Using the Default HAProxy Router

The oc adm router command is provided with the administrator CLI to simplify the tasks of setting up routers in a new installation. The oc adm router command creates the service and deployment configuration objects. Use the --service-account option to specify the service account the router will use to contact the master.

The router service account can be created in advance or created by the oc adm router --service-account command.

Every form of communication between OKD components is secured by TLS and uses various certificates and authentication methods. The --default-certificate .pem format file can be supplied or one is created by the oc adm router command. When routes are created, the user can provide route certificates that the router will use when handling the route.

Routers are deployed on specific nodes. This makes it easier for the cluster administrator and external network manager to coordinate which IP address will run a router and which traffic the router will handle. The routers are deployed on specific nodes by using node selectors.

If the router does not exist, run the following to create a router:

--replicas is usually 1 unless a high availability configuration is being created.

--extended-logging=true configures the router to forward logs that are generated by HAProxy to the syslog container.

To find the host IP address of the router:

You can also use router shards to ensure that the router is filtered to specific namespaces or routes, or set any environment variables after router creation. In this case create a router for each shard.

To check if a default router, named router, already exists:

The following example is the configuration for a router that uses --extended-logging=true:

Use the following commands to view the HAProxy logs:

The HAProxy logs take the following form:

For example, if you want to create a router named router and have it placed on a node labeled with node-role.kubernetes.io/infra=true:

During cluster installation, the openshift_router_selector and openshift_registry_selector Ansible settings are set to node-role.kubernetes.io/infra=true by default. The default router and registry will only be automatically deployed if a node exists that matches the node-role.kubernetes.io/infra=true label.

For information on updating labels, see Updating Labels on Nodes.

Multiple instances are created on different hosts according to the scheduler policy.

For example:

Using the ROUTE_LABELS environment variable, you can filter routes so that they are used only by specific routers.

For example, if you have multiple routers, and 100 routes, you can attach labels to the routes so that a portion of them are handled by one router, whereas the rest are handled by another.

When the router is created, the --max-connections= option sets the desired limit:

Edit the ROUTER_MAX_CONNECTIONS environment variable in the router’s deployment configuration to change the value. The router pods are restarted with the new value. If ROUTER_MAX_CONNECTIONS is not present, the default value of 20000, is used.

The HAProxy strict-sni can be controlled through the ROUTER_STRICT_SNI environment variable in the router’s deployment configuration. It can also be set when the router is created by using the --strict-sni command line option.

Set the router cipher suite using the --ciphers option when creating a router:

The values are: modern, intermediate, or old, with intermediate as the default. Alternatively, a set of ":" separated ciphers can be provided. The ciphers must be from the set displayed by:

Alternatively, use the ROUTER_CIPHERS environment variable for an existing router.

Client access to the router and the backend services can be restricted using mutual TLS authentication. The router will reject requests from clients not in its authenticated set. Mutual TLS authentication is implemented on client certificates and can be controlled based on the certifying authorities (CAs) issuing the certificates, the certificate revocation list and/or any certificate subject filters. Use the mutual tls config options --mutual-tls-auth, --mutual-tls-auth-ca, --mutual-tls-auth-crl and --mutual-tls-auth-filter when creating a router:

The --mutual-tls-auth values are required, optional, or none, with none as the default. The --mutual-tls-auth-ca value specifies a file containing one or more CA certificates. These CA certificates are used by the router to verify a client’s certificate.

The --mutual-tls-auth-crl can be used specify the certificate revocation list to handle cases where certificates (issued by valid certifying authorities) have been revoked.

The --mutual-tls-auth-filter value can be used for fine grain access control based on the certificate subject. The value is a regular expression, which is to used to match up the certificate’s subject.

Using --mutual-tls-auth=required ensures that you only allow authenticated clients access to the backend resources. This means that the client is always required to provide authentication information (aka a client certificate). To make the mutual TLS authentication optional, use --mutual-tls-auth=optional (or use none to disable it - this is the default). Note here that optional means that you do not require a client to present any authentication information and if the client provides any authentication information, that is just passed on to the backend in the X-SSL* HTTP headers.

When mutual TLS authentication support is enabled (either using the required or optional value for the --mutual-tls-auth flag), the client authentication information is passed to the backend in the form of X-SSL* HTTP headers.

Examples of the X-SSL* HTTP headers X-SSL-Client-DN: the full distinguished name (DN) of the certificate subject. X-SSL-Client-NotBefore: the client certificate start date in YYMMDDhhmmss[Z] format. X-SSL-Client-NotAfter: the client certificate end date in YYMMDDhhmmss[Z] format. X-SSL-Client-SHA1: the SHA-1 fingerprint of the client certificate. X-SSL-Client-DER: provides full access to the client certificate. Contains the DER formatted client certificate encoded in base-64 format.

You can set up a highly-available router on your OKD cluster using IP failover. This setup has multiple replicas on different nodes so the failover software can switch to another replica if the current one fails.

You can customize the service ports that a template router binds to by setting the environment variables ROUTER_SERVICE_HTTP_PORT and ROUTER_SERVICE_HTTPS_PORT. This can be done by creating a template router, then editing its deployment configuration.

The following example creates a router deployment with 0 replicas and customizes the router service HTTP and HTTPS ports, then scales it appropriately (to 1 replica).

The following is an example using iptables to open the custom router service ports.

An administrator can create multiple routers with the same definition to serve the same set of routes. Each router will be on a different node and will have a different IP address. The network administrator will need to get the desired traffic to each node.

Multiple routers can be grouped to distribute routing load in the cluster and separate tenants to different routers or shards. Each router or shard in the group admits routes based on the selectors in the router. An administrator can create shards over the whole cluster using ROUTE_LABELS. A user can create shards over a namespace (project) by using NAMESPACE_LABELS.

Making specific routers deploy on specific nodes requires two steps:

Router sharding uses NAMESPACE_LABELS and ROUTE_LABELS, to filter router namespaces and routes. This enables you to distribute subsets of routes over multiple router deployments. By using non-overlapping subsets, you can effectively partition the set of routes. Alternatively, you can define shards comprising overlapping subsets of routes.

By default, a router selects all routes from all projects (namespaces). Sharding involves adding labels to routes or namespaces and label selectors to routers. Each router shard comprises the routes that are selected by a specific set of label selectors or belong to the namespaces that are selected by a specific set of label selectors.

Router Sharding and DNS

Because an external DNS server is needed to route requests to the desired shard, the administrator is responsible for making a separate DNS entry for each router in a project. A router will not forward unknown routes to another router.

Consider the following example:

Separate DNS entries must resolve *.foo.com to the node hosting Router A and *.example.com to the node hosting Router B:

Router Sharding Examples

This section describes router sharding using namespace and route labels.

Adding this label makes the route available in the router.

NAMESPACE_LABELS allows filtering of the projects to service and selecting all the routes from those projects, but you may want to partition routes based on other criteria associated with the routes themselves. The ROUTE_LABELS selector allows you to slice-and-dice the routes themselves.

When exposing a service, a user can use the same route from the DNS name that external users use to access the application. The network administrator of the external network must make sure the host name resolves to the name of a router that has admitted the route. The user can set up their DNS with a CNAME that points to this host name. However, the user may not know the host name of the router. When it is not known, the cluster administrator can provide it.

The cluster administrator can use the --router-canonical-hostname option with the router’s canonical host name when creating the router. For example:

This creates the ROUTER_CANONICAL_HOSTNAME environment variable in the router’s deployment configuration containing the host name of the router.

For routers that already exist, the cluster administrator can edit the router’s deployment configuration and add the ROUTER_CANONICAL_HOSTNAME environment variable:

The ROUTER_CANONICAL_HOSTNAME value is displayed in the route status for all routers that have admitted the route. The route status is refreshed every time the router is reloaded.

When a user creates a route, all of the active routers evaluate the route and, if conditions are met, admit it. When a router that defines the ROUTER_CANONICAL_HOSTNAME environment variable admits the route, the router places the value in the routerCanonicalHostname field in the route status. The user can examine the route status to determine which, if any, routers have admitted the route, select a router from the list, and find the host name of the router to pass along to the network administrator.

oc describe inclues the host name when available:

Using the above information, the user can ask the DNS administrator to set up a CNAME from the route’s host, hello.in.mycloud.com, to the router’s canonical hostname, rtr.example.com. This results in any traffic to hello.in.mycloud.com reaching the user’s application.

You can customize the suffix used as the default routing subdomain for your environment by modifying the master configuration file (the /etc/origin/master/master-config.yaml file by default). Routes that do not specify a host name would have one generated using this default routing subdomain.

The following example shows how you can set the configured suffix to v3.openshift.test:

With the OKD master(s) running the above configuration, the generated host name for the example of a route named no-route-hostname without a host name added to a namespace mynamespace would be:

If an administrator wants to restrict all routes to a specific routing subdomain, they can pass the --force-subdomain option to the oc adm router command. This forces the router to override any host names specified in a route and generate one based on the template provided to the --force-subdomain option.

The following example runs a router, which overrides the route host names using a custom subdomain template ${name}-${namespace}.apps.example.com.

A TLS-enabled route that does not include a certificate uses the router’s default certificate instead. In most cases, this certificate should be provided by a trusted certificate authority, but for convenience you can use the OKD CA to create the certificate. For example:

The router expects the certificate and key to be in PEM format in a single file:

From there you can use the --default-cert flag:

To manually redeploy the router certificates:

Currently, password protected key files are not supported. HAProxy prompts for a password upon starting and does not have a way to automate this process. To remove a passphrase from a keyfile, you can run:

Here is an example of how to use a secure edge terminated route with TLS termination occurring on the router before traffic is proxied to the destination. The secure edge terminated route specifies the TLS certificate and key information. The TLS certificate is served by the router front end.

First, start up a router instance:

Next, create a private key, csr and certificate for our edge secured route. The instructions on how to do that would be specific to your certificate authority and provider. For a simple self-signed certificate for a domain named www.example.test, see the example shown below:

Generate a route using the above certificate and key.

Look at its definition.

Make sure your DNS entry for www.example.test points to your router instance(s) and the route to your domain should be available. The example below uses curl along with a local resolver to simulate the DNS lookup:

The HAProxy router has support for wildcard routes, which are enabled by setting the ROUTER_ALLOW_WILDCARD_ROUTES environment variable to true. Any routes with a wildcard policy of Subdomain that pass the router admission checks will be serviced by the HAProxy router. Then, the HAProxy router exposes the associated service (for the route) per the route’s wildcard policy.

Learn how to configure the web console for wildcard routes.

This example reflects TLS termination occurring on the router before traffic is proxied to the destination. Traffic sent to any hosts in the subdomain example.org (*.example.org) is proxied to the exposed service.

The secure edge terminated route specifies the TLS certificate and key information. The TLS certificate is served by the router front end for all hosts that match the subdomain (*.example.org).

For routers that allow wildcard routes (ROUTER_ALLOW_WILDCARD_ROUTES set to true), there are some caveats to the ownership of a subdomain associated with a wildcard route.

Prior to wildcard routes, ownership was based on the claims made for a host name with the namespace with the oldest route winning against any other claimants. For example, route r1 in namespace ns1 with a claim for one.example.test would win over another route r2 in namespace ns2 for the same host name one.example.test if route r1 was older than route r2.

In addition, routes in other namespaces were allowed to claim non-overlapping hostnames. For example, route rone in namespace ns1 could claim www.example.test and another route rtwo in namespace d2 could claim c3po.example.test.

This is still the case if there are no wildcard routes claiming that same subdomain (example.test in the above example).

However, a wildcard route needs to claim all of the host names within a subdomain (host names of the form \*.example.test). A wildcard route’s claim is allowed or denied based on whether or not the oldest route for that subdomain (example.test) is in the same namespace as the wildcard route. The oldest route can be either a regular route or a wildcard route.

For example, if there is already a route eldest that exists in the ns1 namespace that claimed a host named owner.example.test and, if at a later point in time, a new wildcard route wildthing requesting for routes in that subdomain (example.test) is added, the claim by the wildcard route will only be allowed if it is the same namespace (ns1) as the owning route.

The following examples illustrate various scenarios in which claims for wildcard routes will succeed or fail.

In the example below, a router that allows wildcard routes will allow non-overlapping claims for hosts in the subdomain example.test as long as a wildcard route has not claimed a subdomain.

In the example below, a router that allows wildcard routes will not allow the claim for owner.example.test or aname.example.test to succeed since the owning namespace is ns1.

In the example below, a router that allows wildcard routes will allow the claim for `\*.example.test to succeed since the owning namespace is ns1 and the wildcard route belongs to that same namespace.

In the example below, a router that allows wildcard routes will not allow the claim for `\*.example.test to succeed since the owning namespace is ns1 and the wildcard route belongs to another namespace cyclone.

Similarly, once a namespace with a wildcard route claims a subdomain, only routes within that namespace can claim any hosts in that same subdomain.

In the example below, once a route in namespace ns1 with a wildcard route claims subdomain example.test, only routes in the namespace ns1 are allowed to claim any hosts in that same subdomain.

In the example below, different scenarios are shown, in which the owner routes are deleted and ownership is passed within and across namespaces. While a route claiming host eldest.example.test in the namespace ns1 exists, wildcard routes in that namespace can claim subdomain example.test. When the route for host eldest.example.test is deleted, the next oldest route senior.example.test would become the oldest route and would not affect any other routes. Once the route for host senior.example.test is deleted, the next oldest route junior.example.test becomes the oldest route and block the wildcard route claimant.

The OKD router runs inside a container and the default behavior is to use the network stack of the host (i.e., the node where the router container runs). This default behavior benefits performance because network traffic from remote clients does not need to take multiple hops through user space to reach the target service and container.

Additionally, this default behavior enables the router to get the actual source IP address of the remote connection rather than getting the node’s IP address. This is useful for defining ingress rules based on the originating IP, supporting sticky sessions, and monitoring traffic, among other uses.

This host network behavior is controlled by the --host-network router command line option, and the default behaviour is the equivalent of using --host-network=true. If you wish to run the router with the container network stack, use the --host-network=false option when creating the router. For example:

Internally, this means the router container must publish the 80 and 443 ports in order for the external network to communicate with the router.

You can configure the HAProxy router to support the dynamic configuration manager.

The dynamic configuration manager brings certain types of routes online without requiring HAProxy reload downtime. It handles any route and endpoint life-cycle events such as route and endpoint addition|deletion|update.

Enable the dynamic configuration manager by setting the ROUTER_HAPROXY_CONFIG_MANAGER environment variable to true:

If the dynamic configuration manager cannot dynamically configure HAProxy, it rewrites the configuration and reloads the HAProxy process. For example, if a new route contains custom annotations, such as custom timeouts, or if the route requires custom TLS configuration.

The dynamic configuration internally uses the HAProxy socket and configuration API with a pool of pre-allocated routes and back end servers. The pre-allocated pool of routes is created using route blueprints. The default set of blueprints supports unsecured routes, edge secured routes without any custom TLS configuration, and passthrough routes.

The following procedure assumes you have created three route objects: reencrypt-blueprint, annotated-edge-blueprint, and annotated-unsecured-blueprint. See Route Types for an example of the different route type objects.

The HAProxy router metrics are, by default, exposed or published in Prometheus format for consumption by external metrics collection and aggregation systems (e.g. Prometheus, statsd). Metrics are also available directly from the HAProxy router in its own HTML format for viewing in a browser or CSV download. These metrics include the HAProxy native metrics and some controller metrics.

When you create a router using the following command, OKD makes metrics available in Prometheus format on the stats port, by default 1936.

In OKD clusters with large numbers of routes (greater than the value of net.ipv4.neigh.default.gc_thresh3, which is 65536 by default), you must increase the default values of sysctl variables on each node in the cluster running the router pod to allow more entries in the ARP cache.

When the problem is occuring, the kernel messages would be similar to the following:

When this issue occurs, the oc commands might start to fail with the following error:

To verify the actual amount of ARP entries for IPv4, run the following:

If the number begins to approach the net.ipv4.neigh.default.gc_thresh3 threshold, increase the values. Get the current value by running:

The following sysctl sets the variables to the OKD current default values.

To make these settings permanent, see this document.

Add timeout http-request to the default HAProxy router image to protect the deployment against distributed denial-of-service (DDoS) attacks (for example, slowloris):

Also, when the environment variable ROUTER_SLOWLORIS_TIMEOUT is set, it limits the amount of time a client has to send the whole HTTP request. Otherwise, HAProxy will shut down the connection.

Setting the environment variable allows information to be captured as part of the router’s deployment configuration and does not require manual modification of the template, whereas manually adding the HAProxy setting requires you to rebuild the router pod and maintain your router template file.

Using annotations implements basic DDoS protections in the HAProxy template router, including the ability to limit the:

These are enabled on a per route basis because applications can have extremely different traffic patterns.

Enabled threading with the --threads flag. This flag specifies the number of threads that the HAProxy router will use.