Skip to content

AWS Limits

Usage & Spend group. AWS Limits lists the AWS account’s applied service quotas for the services this stack depends on, with current usage where AWS publishes a usage metric, and lets an admin request an increase. Managed under /admin/aws-limits, scope admin.aws_limits.

These are limits on the AWS account, not the per-user spending limits on the Quotas page.

Off by default while in development: set CDK_AWS_LIMITS_ENABLED=true (app-api, and the Service Quotas IAM grants) and build the SPA with features.awsLimits: true.

Quotas are per account and per Region. The page reads the Region the stack is deployed in. Limits are grouped by the part of the stack that depends on them:

Group Service Shown when
Model inference Bedrock requests and tokens per minute (on-demand, cross-Region, global, and the bedrock-mantle endpoint) Always
Agent runtime AgentCore Runtime sessions, Memory, Code Interpreter, Browser, Gateway Always
Background jobs Lambda concurrent executions Always
App API Fargate On-Demand vCPU, ECS tasks per service Always
Data DynamoDB tables and per-table throughput Always
Knowledge bases (RAG) A link to the S3 Vectors limits, which are not in Service Quotas Always
Fine-tuning SageMaker training and Batch Transform quotas for the G5, G6 and G6e instance types Fine-tuning is on
Dictation Amazon Transcribe concurrent streams Dictation is on

By default, Bedrock rows are limited to the models in the managed-model catalog (the admin Models page). Show all Bedrock quotas lists every Bedrock quota.

Each row shows the applied value (and the AWS default, when it differs), usage, utilization and a link to the quota in the Service Quotas console. Utilization turns amber at 80% and red at 95%.

Usage is the high point of the last 7 days, so a quiet day still shows where the account normally peaks. Hover a value to see the CloudWatch metric behind it.

  • Rates and tokens (requests per second, tokens per minute) use the busiest minute. A per-second quota shows that minute’s average rate, so a burst inside the minute can run higher.
  • Counts (concurrent sessions, Lambda concurrency, tables) use the highest value recorded.
  • Bedrock model quotas have no usage metric in Service Quotas, so they are measured from Bedrock’s own per-model metrics: EstimatedTPMQuotaUsage for tokens per minute (it counts output tokens at their quota burndown rate, as AWS does) and Invocations for requests per minute. A cross-Region quota counts calls through the geographic profile (us.…), a global quota calls through global.…, and an on-demand quota calls to the model itself. A model nobody called that way this week shows 0%.
  • Other quotas without a Service Quotas metric are measured from the service’s own metrics: Code Interpreter and Browser sessions from AgentCore’s ActiveSessionCount, long-term memory extraction from AgentCore’s input and output token usage (both counted, so it errs high), the gateway tool-call rate from gateway Invocations, and DynamoDB per-table throughput from consumed capacity on the busiest table.
  • ECS tasks per service is the busiest service’s RunningTaskCount, which needs Container Insights. Without it the row says so.
  • Counts of what exists (Memories, Gateways, Targets per gateway) are read from the AgentCore APIs and show the current number, since a count has no history. The column heading says which kind a group shows: “7-day peak”, “Current count”, or “Usage” with each row labelled when a group has both.
  • Rows without a number say why: “Limit is 0. Request an increase to use this.” (many SageMaker instance types start at 0), the bedrock-mantle endpoint (not published to CloudWatch), long-context variants (“1M Context Length”, which Bedrock reports with the base model), or “Not measured by AWS”.

Hide unused hides rows with no use in the last 7 days: a peak or count of 0, or nothing measured. Adjustable only hides quotas Service Quotas cannot raise. Both are on when the page opens; clear them to see every monitored quota.

AWS data is cached for about ten minutes. Refresh reads it again.

The first load and each refresh read about 2,000 quotas from Service Quotas (Bedrock and SageMaker alone are about 26 pages each) and a week of CloudWatch data, so they take about 10 to 30 seconds. That time depends on how many quotas AWS lists, not on traffic, so production takes as long as development. The page says so while it loads. The cache belongs to each App API task, so with more than one task a later load can land on a task that has not read AWS yet.

Request AWS increase appears on adjustable quotas for admins who may submit increases. Enter a value higher than the applied one, review it, and submit. The app asks Service Quotas for the increase and shows the status AWS returns.

Every submission is recorded in the audit log as aws_limit.increase_requested with the admin, the quota, the current and requested values and the AWS request ID. AWS itself records the App API’s role as the requester, so the audit record is the only place that says which admin asked. View all on the Open increase requests card opens every request AWS still lists for the account and Region, with that admin, Outside this app for requests made in the console or CLI, or Automatic Management for requests AWS made itself. Approved requests are hidden by default; clear Hide approved requests to see them.

Small increases are often approved automatically within minutes. Larger ones go to a person at AWS and can take several business days. Model tokens-per-minute increases are reviewed against regional capacity: plan about two weeks for high-demand models, and request ahead of a launch rather than during it.

When an increase needs manual review, Service Quotas opens an AWS Support case for it. The app sends SupportCaseAllowed=true; with false, AWS would mark such a request NOT_APPROVED instead of opening a case. The case ID appears in the history once AWS assigns it. Follow up on the case in the AWS Support console.

The App API role has no support:* permissions. Service Quotas opens the case through its own service-linked role.

Each Region has its own quotas. Cross-Region and global inference profiles have their own quotas, separate from on-demand ones, counted in the Region the request is sent from (the stack’s Region). Raise the quota for the inference type the platform calls: a managed model with a us. or global. prefix uses the cross-Region or global quota, not the on-demand one.

The first increase anyone requests in an account creates the Service Quotas service-linked role, AWSServiceRoleForServiceQuotas. The App API role may create that one role (AwsLimitsServiceLinkedRole). If the request still fails with DependencyAccessDeniedException, the page says so; create the role once by requesting any increase from the Service Quotas console, then retry.

Some quotas cannot be raised through Service Quotas, and their rows have no request button. To ask for an exception, open a case in the AWS Support Center (Service limit increase), naming the quota and the Region, or use the console link on the row to check whether the quota has become adjustable. A non-adjustable limit is usually a design constraint: the AWS documentation for the service describes how to work within it.

The banner shows whether Service Quotas Automatic Management is on for the account. When on, AWS watches utilization and notifies you before a quota runs out. Turn it on in the Service Quotas console; the app does not configure it.