This section describes the common reload and relaunch scenarios for Edge-LB pools and how to restore load balancing for a service after a pod failure.

Before you begin

Before you create Edge-LB pools and pool configuration files, you should have DC/OS™ Enterprise cluster nodes installed and ready to use, and have previously downloaded and installed the latest Edge-LB packages.

You must have:

• Edge-LB installed as described in the Edge-LB installation instructions.
• The core DC/OS command-line interface (CLI) installed and configured to communicate with the DC/OS cluster.
• The edgelb command-line interface (CLI) installed.
• An active and properly-configured DC/OS Enterprise cluster, with at least one DC/OS private agent node to run the load-balanced service and at least one DC/OS public agent node for exposing the load-balanced service.

For information about installing Edge-LB packages, see the installation instructions.

Normal reload scenario

A change to a service (such as scaling up) that is load balanced by a pool, automatically triggers a reload of the pool configuration file, and restarts the load balancer instances for that service.

During this type of normal reload scenario, you can expect the following behavior:

• No traffic is dropped unless the service instance that was serving the request was killed.

• The load balancer will keep any existing long-running connections intact.

• All the new connections will be proxied using the new reloaded configuration.

This type of reload occurs at most once every 3 seconds. Normal reload operations play a key role in enabling zero-downtime deployment strategies like the one described in the Blue/Green service deployment tutorial.

Load balancer relaunch scenario

A change to the load balancer pool (such as adding a secret) triggers a relaunch of all load balancers in the pool.

During this type of load balancer relaunch scenario, you can expect the following behavior:

• Traffic currently flowing through the given load balancer is dropped when it is stopped. To minimize the impact of stopping and restarting a load balancer instance, you should run more than one load balancer instance within the pool.

• Only one load balancer is stopped in the pool during the update at a time.

• The load balancer will be relaunched on the same node (unless the node itself has failed).

Replacing a failed pod

By default, Edge-LB load balancer instances are tied to a given node. If the node goes down, Edge-LB does not automatically relocate the pod containing the Edge-LB load balancer instance to a new node. You must issue a pod replace command to the pool scheduler to tell it to start the load balancer instance on a new node. If a machine hosting a pod is permanently lost, manual intervention is required to discard the missing pod and start it on a new node.

You can perform the necessary steps using the edgelb-pool command-line interface subcommand. The edgelb-pool command-line interface is distinct from the edgelb subcommand that is installed with the Edge-LB package. You must install the edgelb-pool command-line interface separately if you have not previously installed this package.

1. Install the edgelb-pool CLI subcommand:

   dcos package install edgelb-pool --cli --yes
   

2. Get the name of the pool that owns the pod you need to relocate:

   dcos edgelb list
   

   This command lists all of the Edge-LB pool configurations you have deployed. The pool that has a missing pod in this list identifies the <pool-name> you specify in the next step.

3. Get the name of the pod you need to replace (the pod that was running on the removed public agent):

   dcos edgelb-pool --name=/dcos-edgelb/pools/<pool-name> pod list
   

   This command returns the pod name you will specify for the <pod-id> in the next step.

4. Replace the <pod-id> that is no longer running using the pod replace command as follow:

   dcos edgelb-pool --name=/dcos-edgelb/pools/<pool-name> pod replace <pod-id>
   

   This command destroys the pool server and relaunchs a new Edge-LB pool on the new public agent.