Security
This section contains information about security for CloudNativePG, that are analyzed at 3 different layers: Code, Container and Cluster.
Warning
The information contained in this page must not exonerate you from performing regular InfoSec duties on your Kubernetes cluster. Please familiarize yourself with the "Overview of Cloud Native Security" page from the Kubernetes documentation.
About the 4C's Security Model
Please refer to "The 4C’s Security Model in Kubernetes" blog article to get a better understanding and context of the approach EDB has taken with security in CloudNativePG.
Code
CloudNativePG's source code undergoes systematic static analysis, including checks for security vulnerabilities, using the popular open-source linter for Go, GolangCI-Lint, directly integrated into the CI/CD pipeline. GolangCI-Lint can run multiple linters on the same source code.
The following tools are used to identify security issues:
-
Golang Security Checker (
gosec
): A linter that scans the abstract syntax tree of the source code against a set of rules designed to detect known vulnerabilities, threats, and weaknesses, such as hard-coded credentials, integer overflows, and SQL injections. GolangCI-Lint runsgosec
as part of its suite. -
govulncheck: This tool runs in the CI/CD pipeline and reports known vulnerabilities affecting Go code or the compiler. If the operator is built with a version of the Go compiler containing a known vulnerability,
govulncheck
will detect it. -
CodeQL: Provided by GitHub, this tool scans for security issues and blocks any pull request with detected vulnerabilities. CodeQL is configured to review only Go code, excluding other languages in the repository such as Python or Bash.
-
Snyk: Conducts nightly code scans in a scheduled job and generates weekly reports highlighting any new findings related to code security and licensing issues.
The CloudNativePG repository has the "Private vulnerability reporting" option enabled in the Security section. This feature allows users to safely report security issues that require careful handling before being publicly disclosed. If you discover any security bug, please use this medium to report it.
Important
A failure in the static code analysis phase of the CI/CD pipeline will block the entire delivery process of CloudNativePG. Every commit must pass all the linters defined by GolangCI-Lint.
Container
Every container image in CloudNativePG is automatically built via CI/CD pipelines following every commit. These images include not only the operator's image but also the operands' images, specifically for every supported PostgreSQL version. During the CI/CD process, images undergo scanning with the following tools:
- Dockle: Ensures best practices in the container build process.
- Snyk: Detects security issues within the container and reports findings via the GitHub interface.
Important
All operand images are automatically rebuilt daily by our pipelines to incorporate security updates at the base image and package level, providing patch-level updates for the container images distributed to the community.
Guidelines and Frameworks for Container Security
The following guidelines and frameworks have been considered for ensuring container-level security:
- "Container Image Creation and Deployment Guide": Developed by the Defense Information Systems Agency (DISA) of the United States Department of Defense (DoD).
- "CIS Benchmark for Docker": Developed by the Center for Internet Security (CIS).
About Container-Level Security
For more information on the approach that EDB has taken regarding security at the container level in CloudNativePG, please refer to the blog article "Security and Containers in CloudNativePG".
Cluster
Security at the cluster level takes into account all Kubernetes components that form both the control plane and the nodes, as well as the applications that run in the cluster (PostgreSQL included).
Role Based Access Control (RBAC)
The operator interacts with the Kubernetes API server using a dedicated service
account named cnpg-manager
. This service account is typically installed in
the operator namespace, commonly cnpg-system
. However, the namespace may vary
based on the deployment method (see the subsection below).
In the same namespace, there is a binding between the cnpg-manager
service
account and a role. The specific name and type of this role (either Role
or
ClusterRole
) also depend on the deployment method. This role defines the
necessary permissions required by the operator to function correctly. To learn
more about these roles, you can use the kubectl describe clusterrole
or
kubectl describe role
commands, depending on the deployment method.
Important
The above permissions are exclusively reserved for the operator's service
account to interact with the Kubernetes API server. They are not directly
accessible by the users of the operator that interact only with Cluster
,
Pooler
, Backup
, ScheduledBackup
, Database
, Publication
,
Subscription
, ImageCatalog
and ClusterImageCatalog
resources.
Below we provide some examples and, most importantly, the reasons why CloudNativePG requires full or partial management of standard Kubernetes namespaced or non-namespaced resources.
configmaps
- The operator needs to create and manage default config maps for the Prometheus exporter monitoring metrics.
deployments
- The operator needs to manage a PgBouncer connection pooler
using a standard Kubernetes
Deployment
resource. jobs
- The operator needs to handle jobs to manage different
Cluster
's phases. persistentvolumeclaims
- The volume where the
PGDATA
resides is the central element of a PostgreSQLCluster
resource; the operator needs to interact with the selected storage class to dynamically provision the requested volumes, based on the defined scheduling policies. pods
- The operator needs to manage
Cluster
's instances. secrets
- Unless you provide certificates and passwords to your
Cluster
objects, the operator adopts the "convention over configuration" paradigm by self-provisioning random generated passwords and TLS certificates, and by storing them in secrets. serviceaccounts
- The operator needs to create a service account that
enables the instance manager (which is the PID 1 process of the container
that controls the PostgreSQL server) to safely communicate with the
Kubernetes API server to coordinate actions and continuously provide
a reliable status of the
Cluster
. services
- The operator needs to control network access to the PostgreSQL cluster (or the connection pooler) from applications, and properly manage failover/switchover operations in an automated way (by assigning, for example, the correct end-point of a service to the proper primary PostgreSQL instance).
validatingwebhookconfigurations
andmutatingwebhookconfigurations
- The operator injects its self-signed webhook CA into both webhook configurations, which are needed to validate and mutate all the resources it manages. For more details, please see the Kubernetes documentation.
volumesnapshots
- The operator needs to generate
VolumeSnapshots
objects in order to take backups of a PostgreSQL server. VolumeSnapshots are read too in order to validate them before starting the restore process. nodes
- The operator needs to get the labels for Affinity and AntiAffinity so it can decide in which nodes a pod can be scheduled. This is useful, for example, to prevent the replicas from being scheduled in the same node - especially important if nodes are in different availability zones. This permission is also used to determine whether a node is scheduled, preventing the creation of pods on unscheduled nodes, or triggering a switchover if the primary lives in an unscheduled node.
Deployments and ClusterRole
Resources
As mentioned above, each deployment method may have variations in the namespace location of the service account, as well as the names and types of role bindings and respective roles.
Via Kubernetes Manifest
When installing CloudNativePG using the Kubernetes manifest, permissions are
set to ClusterRoleBinding
by default. You can inspect the permissions
required by the operator by running:
kubectl describe clusterrole cnpg-manager
Via OLM
From a security perspective, the Operator Lifecycle Manager (OLM) provides a more flexible deployment method. It allows you to configure the operator to watch either all namespaces or specific namespaces, enabling more granular permission management.
Info
OLM allows you to deploy the operator in its own namespace and configure it to watch specific namespaces used for CloudNativePG clusters. This setup helps to contain permissions and restrict access more effectively.
Why Are ClusterRole Permissions Needed?
The operator currently requires ClusterRole
permissions to read nodes
and
ClusterImageCatalog
objects. All other permissions can be namespace-scoped (i.e., Role
) or
cluster-wide (i.e., ClusterRole
).
Even with these permissions, if someone gains access to the ServiceAccount
,
they will only have get
, list
, and watch
permissions, which are limited
to viewing resources. However, if an unauthorized user gains access to the
ServiceAccount
, it indicates a more significant security issue.
Therefore, it's crucial to prevent users from accessing the operator's
ServiceAccount
and any other ServiceAccount
with elevated permissions.
Calls to the API server made by the instance manager
The instance manager, which is the entry point of the operand container, needs
to make some calls to the Kubernetes API server to ensure that the status of
some resources is correctly updated and to access the config maps and secrets
that are associated with that Postgres cluster. Such calls are performed through
a dedicated ServiceAccount
created by the operator that shares the same
PostgreSQL Cluster
resource name.
Important
The operand can only access a specific and limited subset of resources through the API server. A service account is the recommended way to access the API server from within a Pod.
For transparency, the permissions associated with the service account are defined in the
roles.go
file. For example, to retrieve the permissions of a generic mypg
cluster in the
myns
namespace, you can type the following command:
kubectl get role -n myns mypg -o yaml
Then verify that the role is bound to the service account:
kubectl get rolebinding -n myns mypg -o yaml
Important
Remember that roles are limited to a given namespace.
Below we provide a quick summary of the permissions associated with the service account for generic Kubernetes resources.
configmaps
- The instance manager can only read config maps that are related to the same cluster, such as custom monitoring queries
secrets
- The instance manager can only read secrets that are related to the same cluster, namely: streaming replication user, application user, super user, LDAP authentication user, client CA, server CA, server certificate, backup credentials, custom monitoring queries
events
- The instance manager can create an event for the cluster, informing the API server about a particular aspect of the PostgreSQL instance lifecycle
Here instead, we provide the same summary for resources specific to CloudNativePG.
clusters
- The instance manager requires read-only permissions, namely
get
,list
andwatch
, just for its ownCluster
resource clusters/status
- The instance manager requires to
update
andpatch
the status of just its ownCluster
resource backups
- The instance manager requires
get
andlist
permissions to read anyBackup
resource in the namespace. Additionally, it requires thedelete
permission to clean up the Kubernetes cluster by removing theBackup
objects that do not have a counterpart in the object store - typically because of retention policies backups/status
- The instance manager requires to
update
andpatch
the status of anyBackup
resource in the namespace
Pod Security Policies
Important
Starting from Kubernetes v1.21, the use of PodSecurityPolicy
has been
deprecated, and as of Kubernetes v1.25, it has been completely removed. Despite
this deprecation, we acknowledge that the operator is currently undergoing
testing in older and unsupported versions of Kubernetes. Therefore, this
section is retained for those specific scenarios.
A Pod Security Policy is the Kubernetes way to define security rules and specifications that a pod needs to meet to run in a cluster. For InfoSec reasons, every Kubernetes platform should implement them.
CloudNativePG does not require privileged mode for containers execution.
The PostgreSQL containers run as postgres
system user. No component whatsoever requires running as root
.
Likewise, Volumes access does not require privileges mode or root
privileges either.
Proper permissions must be properly assigned by the Kubernetes platform and/or administrators.
The PostgreSQL containers run with a read-only root filesystem (i.e. no writable layer).
The operator explicitly sets the required security contexts.
Restricting Pod access using AppArmor
You can assign an
AppArmor profile to
the postgres
, initdb
, join
, full-recovery
and bootstrap-controller
containers inside every Cluster
pod through the
container.apparmor.security.beta.kubernetes.io
annotation.
Example of cluster annotations
kind: Cluster
metadata:
name: cluster-apparmor
annotations:
container.apparmor.security.beta.kubernetes.io/postgres: runtime/default
container.apparmor.security.beta.kubernetes.io/initdb: runtime/default
container.apparmor.security.beta.kubernetes.io/join: runtime/default
Warning
Using this kind of annotations can result in your cluster to stop working.
If this is the case, the annotation can be safely removed from the Cluster
.
The AppArmor configuration must be at Kubernetes node level, meaning that the underlying operating system must have this option enable and properly configured.
In case this is not the situation, and the annotations were added at the
Cluster
creation time, pods will not be created. On the other hand, if you
add the annotations after the Cluster
was created the pods in the cluster will
be unable to start and you will get an error like this:
metadata.annotations[container.apparmor.security.beta.kubernetes.io/postgres]: Forbidden: may not add AppArmor annotations]
In such cases, please refer to your Kubernetes administrators and ask for the proper AppArmor profile to use.
Network Policies
The pods created by the Cluster
resource can be controlled by Kubernetes
network policies
to enable/disable inbound and outbound network access at IP and TCP level.
You can find more information in the networking document.
Important
The operator needs to communicate to each instance on TCP port 8000 to get information about the status of the PostgreSQL server. Please make sure you keep this in mind in case you add any network policy, and refer to the "Exposed Ports" section below for a list of ports used by CloudNativePG for finer control.
Network policies are beyond the scope of this document. Please refer to the "Network policies" section of the Kubernetes documentation for further information.
Exposed Ports
CloudNativePG exposes ports at operator, instance manager and operand levels, as listed in the table below:
System | Port number | Exposing | Name | TLS | Authentication |
---|---|---|---|---|---|
operator | 9443 | webhook server | webhook-server |
Yes | Yes |
operator | 8080 | metrics | metrics |
No | No |
instance manager | 9187 | metrics | metrics |
Optional | No |
instance manager | 8000 | status | status |
Yes | No |
operand | 5432 | PostgreSQL instance | postgresql |
Optional | Yes |
PostgreSQL
The current implementation of CloudNativePG automatically creates
passwords and .pgpass
files for the database owner and, only
if requested by setting enableSuperuserAccess
to true
, for the
postgres
superuser.
Warning
enableSuperuserAccess
is set to false
by default to improve the
security-by-default posture of the operator, fostering a microservice approach
where changes to PostgreSQL are performed in a declarative way through the
spec
of the Cluster
resource, while providing developers with full powers
inside the database through the database owner user.
As far as password encryption is concerned, CloudNativePG follows
the default behavior of PostgreSQL: starting from PostgreSQL 14,
password_encryption
is by default set to scram-sha-256
, while on earlier
versions it is set to md5
.
Important
Please refer to the "Password authentication" section in the PostgreSQL documentation for details.
Note
The operator supports toggling the enableSuperuserAccess
option. When you
disable it on a running cluster, the operator will ignore the content of the secret,
remove it (if previously generated by the operator) and set the password of the
postgres
user to NULL
(de facto disabling remote access through password authentication).
See the "Secrets" section in the "Connecting from an application" page for more information.
You can use those files to configure application access to the database.
By default, every replica is automatically configured to connect in physical
async streaming replication with the current primary instance, with a special
user called streaming_replica
. The connection between nodes is encrypted
and authentication is via TLS client certificates (please refer to the
"Client TLS/SSL Connections" page
for details). By default, the operator requires TLS v1.3 connections.
Currently, the operator allows administrators to add pg_hba.conf
lines directly in the manifest
as part of the pg_hba
section of the postgresql
configuration. The lines defined in the
manifest are added to a default pg_hba.conf
.
For further detail on how pg_hba.conf
is managed by the operator, see the
"PostgreSQL Configuration" page of the documentation.
The administrator can also customize the content of the pg_ident.conf
file that by default
only maps the local postgres user to the postgres user in the database.
For further detail on how pg_ident.conf
is managed by the operator, see the
"PostgreSQL Configuration" page of the documentation.
Important
Examples assume that the Kubernetes cluster runs in a private and secure network.
Storage
CloudNativePG delegates encryption at rest to the underlying storage class. For data protection in production environments, we highly recommend that you choose a storage class that supports encryption at rest.