Skip to main content

ECS update container timeout

Last updated on

ECS update container timeout is an AWS chaos fault that registers a new revision of the target service's task definition with the container startTimeout set to START_TIMEOUT seconds and stopTimeout set to STOP_TIMEOUT seconds, updates the service to that revision for a configurable duration, then re-registers and restores the original revision. The chaos timeouts apply to new tasks scheduled during the chaos window: ECS gives them only START_TIMEOUT seconds to become healthy, and only STOP_TIMEOUT seconds to drain on SIGTERM before being killed.

Use this fault to test how a workload behaves when ECS does not wait long enough for containers to start (cold starts, dependency-readiness probes) or to drain (in-flight requests, connection close, message commits), and to validate that your application's startup and shutdown handlers are correctly sized.

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:

  • Cold-start failures: When startTimeout is shorter than the application's startup time, do tasks fail health checks and get killed, and does the service deployment surface this clearly?
  • Hard drain: When stopTimeout is shorter than in-flight request duration, do clients see resets, and do dependent systems (queues, databases) handle the partial commits?
  • Sidecar startup ordering: Do dependent containers in the task (proxies, log shippers) survive a shorter startTimeout, or do they cause the task to fail entirely?
  • Deployment safety: Does the deployment circuit breaker catch repeated start failures and roll back automatically?
  • Recovery from configuration drift: If someone accidentally lowered the timeouts in production, would alarms catch it, and is the fix simple?

Prerequisites​

  • Kubernetes version: 1.21 or later for the chaos infrastructure cluster. Go to What's supported to confirm distribution support.
  • Target ECS service: SERVICE_NAME exists in CLUSTER_NAME within REGION and has at least one running task.
  • AWS credentials available: Either an AWS credentials file uploaded as a File Secret in Harness Secret Manager (see Authentication below) or an IAM role for service accounts (IRSA) bound to the chaos infrastructure service account.
  • IAM permissions granted: The credentials or role include the permissions listed below, including iam:PassRole.

Supported environments​

PlatformSupport status
Amazon ECS on EC2 launch typeSupported
Amazon ECS on Fargate launch typeSupported (timeouts are bounded by the Fargate platform; values above platform limits are silently clamped)
Tasks with sidecar containersSupported (every container gets the new startTimeout/stopTimeout)
AWS regionsSupported in every commercial region; pass the region in REGION

Permissions required​

The IAM principal that the chaos pod uses (the credentials mounted from the Harness Secret Manager file secret, the IRSA role 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:DescribeServices",
        "ecs:DescribeTaskDefinition",
        "ecs:RegisterTaskDefinition",
        "ecs:DeregisterTaskDefinition",
        "ecs:UpdateService",
        "ecs:ListTasks",
        "ecs:DescribeTasks"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "iam:PassRole"
      ],
      "Resource": "*"
    }
  ]
}

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

Authentication​

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

MethodWhen to use itHow to configure
Harness Secret Manager file secretChaos infrastructure runs outside EKS, or you want explicit static credentialsUpload the AWS credentials file as a File Secret in Harness Secret Manager and reference its identifier via AWS_AUTHENTICATION_SECRET
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

When using the Harness Secret Manager method, the contents of the File Secret should be the AWS credentials file in the standard ~/.aws/credentials format:

[default]
aws_access_key_id = REPLACE_WITH_ACCESS_KEY_ID
aws_secret_access_key = REPLACE_WITH_SECRET_ACCESS_KEY

Upload this file as a File Secret in Harness Secret Manager (Project Setup → Secrets → New File Secret), and pass the secret identifier in AWS_AUTHENTICATION_SECRET when configuring the fault.

Go to AWS named profile for chaos to switch between profiles inside a single credentials file.

Fault tunables​

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

Required parameters

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

Chaos parameters

TunableDescriptionDefault
START_TIMEOUTContainer startTimeout (in seconds) to apply during the chaos window. ECS kills the task if any container does not reach HEALTHY within this window.30
STOP_TIMEOUTContainer stopTimeout (in seconds) to apply during the chaos window. ECS waits this long after SIGTERM before sending SIGKILL.30
TASK_REPLICA_AFFECTED_PERCPercentage of task replicas to affect through the deployment.100
TOTAL_CHAOS_DURATIONDuration of the fault in seconds. The service runs on the chaos revision for this period.30
CHAOS_INTERVALDelay in seconds between successive iterations when running for more than one cycle.30
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.""
AWS_AUTHENTICATION_SECRETIdentifier of the File Secret in Harness Secret Manager that contains the AWS credentials file. Not required when using IRSA.""

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​

Reads the current task definition for SERVICE_NAME, registers a new revision with each container's startTimeout and stopTimeout set to START_TIMEOUT and STOP_TIMEOUT, updates the service to that revision, waits for TOTAL_CHAOS_DURATION seconds, then restores the original revision and deregisters the chaos revision.

Expected behavior during fault execution​

  • A new task definition revision is registered with the chaos timeouts.
  • The service deploys new tasks on the chaos revision; old tasks drain according to the deployment configuration. If STOP_TIMEOUT is shorter than the old stopTimeout, the draining is harsher.
  • New tasks that cannot reach HEALTHY within START_TIMEOUT are killed by ECS; the service deployment may stall or roll back.
  • Stopping tasks receive SIGTERM, get STOP_TIMEOUT seconds to drain, then receive SIGKILL. Long-lived requests in flight may be reset.
  • ECS service events show "task did not transition to RUNNING state in time" and "task started" / "task stopped" cycles.
When the fault ends

The chaos pod restores the original task definition revision and waits for the rollout to begin. Subsequent tasks use the original timeouts; running tasks on the chaos revision drain as the rollout completes.

Signals to watch​

Attach resilience probes to assert each layer:

  • Application availability: Use an HTTP probe against the service endpoint to detect downtime during the chaos window.
  • Task start failures: Use a Prometheus probe on aws_ecs_pending_task_count to detect a build-up of failed start attempts.
  • Deployment status: Use a command probe that runs aws ecs describe-services --query "services[].deployments" and asserts the deployment circuit breaker fires when START_TIMEOUT is too short.
  • Request resets: Use a Prometheus probe on application-level error counters (HTTP 499/502, custom client-reset counters) to detect aggressive drains.

Verify the fault execution effect​

While the experiment is running, confirm the timeouts change and then restore:

  1. Inspect the active task definition.

    aws ecs describe-task-definition \
      --task-definition <active-arn> \
      --region <region> \
      --query "taskDefinition.containerDefinitions[].[name,startTimeout,stopTimeout]"

    During the fault these should equal START_TIMEOUT and STOP_TIMEOUT. After recovery they should return to the original values.

  2. Inspect deployment events.

    aws ecs describe-services \
      --cluster <cluster> \
      --services <service> \
      --region <region> \
      --query "services[].events[0:10]"

    Look for task-start-timeout events during the chaos window if START_TIMEOUT is too short.

  3. Inspect stopped tasks.

    aws ecs list-tasks --cluster <cluster> --service-name <service> --desired-status STOPPED --region <region>
    aws ecs describe-tasks --cluster <cluster> --tasks <arn> --region <region> \
      --query "tasks[].[stoppedReason,containers[0].reason]"

    The reason should indicate timeout-driven termination.

Recovery and cleanup​

  • End of duration: The chaos pod restores the original revision and deregisters the chaos revision.
  • Abort the experiment: Stopping the experiment from Chaos Studio also triggers the restore-and-deregister cleanup.
  • Manual recovery: If the fault exits before cleanup, set the service back to the original revision manually with aws ecs update-service --task-definition <original-arn>.
  • Workload recovery: Requests reset during the harsh drain are not replayed by the fault; the client is responsible for retrying.

Limitations​

  • Fargate platform clamps: On Fargate, startTimeout and stopTimeout are bounded by the platform version's maximums; values above the cap are silently clamped.
  • Existing tasks keep original timeouts: Running tasks are not retroactively retimed; only new tasks scheduled during the chaos window get the new timeouts.
  • Health checks must be defined: startTimeout is enforced only when a container has a healthCheck configured. Containers without health checks ignore the value.
  • Single service per experiment: Use multiple experiments to target multiple services.
  • Cross-region targeting: A single experiment targets one region (the value of REGION).

Troubleshooting​

ECS update container timeout deployment never reaches steady state

START_TIMEOUT is too short for the application to reach HEALTHY. Increase START_TIMEOUT (or temporarily relax the health check command), or pick a longer value that still constrains startup. If the deployment circuit breaker is enabled, the service will roll back automatically; check 'aws ecs describe-services' deployments for rolledback events.

STOP_TIMEOUT change does not appear to affect drain time

The most common causes are: STOP_TIMEOUT is larger than the original stopTimeout (no observable change because the original already gave the app enough time); the load balancer's deregistration delay is the dominant factor (set this to a value compatible with STOP_TIMEOUT); or the application exits faster than STOP_TIMEOUT when it receives SIGTERM. Test with a smaller STOP_TIMEOUT and an application that holds open long-lived connections.

Sidecar containers fail to start within START_TIMEOUT

START_TIMEOUT applies to every container in the task. Sidecars with slow startup (for example Envoy initial xDS sync) may not become HEALTHY within the chaos window even though the main container does. Use a value that respects the slowest sidecar, or remove the sidecar's healthCheck so it is not gated by startTimeout.

Service is left on the chaos task definition after the experiment

If the experiment was killed mid-cleanup, the service may still be on the chaos revision. Run 'aws ecs update-service --cluster <c> --service <s> --task-definition <original-arn>' to roll back, then 'aws ecs deregister-task-definition --task-definition <chaos-arn>' to clean up. The original ARN is recorded in the chaos pod logs.