Service Quotas and Throttling

Topics Covered

How Service Quotas Work

Quotas are per account and per region

Adjustable and fixed quotas

New accounts start low

Seeing current usage

Requesting increases

Shared quotas inside an account

Every AWS service has limits. Some cap how many resources an account can create, such as VPCs per region or vCPUs of on demand instances; others cap how fast an account can call an API, such as requests per second to KMS or to the EC2 control plane. These limits protect the shared infrastructure and protect customers from runaway automation, but they also cause outages that look mysterious to teams who do not expect them: a deployment that cannot start new tasks, a launch day that throttles at exactly the same request rate every minute, a new account that cannot run more than a handful of Lambda functions at once. Knowing where the limits are, and designing so they are rarely reached, is part of running anything serious on AWS.

Quotas are per account and per region

Almost every quota applies to one account in one region. An account with a limit of 1,000 concurrent Lambda executions in us-east-1 has a separate 1,000 in eu-west-1, and a second account has its own allowance in each. This has two consequences. Spreading workloads across accounts, which most organisations do for security reasons, also spreads them across separate quota pools. And failing over to another region only works if that region's quotas were raised to match, which teams frequently forget until the day they need it, when an increase request takes hours or days to approve.

Some quotas are global rather than regional, which adds a second trap. IAM limits, such as the number of roles or managed policies per account, apply to the whole account, as do Route 53 hosted zone limits and some Organizations limits. Teams that generate roles per customer or per environment through automation can exhaust these global quotas long after regional ones look fine. The Service Quotas console lists each quota with its scope, and reading the scope is as important as reading the value when planning a design that creates resources in proportion to customers.

Adjustable and fixed quotas

Quotas come in two kinds. Adjustable quotas, sometimes called soft limits, can be raised through the Service Quotas console or API, with many small increases approved automatically and larger ones reviewed by AWS. Fixed quotas, or hard limits, cannot be changed: the maximum item size in DynamoDB is 400 KB, a Lambda function can run for at most fifteen minutes, and an SQS message can be at most 256 KB. Fixed limits shape design, because no request will lift them, while adjustable limits shape operations, because they need to be raised ahead of growth.

Scale out flattened at the 4,000 vCPU Fargate quota. Tasks sat pending for forty minutes until a support case raised it to 8,000.

Resource quotas also interact with each other. An instance launch needs vCPU quota, an Elastic IP quota if it uses one, network interface quota in the VPC, and free addresses in the subnet, which is not an AWS quota at all but behaves like one. A container platform scaling out can hit any of these first. When a scale out stalls, checking only the obvious quota and finding it fine is a common way to lose an hour, and a checklist of the related limits for each type of resource shortens the search considerably.

New accounts start low

Newly created accounts often begin with quotas well below the documented defaults. A new account might allow only a small number of concurrent Lambda executions or a few dozen vCPUs of on demand capacity, rising automatically as the account builds a history of usage. Teams that create a fresh account for a new product, run load tests in staging that pass, and then launch in a new production account sometimes discover the gap on launch day. Checking the applied quota values in the new account, rather than assuming the published defaults, is a cheap step that prevents this.

The gap between applied values and defaults also matters when teams copy architecture between accounts. A design proven in a mature account that has accumulated increases over years will run into the defaults again when it is deployed into a fresh account, whether for a new region, a new customer or a disaster recovery copy. Exporting the applied quota values from the mature account and turning them into a request template is a simple way to carry the operational history along with the infrastructure code.

Seeing current usage

Service Quotas shows the applied value for each quota, and for many of them CloudWatch publishes usage metrics in the AWS/Usage namespace. Metric math includes a SERVICE_QUOTA function that returns the current quota for a usage metric, so an alarm can fire when usage passes, for example, 80 percent of the quota, adjusting automatically when the quota is raised. Trusted Advisor also checks a set of common limits. Alarming on usage as a fraction of quota turns a surprise outage into a ticket raised days in advance.

Usage alarms work best when they go to the team that owns the workload rather than to a central platform inbox. The owning team knows whether growth is expected, whether a sudden jump comes from a bug, and whether the right response is an increase or a fix. A central team can still review all quota alarms weekly to spot patterns across accounts.

Requesting increases

Increases are requested per quota, per account, per region. Small increases on common quotas are often approved within minutes; large ones, or increases on quotas for scarce resources such as GPU instances, can take days and may require a justification. Organizations supports quota request templates that apply a set of increases automatically to every new account, which removes the new account trap for standard workloads. For predictable events such as a product launch or a seasonal peak, raising quotas weeks ahead, in every region the system might fail over to, is part of the launch checklist.

Usage grew steadily toward the 1,000 execution quota. The 80 percent alarm fired nine days ahead, and the quota was raised to 3,000 before anyone was throttled.

Increases do not always take effect instantly or completely. Some quotas are raised in stages, some depend on capacity in a specific availability zone, and for scarce instance types a higher quota does not guarantee that capacity will be available when a launch is attempted. On demand capacity reservations solve the second problem for workloads that must be able to launch at a specific moment, such as a failover, by holding capacity in a chosen zone at the normal on demand price, whether or not instances are running in it.

Shared quotas inside an account

Many quotas are shared by every workload in an account and region. Lambda's concurrency limit is shared by all functions, so a batch job that fans out to thousands of invocations can throttle the customer facing API in the same account. KMS request quotas are shared by every service that encrypts or decrypts with keys in that account, including S3, DynamoDB and Secrets Manager calls made on the team's behalf. Understanding which quotas are shared is often the key to explaining an incident in which one service failed because of traffic in another, and it is a strong argument for separating unrelated workloads into separate accounts.

Shared quotas also explain incidents that cross team boundaries. When two teams share an account, a spike in one team's work can throttle the other, and neither may see the cause in their own dashboards. Account level usage metrics, broken down by function or by caller where possible, make the cause visible. Where separation by account is not practical, reserved concurrency and agreed per team budgets for shared quotas provide some of the same protection, as long as someone owns the account level view and reviews it regularly.

Quotas are per account, per region

Check applied values in every new account, alarm on usage as a share of quota with SERVICE_QUOTA metric math, raise quotas ahead of launches in every failover region, use quota request templates for new accounts, and remember that concurrency and KMS quotas are shared across workloads in an account.