Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

App Mesh best practices

To achieve the goal of zero failed requests during planned deployments and during the unplanned loss of some hosts, the best practices in this topic implement the following strategy:

To significantly reduce or eliminate failures, we recommend that you implement the recommendations in all of the following practices.

Instrument all routes with retries

Configure all virtual services to use a virtual router and set a default retry policy for all routes. This will mitigate failed requests by reselecting a host and sending a new request. For retry policies, we recommend a value of at least two for maxRetries, and specifying the following options for each type of retry event in each route type that supports the retry event type:

Other retry events need to be considered on a case-by-case basis as they may not be safe, such as if the request isn’t idempotent. You will need to consider and test values for maxRetries and perRetryTimeout that make the appropriate trade off between the maximum latency of a request (maxRetries * perRetryTimeout) versus the increased success rate of more retries. Additionally, when Envoy attempts to connect to an endpoint that is no longer present, you should expect that request to consume the full perRetryTimeout. To configure a retry policy, see Creating a route and then select the protocol that you want to route.

Adjust deployment velocity

When using rolling deployments, reduce the overall deployment velocity. By default, Amazon ECS configures a deployment strategy of a minimum of 100 percent healthy tasks and 200 percent total tasks. On deployment, this results in two points of high drift:

When configured with these deployment constraints, container orchestrators may enter a state where they are simultaneously hiding all old destinations and making all new destinations visible. Because your Envoy dataplane is eventually consistent, this can result in periods where the set of destinations visible in your dataplane have diverged from the orchestrator’s point of view. To mitigate this, we recommend maintaining a minimum of 100 percent healthy tasks, but lowering total tasks to 125 percent. This will reduce divergence and improve the reliability of retries. We recommend the following settings for different container runtimes:

Amazon ECS

If your service has a desired count of two or three, set maximumPercent to 150 percent. Otherwise, set maximumPercent to 125 percent.

Kubernetes

Configure your deployment's update strategy, setting maxUnavailable to 0 percent and maxSurge to 25 percent. For more information on deployments, see Kubernetes Deployments documentation.

Scale out before scale in

Scale out and scale in can both result in some probability of failed requests in retries. While there are task recommendations that mitigate scale out, the only recommendation for scale in is to minimize the percentage of scaled in tasks at any one time. We recommend that you use a deployment strategy that scales out new Amazon ECS tasks or Kubernetes deployments prior to scaling in old tasks or deployments. This scaling strategy keeps your percentage of scaled in tasks or deployments lower, while maintaining the same velocity. This practice applies to both Amazon ECS tasks and Kubernetes deployments.

Implement container health checks

In the scale up scenario, containers in an Amazon ECS task may come up out of order and may not be initially responsive. We recommend the following suggestions for different container runtimes:

Amazon ECS

To mitigate this, we recommend using container health checks and container dependency ordering to ensure that Envoy is running and healthy prior to any containers requiring outbound network connectivity starting. To correctly configure an application container and Envoy container in a task definition, see Container dependency.

Kubernetes

None, because Kubernetes liveness and readiness probes are not being considered in registration and de-registration of Amazon Cloud Map instances in the App Mesh controller for Kubernetes. For more information, see GitHub issue #132.

Optimize DNS resolution

If you're using DNS for service discovery, it's essential to select the appropriate IP protocol to optimize DNS resolution when configuring your meshes. App Mesh supports both IPv4 and IPv6, and your choice can impact your service's performance and compatibility. If your infrastructure doesn't support IPv6, we recommended you specify an IP setting that aligns with your infrastructure rather than relying on the default IPv6_PREFERRED behavior. The default IPv6_PREFERRED behavior can degrade service performance.

You can set IP version preferences at both the mesh level and the virtual node level, with virtual node settings overriding those at the mesh level.

For more information, see Service Meshes and Virtual Nodes.