Skip to main content

ECS agent stop

Last updated on

ECS agent stop is an AWS chaos fault that stops the ecs-agent process on every container instance registered to a target ECS cluster for a configurable duration. While the agent is down each container instance is reported as AGENT_DISCONNECTED and the ECS control plane can no longer place new tasks, send signals, or collect status from those instances. The tasks already running on the instances continue to run, but they become invisible to the scheduler until the agent is restored.

Use this fault to test how an ECS cluster behaves when the control plane temporarily loses contact with its data plane: whether the scheduler waits or fails over, whether a deployment in flight stalls, whether dependent CI or autoscaling logic times out gracefully, and whether the agent reconnects cleanly when the fault ends.

Run your first experiment

If you have not configured the chaos infrastructure yet, go to Quickstart to install the chaos infrastructure and run an experiment end to end.

Use cases​

Run this fault when you want to answer concrete questions like:

  • Scheduler resilience: When every agent in a cluster disconnects, does the ECS scheduler wait for reconnection, or does it move tasks to a different cluster?
  • In-flight deployment behaviour: If a service update is rolling out when agents disconnect, does the deployment stall safely and resume when agents return?
  • Container instance lifecycle: Do container instances reconnect cleanly when the agent is restored, and do running tasks resume reporting normally?
  • Operational alerting: Does your alerting fire when container instances report AGENT_DISCONNECTED, and is the runbook for "ECS agent down" still accurate?
  • CI / autoscaling backpressure: Does upstream CI or autoscaling that relies on ECS API responses degrade gracefully when the agent is unavailable, or does it cascade?

Prerequisites​

  • Kubernetes version: 1.21 or later for the chaos infrastructure cluster. Go to What's supported to confirm distribution support.
  • Target ECS cluster: CLUSTER_NAME exists in REGION and uses the EC2 launch type (this fault is not applicable to Fargate-only clusters, which have no agent process to stop).
  • Container instances are SSM-managed: Every EC2 container instance is registered with AWS Systems Manager (the SSM Agent is running and the host has an instance profile with SSM permissions). Go to Setting up Systems Manager to enable SSM.
  • AWS credentials available: Either an IAM role for service accounts (IRSA) bound to the chaos infrastructure service account, or an explicit role assumed via ASSUME_ROLE_ARN. This fault does not expose an AWS_AUTHENTICATION_SECRET tunable; if you need static credentials, configure them at the chaos infrastructure level.
  • IAM permissions granted: The credentials or role include the permissions listed below.

Supported environments​

PlatformSupport status
Amazon ECS on EC2 launch typeSupported
Amazon ECS on Fargate launch typeNot supported (no agent process to stop)
Linux container instancesSupported
Windows container instancesSupported
AWS regionsSupported in every commercial region; pass the region in REGION
Mixed clusters (EC2 + Fargate)Supported; only the EC2 container instances are affected

Permissions required​

The IAM principal that the chaos pod uses (IRSA on the chaos service account, or the role assumed via ASSUME_ROLE_ARN) needs the following AWS actions.

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "ecs:DescribeClusters",
        "ecs:ListContainerInstances",
        "ecs:DescribeContainerInstances"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "ssm:SendCommand",
        "ssm:CancelCommand",
        "ssm:GetCommandInvocation",
        "ssm:DescribeInstanceInformation"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "ec2messages:AcknowledgeMessage",
        "ec2messages:DeleteMessage",
        "ec2messages:FailMessage",
        "ec2messages:GetEndpoint",
        "ec2messages:GetMessages",
        "ec2messages:SendReply"
      ],
      "Resource": "*"
    }
  ]
}
  • ecs:DescribeClusters, ecs:ListContainerInstances, and ecs:DescribeContainerInstances are used to discover the EC2 hosts behind the cluster.
  • ssm:SendCommand and related SSM actions are used to stop the ecs-agent service on each host.
  • ec2messages:* are the legacy messaging permissions still required for SSM Run Command on older instance roles.

Go to common policy for all AWS faults to use a single superset IAM policy across every AWS fault.

Authentication​

The fault supports two credential delivery models. Pick one based on how your chaos infrastructure is deployed.

MethodWhen to use itHow to configure
IAM Roles for Service Accounts (IRSA)Chaos infrastructure runs in EKS and uses an OIDC-bound service accountNo tunable changes; the chaos pod inherits the role automatically. Go to AWS IAM integration to set it up
Assume roleThe fault needs to act in a different account or with elevated permissionsSet ASSUME_ROLE_ARN to the role ARN; the chaos pod assumes the role on top of its base credentials

If you need to inject static AWS credentials, configure them at the chaos infrastructure level (the chaos service account environment) rather than per fault, because this fault does not expose an AWS_AUTHENTICATION_SECRET tunable.

Fault tunables​

Configure the following fault parameters when you add ECS agent stop to an experiment in Chaos Studio. Defaults are shown for reference.

Required parameters

TunableDescriptionDefault
CLUSTER_NAMEName of the target ECS cluster.(required)
REGIONAWS region that hosts the ECS cluster (for example us-east-1).(required)

Chaos parameters

TunableDescriptionDefault
TOTAL_CHAOS_DURATIONDuration of the fault in seconds. The agent stays stopped for this period.30
SEQUENCEOrder in which multiple container instances are targeted: parallel stops the agent on all hosts at once; serial stops them one host at a time.parallel
RAMP_TIMEWait period in seconds before and after the fault. Go to ramp time to read how it is applied.0

Authentication

TunableDescriptionDefault
ASSUME_ROLE_ARNARN of an IAM role to assume on top of the base credentials. Leave empty to use the base credentials directly.""

Tunables that apply to every fault are documented in common tunables for all faults. AWS-specific shared tunables are documented in common AWS fault tunables.

Fault execution in brief​

Discovers every container instance in CLUSTER_NAME within REGION, stops the ecs-agent service on each underlying EC2 host for TOTAL_CHAOS_DURATION seconds, then restarts the agent and waits for each container instance to return to the ACTIVE connection state.

Expected behavior during fault execution​

  • Container instances transition from ACTIVE with agentConnected=true to ACTIVE with agentConnected=false (also reported in the ECS console as AGENT_DISCONNECTED).
  • Already-running tasks continue to run on each container instance because the workload itself is unaffected; only the control-plane channel is severed.
  • New RunTask or StartTask calls targeting these container instances fail or queue, because the scheduler cannot reach the agent.
  • A service deployment that was already rolling out pauses; deployments that begin during the fault are blocked.
  • ECS CloudWatch metrics (for example ContainerInstance health and Service event counts) reflect the disconnection.
When the fault ends

The chaos pod issues a start of the ecs-agent service on every host and polls until each container instance reports agentConnected=true again. The scheduler resumes placement automatically; no manual intervention is needed.

Signals to watch​

Attach resilience probes to assert each layer:

  • Agent connectivity: Use a command probe that runs aws ecs describe-container-instances --cluster <cluster> --container-instances <arn> and asserts on agentConnected.
  • Scheduler responsiveness: Use a command probe that calls aws ecs run-task against the cluster and asserts the call fails during the fault and succeeds afterwards.
  • In-flight deployments: Use a Prometheus probe on aws_ecs_running_task_count or aws_ecs_pending_task_count to confirm the deployment paused and recovered.
  • Application availability: Use an HTTP probe against the service endpoint to confirm running tasks continue to serve traffic during the fault.

Verify the fault execution effect​

While the experiment is running, confirm the agent is disconnected and then reconnected:

  1. Inspect container instance connectivity.

    aws ecs list-container-instances --cluster <cluster> --region <region>
    aws ecs describe-container-instances \
      --cluster <cluster> \
      --region <region> \
      --container-instances <arn-1> <arn-2> \
      --query "containerInstances[].[ec2InstanceId,agentConnected]"

    During the fault agentConnected should be false. After TOTAL_CHAOS_DURATION it should return to true.

  2. Confirm tasks keep running.

    aws ecs list-tasks --cluster <cluster> --region <region> --desired-status RUNNING

    The running task list should remain stable during the fault.

  3. Verify the scheduler regains placement after recovery.

    aws ecs run-task \
      --cluster <cluster> \
      --task-definition <task-def> \
      --region <region>

    This call should succeed only after the agent has reconnected.

Recovery and cleanup​

  • End of duration: The chaos pod starts the ecs-agent service on each container instance and polls until every host reports agentConnected=true.
  • Abort the experiment: Stopping the experiment from Chaos Studio triggers the same start call.
  • Manual recovery: If the fault exits before issuing the start call, restart the agent on each affected host with sudo systemctl start ecs (Amazon Linux 2 / Amazon Linux 2023) or sudo start ecs (older AMIs). The agent reconnects automatically once running.
  • Workload recovery: Running tasks should not need recovery because they were never interrupted; verify deployment status if a rollout was in flight.

Limitations​

  • Fargate is not affected: Fargate tasks have no exposed agent process; this fault is meaningful only on container instances under the EC2 launch type.
  • All container instances in the cluster are affected: The fault does not filter by tag, AZ, or service. Use a dedicated test cluster or a serial sequence if you need a smaller blast radius.
  • SSM-managed hosts only: Container instances that are not registered with SSM cannot be reached; the fault skips them and reports the host as unaffected.
  • Already-running tasks keep running: This fault tests scheduler / control-plane resilience, not workload availability. Pair it with a workload-level fault if you also want to test task-level disruption.

Troubleshooting​

ECS agent stop experiment fails with AccessDeniedException in Harness Chaos Engineering

The credentials supplied to the chaos pod do not have the required ECS and SSM permissions. Confirm the IAM policy attached to the IRSA role (or the role assumed via ASSUME_ROLE_ARN) includes ecs:DescribeClusters, ecs:ListContainerInstances, ecs:DescribeContainerInstances, ssm:SendCommand, ssm:GetCommandInvocation, and ssm:DescribeInstanceInformation. When using ASSUME_ROLE_ARN, also confirm the source identity is trusted to assume the role.

ECS agent stop reports container instance is not SSM-managed

The fault uses AWS Systems Manager Run Command to stop the agent process; the underlying EC2 container instance must be registered with SSM. Confirm the SSM Agent is installed and running on the host, the host has an instance profile with AmazonSSMManagedInstanceCore (or equivalent), and the host appears in 'aws ssm describe-instance-information'. After enabling SSM the instance may take a few minutes to register.

ECS agent stop completes but container instance still shows agentConnected=true

The agent restart was so fast that the disconnection window was shorter than the ECS heartbeat interval (typically 30 seconds). Increase TOTAL_CHAOS_DURATION to at least 60 seconds and rerun the experiment. You can also confirm the SSM command actually ran by inspecting the command invocation in the AWS console under Systems Manager → Run Command.

Running tasks failed during ECS agent stop and did not recover

Tasks should not be terminated by this fault. If tasks died during the experiment, the most likely cause is a health check on the service that flapped because the deployment controller could not reach the agent; the service then replaced tasks once the agent reconnected. Verify the service events with 'aws ecs describe-services' and consider lowering deployment health check sensitivity or pairing this fault with longer health-check grace periods.