How NGINX Ingress Controller Works
This document explains how NGINX Ingress Controller works. The target audience includes the following two main groups:
- Operators who would like to know how the software works and also better understand how it can fail.
- Developers who would like to contribute to the project.
We assume that the reader is familiar with core Kubernetes concepts, such as Pod, Deployment, Service, and Endpoints. Additionally, we recommend reading this blog post for an overview of the NGINX architecture.
An Ingress Controller is a component in a Kubernetes cluster that configures an HTTP load balancer according to Ingress resources created by the cluster user.
If you’d like to read more about the Ingress resource, refer to the official Kubernetes documentation.
This document is specific to NGINX Ingress Controller, referred to as Ingress Controller or IC, which is built upon NGINX and NGINX Plus capabilities
Let’s begin with a high-level examination of the Ingress Controller (IC). The following figure depicts an example of how the IC exposes two web applications running in a Kubernetes cluster to clients on the internet:
The figure shows:
- A Kubernetes cluster.
- Cluster users Admin, User A and User B, which use the cluster via the Kubernetes API.
- Clients A and Clients B, which connect to the Applications A and B deployed by the corresponding users.
- IC, deployed by Admin in a pod in the namespace nginx-ingress and configured via the ConfigMap nginx-ingress. For simplicity, we depict only one IC pod; however, Admin typically deploys at least two pods for redundancy. The IC uses the Kubernetes API to get the latest Ingress resources created in the cluster and then configures NGINX according to those resources.
- Application A with two pods deployed in the namespace A by User A. To expose the application to its clients (Clients A) via the host
a.example.com, User A creates Ingress A.
- Application B with one pod deployed in the namespace B by User B. To expose the application to its clients (Clients B) via the host
b.example.com, User B creates VirtualServer B.
- Public Endpoint, which fronts the IC pod(s). This is typically a TCP load balancer (cloud, software, or hardware) or a combination of such load balancer with a NodePort service. Clients A and B connect to their applications via the Public Endpoint.
The yellow and purple arrows represent connections related to the client traffic, and the black arrows represent access to the Kubernetes API.
For simplicity, many necessary Kubernetes resources like Deployment and Services aren’t shown, which Admin and the users also need to create.
Next, let’s explore the IC pod.
The IC pod consists of a single container, which in turn includes the following:
- IC process, which configures NGINX according to Ingress and other resources created in the cluster.
- NGINX master process, which controls NGINX worker processes.
- NGINX worker processes, which handle the client traffic and load balance the traffic to the backend applications.
The following is an architectural diagram that shows how those processes interact together and with some external processes/entities:
For brevity, we’ve omitted the suffix process from the description of the processes.
The numbered list that follows describes each connection with its type in curly brackets:
- (HTTP) Prometheus fetches the IC and NGINX metrics via an HTTP endpoint that the IC exposes. The default is
:9113/metrics. Note: Prometheus is not required by the IC, the endpoint can be turned off.
- (HTTPS) The IC reads the Kubernetes API to get the latest versions of the resources in the cluster and writes to the API to update the handled resources' statuses and emit events.
- (HTTP) Kubelet probes the IC readiness probe (the default is
:8081/nginx-ready) to consider the IC pod ready.
- (File I/O) When the IC starts, it reads the configuration templates necessary for config generation from the filesystem. The templates are located in the
/directory of the container and have the
- (File I/O) The IC writes logs to its stdout and stderr, which are collected by the container runtime.
- (File I/O) The IC generates NGINX configuration based on the resources created in the cluster (refer to The Ingress Controller is a Kubernetes Controller section for the list of resources) and writes it on the filesystem in the
/etc/nginxfolder. The configuration files have a
- (File I/O) The IC writes TLS certificates and keys from any TLS Secrets referenced in the Ingress and other resources to the filesystem.
- (HTTP) The IC fetches the NGINX metrics via the
unix:/var/lib/nginx/nginx-status.sockUNIX socket and converts it to Prometheus format used in #1.
- (HTTP) To consider a configuration reload a success, the IC ensures that at least one NGINX worker has the new configuration. To do that, the IC checks a particular endpoint via the
- (N/A) To start NGINX, the IC runs the
nginxcommand, which launches the NGINX master.
- (Signal) To reload NGINX, the IC runs the
nginx -s reloadcommand, which validates the configuration and sends the reload signal to the NGINX master.
- (Signal) To shutdown NGINX, the IC executes
nginx -s quitcommand, which sends the graceful shutdown signal to the NGINX master.
- (File I/O) The NGINX master sends logs to its stdout and stderr, which are collected by the container runtime.
- (File I/O) The NGINX master reads the TLS cert and keys referenced in the configuration when it starts or reloads.
- (File I/O) The NGINX master reads configuration files when it starts or during a reload.
- (Signal) The NGINX master controls the lifecycle of NGINX workers it creates workers with the new configuration and shutdowns workers with the old configuration.
- (File I/O) An NGINX worker writes logs to its stdout and stderr, which are collected by the container runtime.
- (UDP) An NGINX worker sends the HTTP upstream server response latency logs via the Syslog protocol over the UNIX socket
/var/lib/nginx/nginx-syslog.sockto the IC. In turn, the IC analyzes and transforms the logs into Prometheus metrics.
- (HTTP,HTTPS,TCP,UDP) A client sends traffic to and receives traffic from any of the NGINX workers on ports 80 and 443 and any additional ports exposed by the GlobalConfiguration resource.
- (HTTP,HTTPS,TCP,UDP) An NGINX worker sends traffic to and receives traffic from the backends.
- (HTTP) Admin can connect to the NGINX stub_status using port 8080 via an NGINX worker. Note: By default, NGINX only allows connections from
The preceding diagram depicts the IC with NGINX. The IC also supports NGINX Plus with the following important differences:
- To configure NGINX Plus, in addition to configuration reloads, the IC uses the NGINX Plus API, which allows the IC to dynamically change the upstream servers of an upstream.
- Instead of the stub status metrics, the extended metrics are used, which are available via NGINX Plus API.
- In addition to TLS certs and keys, the IC writes JWKs from the secrets of the type
nginx.org/jwk, and NGINX workers read them.
This section covers the architecture of the IC process, including:
- How the IC processes a new Ingress resource created by a user.
- The summary of how the IC works and how it relates to Kubernetes Controllers.
- The different components of the IC process.
The following diagram depicts how the IC processes a new Ingress resource. We represent the NGINX master and worker processes as a single rectangle NGINX for simplicity. Also, note that VirtualServer and VirtualServerRoute resources are processed similarly.
Processing a new Ingress resource involves the following steps, where each step corresponds to the arrow on the diagram with the same number:
- User creates a new Ingress resource.
- The IC process has a Cache of the resources in the cluster. The Cache includes only the resources the IC is interested in, such as Ingresses. The Cache stays in sync with the Kubernetes API by watching for changes to the resources.
- Once the Cache has the new Ingress resource, it notifies the Control loop about the changed resource.
- The Control loop gets the latest version of the Ingress resource from the Cache. Because the Ingress resource references other resources, such as TLS Secrets, the Control loop gets the latest versions of any referenced resources as well.
- The Control loop generates TLS certificates and keys from the TLS Secrets and writes them to the filesystem.
- The Control loop generates and writes the NGINX configuration files, which correspond to the Ingress resource, and writes them to the filesystem.
- The Control loop reloads NGINX and waits for NGINX to successfully reload. As part of the reload:
- NGINX reads the TLS certs and keys.
- NGINX reads the configuration files.
- The Control loop emits an event for the Ingress resource and updates its status. If the reload fails, the event includes the error message.
Based on the example from the previous section, we can generalize how the IC works:
The IC constantly processes both new resources and changes to the existing resources in the cluster. As a result, the NGINX configuration stays up-to-date with the resources in the cluster.
The IC is an example of a Kubernetes controller: the IC runs a control loop that ensures NGINX is configured according to the desired state (Ingresses and other resources).
The desired state is concentrated in the following built-in Kubernetes resources and Custom Resources (CRs):
- Layer 7 Load balancing configuration:
- VirtualServers (CR)
- VirtualServerRoutes (CR)
- Layer 7 policies:
- Policies (CR)
- Layer 4 load balancing configuration:
- TransportServers (CR)
- Service discovery:
- Secret configuration:
- Global Configuration:
- ConfigMap (only one resource)
- GlobalConfiguration (CR, only one resource)
The IC can watch additional Custom Resources, which are less common and not enabled by default:
- NGINX App Protect resources (APPolicies, APLogConfs, APUserSigs)
- IngressLink resource (only one resource)
In the next section, we examine the different components of the IC process.
In this section, we describe the components of the IC process and how they interact, including:
- How the IC watches for resources changes.
- The main components of the IC control loop.
- How those components process a resource change.
- A few additional components, which are crucial for processing changes.
In the section Processing a New Ingress Resource, we mentioned that the IC has a cache of the resources in the cluster that stays in sync with the Kubernetes API by watching for changes to the resources. We also mentioned that once cache is updated, it notifies the control loop about the changed resource.
The cache is actually a collection of informers. The following diagram shows how changes to resources are processed by the IC.
- For every resource type the IC monitors, it creates an Informer. The Informer includes a Store that holds the resources of that type. To keep the Store in sync with the latest versions of the resources in the cluster, the Informer calls the Watch and List Kubernetes APIs for that resource type (see the arrow 1. Watch and List on the diagram).
- When a change happens in the cluster (for example, a new resource is created), the Informer updates its Store and invokes Handlers (see the arrow 2. Invoke) for that Informer.
- The IC registers handlers for every Informer. Most of the time, a Handler creates an entry for the affected resource in the Workqueue where a workqueue element includes the type of the resource and its namespace and name. (See the arrow 3. Put.)
- The Workqueue always tries to drain itself: if there is an element at the front, the queue will remove the element and send it to the Controller by calling a callback function. (See the arrow 4. Send.)
- The Controller is the primary component in the IC, which represents the control loop. We explain the components in The Control Loop section. For now, it suffices to know that to process a workqueue element, the Controller component gets the latest version of the resource from the Store (see the arrow 5. Get), reconfigures NGINX according to the resource (see the arrow 6. Reconfigure), updates the resource status, and emits an event via the Kubernetes API (see the arrow 7. Update status and emit event).
This section discusses the main components of the IC, which comprise the control loop:
- Runs the IC control loop.
- Instantiates Informers, Handlers, the Workqueue and additional helper components.
- Includes the sync method (see the next section), which is called by the Workqueue to process a changed resource.
- Passes changed resources to Configurator to re-configure NGINX.
- Generates NGINX configuration files, TLS and cert keys, and JWKs based on the Kubernetes resource.
- Uses Manager to write the generated files and reload NGINX.
- Controls the lifecycle of NGINX (starting, reloading, quitting).
- Manages the configuration files, TLS keys and certs, and JWKs.
The following diagram shows how the three components interact:
The Controller sync method is called by the Workqueue to process a change of a resource. The method determines the kind of the resource and calls the appropriate sync method (for example, syncIngress for Ingresses).
Rather than show how all the various sync methods work, we focus on the most important one – the syncIngress method – and look at how it processes a new Ingress resource, illustrated in the diagram below.
- The Workqueue calls the sync method and passes a workqueue element to it that includes the changed resource kind and key (the key is the resource namespace/name like “default/cafe-ingress”).
- Using the kind, the sync method calls the appropriate sync method and passes the resource key. For Ingresses, that method is syncIngress.
- syncIngress gets the Ingress resource from the Ingress Store using the key. The Store is controlled by the Ingress Informer, as mentioned in the section Resource Caches. Note: In the code, we use the helper storeToIngressLister type that wraps the Store.
- syncIngress calls AddOrUpdateIngress of the Configuration, passing the Ingress along. The Configuration is a component that represents a valid collection of load balancing configuration resources (Ingresses, VirtualServers, VirtualServerRoutes, TransportServers), ready to be converted to the NGINX configuration (see the Configuration section for more details). AddOrUpdateIngress returns a list of ResourceChanges, which must be reflected in the NGINX config. Typically, for a new Ingress resource, the Configuration returns only a single ResourceChange.
- syncIngress calls processChanges, which processes the single Ingress ResourceChange.
- processChanges creates an extended Ingress resource (IngressEx) that includes the original Ingress resource and its dependencies, such as Endpoints and Secrets, to generate the NGINX configuration. For simplicity, we don’t show this step on the diagram.
- processChanges calls AddOrUpdateIngress of the Configurator and passes the extended Ingress resource.
- Configurator generates an NGINX configuration file based on the extended Ingress resource and then:
- Calls Manager’s CreateConfig() to update the config for the Ingress resource.
- Calls Manager’s Reload() to reload NGINX.
- The reload status is propagated from Manager to processChanges. The status is either a success or a failure with an error message.
- processChanges calls updateRegularIngressStatusAndEvent to update the status of the Ingress resource and emit an event with the status of the reload. Both involve making an API call to the Kubernetes API.
- Some details weren’t discussed for simplicity. You can view the source code if you want a fuller picture.
- The syncVirtualServer, syncVirtualServerRoute, and syncTransportServer methods are similar to syncIngress. The other sync methods are different. However, those methods typically involve finding the affected Ingress, VirtualServer, and TransportServer resources and regenerating a configuration for them.
- The Workqueue has only a single worker thread that calls the sync method synchronously. This means that the control loop processes only one change at a time.
There are two additional helper components crucial for processing changes: Configuration and LocalSecretStore.
Configuration holds the latest valid state of the IC load balancing configuration resources: Ingresses, VirtualServers, VirtualServerRoutes, TransportServers, and GlobalConfiguration.
The Configuration supports add (for add/update) and delete operations on the resources. When you add/update/delete a resource in the Configuration, it performs the following:
- Validates the object (for add/update)
- Calculates the changes to the affected resources that are necessary to propagate to the NGINX config, returning the changes to the caller.
For example, when you add a new Ingress resource, the Configuration returns a change requiring the IC to add the configuration for that Ingress to the NGINX config files. Another example: if you make an existing Ingress resource invalid, the Configuration returns a change requiring the IC to remove the configuration for that Ingress from the NGINX config files.
Additionally, the Configuration ensures that only one Ingress/VirtualServer/TransportServer (TLS Passthrough) holds a particular host (for example, cafe.example.com) and only one TransportServer (TCP/UDP) holds a particular listener (for example, port 53 for UDP). This ensures that no host or listener collisions happen in the NGINX config.
Ultimately, the IC ensures the NGINX config on the filesystem reflects the state of the objects in the Configuration at any point in time.
LocalSecretStore (of the SecretStore interface) holds the valid Secret resources and keeps the corresponding files on the filesystem in sync with them. Secrets are used to hold TLS certificates and keys (type
kubernetes.io/tls), CAs (
nginx.org/ca), JWKs (
nginx.org/jwk), and client secrets for an OIDC provider (
When Controller processes a change to a configuration resource like Ingress, it creates an extended version of a resource that includes the dependencies – such as Secrets – necessary to generate the NGINX configuration. LocalSecretStore allows Controller to get a reference on the filesystem for a secret by the secret key (namespace/name).