately bring back the desired number of replicas.
The second solution is only feasible when you're not using the storage of the local node, and re-mounting the PVC in another host is possible in a reasonable amount of time (which only you and your organization know).
The third solution is appropriate when the database is big and uses local node storage for maximum performance and data durability.
The EDB Postgres for Kubernetes controller implements all these strategies so that the user can select the preferred behavior at the cluster level (read the "Kubernetes upgrade" section for details).
Being generic, the
StatefulSet doesn't allow this level of