Maaz Zindani

Loading the index…

↑↓ move↵ openesc close4072 pages

Cloud and platforms

Fargate vs Container Apps vs Cloud Run: Idle and Scale to Zero

In short

Cloud Run and Container Apps scale to zero by default, and neither bills compute at zero. Fargate scales to zero only if you set the ECS minimum capacity to 0. Each also meters idle capacity differently and has its own request timeout and concurrency control, so check those before you port a workload.

Which of the three scales to zero without extra setup?

Cloud Run does. By default it scales instances with incoming requests and scales to zero when idle, and a minimum-instances setting is what keeps instances warm. Google says scaling from zero can only be triggered by a request.

Azure Container Apps does too. The scale table gives a default minimum of 0 replicas and a default maximum of 10 per revision. The docs warn about one trap: if ingress is off and you define no scale rule or minimum, the app scales to zero and has no way of starting back up.

Fargate is different. The pricing page says it runs Amazon ECS tasks and EKS pods, and an ECS service keeps its desired task count. The ECS docs say that to scale to zero you set a minimum capacity of 0 in Application Auto Scaling. Zero is something you configure, not the starting state.

Checked against: source 1, source 5, source 9

What wakes a service up, and what do the docs say about cold starts?

None of the three pages I read gives a cold-start time in seconds. What they do give is the wake-up mechanism, and it differs.

On Cloud Run a request wakes the service. A request that cannot be served at once waits up to 3.5 times the predicted cold start time, or 10 seconds, whichever is greater. Idle instances may be kept up to 15 minutes, or 10 minutes for GPUs, to reduce cold starts. If a container never listens on the expected port, see the Cloud Run startup fix.

Container Apps decides from rules. HTTP and TCP rules count the past 15 seconds of traffic. Custom KEDA rules poll every 30 seconds by default, and a 300-second cool-down applies only when scaling from the last replica to zero.

Fargate reacts through CloudWatch. ECS sends metrics to CloudWatch in 1-minute intervals. At zero tasks, target tracking waits for one data point and then scales out by the minimum possible amount.

Checked against: source 1, source 5, source 8, source 9

What is still billed while nothing is running or idle?

The meters are the real difference, so here they are without prices.

Fargate bills five dimensions: vCPU, memory, operating system, CPU architecture and storage above the default 20 GB. Billing is per second with a 1-minute minimum for Linux, and runs from the start of the image pull until the task terminates. It is based on requested resources, not measured use. The pricing page I read names no idle rate, so a running idle task is billed like any other until it terminates.

Container Apps (Consumption plan) bills vCPU-seconds, GiB-seconds and HTTP requests, plus GPU-seconds. At zero replicas there are no resource charges. A replica above zero can be billed at a reduced idle rate, but only when the revision sits at its minimum replica count and the replica has all containers running, handles no requests, uses under 0.01 vCPU and receives under 1,000 bytes per second.

Cloud Run depends on the billing setting. Under request-based billing, idle minimum instances are billed at a lower rate, and with a minimum of 0 you are not billed for idle. Under instance-based billing, the full instance lifecycle is billed at the default rate, even with a minimum of 0.

Checked against: source 3, source 6, source 8, source 10

How do request timeouts differ?

Cloud Run has a request timeout setting. The default is 300 seconds and the maximum is 3600. At the limit it closes the connection and returns an HTTP 504, but the instance is not terminated and may keep working on the aborted request.

Container Apps states a request timeout of 240 seconds on HTTP ingress. The ingress page I read shows no setting to change it.

Fargate has no request timeout of its own in the pages I read. If you front it with an Application Load Balancer, the connection idle timeout applies: 60 seconds by default, with a valid range of 1 to 4000 seconds. That timer measures inactivity, not total request time, so a slow response that keeps sending bytes is treated differently from a silent one.

Checked against: source 4, source 7, source 11

What does concurrency control in each one?

Cloud Run concurrency is a per-instance cap on simultaneous requests. The console default is 80, the maximum is 1,000, and 1 is allowed. The concurrency overview says the CLI and Terraform default is 80 times the vCPU count, applied only when a service is created, and that the value is a maximum, not a target.

In Container Apps the number is a scaling threshold, not a cap. The HTTP rule concurrentRequests defaults to 10 with a minimum of 1. When requests exceed it, the app adds a replica, up to the maximum.

Fargate has no request-concurrency setting on the pages I read. You scale tasks on metrics. ECS publishes average CPU and memory usage, and the target tracking page also names ALBRequestCountPerTarget.

Checked against: source 2, source 5, source 12, source 13

Which difference should decide the port?

I would check these in order, because each can break a workload that ran fine on the other platform.

  • Long requests: past 240 seconds, Container Apps ingress is the constraint; past 3600, Cloud Run is.
  • Idle meters: Fargate names no idle rate, Container Apps has a narrow one, Cloud Run has two billing modes.
  • Wake-up path: Cloud Run and Container Apps scale from traffic or rules, while Fargate relies on CloudWatch metrics.
  • Concurrency: a Cloud Run value of 80 and a Container Apps value of 10 are different kinds of number.

Checked against: source 1, source 5, source 8, source 11

What else do people ask?

Does Azure Container Apps charge when scaled to zero?

Its billing page says no resource consumption charges are incurred when a revision is scaled to zero replicas. HTTP requests are a separate meter.

Can I keep a Cloud Run instance warm?

Yes. Set minimum instances above 0. Google calls it a best-effort target, and the instances still bill according to your billing setting.

Sources

How was this page made?

I drafted this page with AI assistance from the official documents listed under Sources. This page compares vendor documents that I fetched on 19 Nov 2026. I did not deploy to Fargate, Container Apps or Cloud Run for it, so it reports documented behaviour, not measured cold starts. It does not cover prices, GPU workloads, jobs or App Runner. Found a mistake? Tell me.

Keep reading

Next step

Have a project in mind?

Tell me about it. A discovery call within 48 hours.