Single Interlock deployment

When an application image is updated, the following actions occur:

  1. The service is updated with a new version of the application.

  2. The default “stop-first” policy stops the first replica before scheduling the second. The interlock proxies remove ip1.0 out of the backend pool as the app.1 task is removed.

  3. The first application task is rescheduled with the new image after the first task stops.

  4. The interlock proxy.1 is then rescheduled with the new NGINX configuration that contains the update for the new app.1 task.

  5. After proxy.1 is complete, proxy.2 redeploys with the updated ngnix configuration for the app.1 task.

  6. In this scenario, the amount of time that the service is unavailable is less than 30 seconds.

Optimizing Interlock for applications

Application update order

Swarm provides control over the order in which old tasks are removed while new ones are created. This is controlled on the service-level with --update-order.

  • stop-first (default)- Configures the currently updating task to stop before the new task is scheduled.

  • start-first - Configures the current task to stop after the new task has scheduled. This guarantees that the new task is running before the old task has shut down.

Use start-first if …

  • You have a single application replica and you cannot have service interruption. Both the old and new tasks run simultaneously during the update, but this ensurse that there is no gap in service during the update.

Use stop-first if …

  • Old and new tasks of your service cannot serve clients simultaneously.

  • You do not have enough cluster resourcing to run old and new replicas simultaneously.

In most cases, start-first is the best choice because it optimizes for high availability during updates.

Application update delay

Swarm services use update-delay to control the speed at which a service is updated. This adds a timed delay between application tasks as they are updated. The delay controls the time from when the first task of a service transitions to healthy state and the time that the second task begins its update. The default is 0 seconds, which means that a replica task begins updating as soon as the previous updated task transitions in to a healthy state.

Use update-delay if …

  • You are optimizing for the least number of dropped connections and a longer update cycle as an acceptable tradeoff.

  • Interlock update convergence takes a long time in your environment (can occur when having large amount of overlay networks).

Do not use update-delay if …

  • Service updates must occur rapidly.

  • Old and new tasks of your service cannot serve clients simultaneously.

Use application health checks

Swarm uses application health checks extensively to ensure that its updates do not cause service interruption. health-cmd can be configured in a Dockerfile or compose file to define a method for health checking an application. Without health checks, Swarm cannot determine when an application is truly ready to service traffic and will mark it as healthy as soon as the container process is running. This can potentially send traffic to an application before it is capable of serving clients, leading to dropped connections.

Application stop grace period

Use stop-grace-period to configure the maximum time period delay prior to force killing of the task (default: 10 seconds). In short, under the default setting a task can continue to run for no more than 10 seconds once its shutdown cycle has been initiated. This benefits applications that require long periods to process requests, allowing connection to terminate normally.

Interlock optimizations

Use service clusters for Interlock segmentation

Interlock service clusters allow Interlock to be segmented into multiple logical instances called “service clusters”, which have independently managed proxies. Application traffic only uses the proxies for a specific service cluster, allowing the full segmentation of traffic. Each service cluster only connects to the networks using that specific service cluster, which reduces the number of overlay networks to which proxies connect. Because service clusters also deploy separate proxies, this also reduces the amount of churn in LB configs when there are service updates.

Minimizing number of overlay networks

Interlock proxy containers connect to the overlay network of every Swarm service. Having many networks connected to Interlock adds incremental delay when Interlock updates its load balancer configuration. Each network connected to Interlock generally adds 1-2 seconds of update delay. With many networks, the Interlock update delay causes the LB config to be out of date for too long, which can cause traffic to be dropped.

Minimizing the number of overlay networks that Interlock connects to can be accomplished in two ways:

  • Reduce the number of networks. If the architecture permits it, applications can be grouped together to use the same networks.

  • Use Interlock service clusters. By segmenting Interlock, service clusters also segment which networks are connected to Interlock, reducing the number of networks to which each proxy is connected.

  • Use admin-defined networks and limit the number of networks per service cluster.

Use Interlock VIP Mode

VIP Mode can be used to reduce the impact of application updates on the Interlock proxies. It utilizes the Swarm L4 load balancing VIPs instead of individual task IPs to load balance traffic to a more stable internal endpoint. This prevents the proxy LB configs from changing for most kinds of app service updates reducing churn for Interlock. The following features are not supported in VIP mode:

  • Sticky sessions

  • Websockets

  • Canary deployments

The following features are supported in VIP mode:

  • Host & context routing

  • Context root rewrites

  • Interlock TLS termination

  • TLS passthrough

  • Service clusters

See also