# Capacity Planning

Capacity planning is the process of estimating workload demand and arranging sufficient resources to meet performance, growth, and resilience requirements within cost constraints.

Cloud concept. Updated 2026-10-29. Page: https://blowstack.com/tools/cloud-concepts-hub/capacity-planning

> Capacity planning translates expected traffic, data growth, and processing requirements into measurable resource needs. In the cloud, it combines forecasting, benchmarking, scaling policies, service quota management, and failure testing because elastic infrastructure is neither unlimited nor instantaneous.

Practice exams covering Capacity Planning: [ANS-C01 AWS Certified Advanced Networking - Specialty](https://blowstack.com/practice-exams/aws-certified-advanced-networking-specialty), [DVA-C02 AWS Certified Developer - Associate](https://blowstack.com/practice-exams/aws-certified-developer-associate), [AIP-C01 AWS Certified Generative AI Developer - Professional](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional), [MLA-C02 (beta) AWS Certified Machine Learning Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate), [SAA-C03 AWS Certified Solutions Architect - Associate](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate), [SAP-C03 AWS Certified Solutions Architect - Professional](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional).

## 6 key aspects of Capacity Planning

Capacity planning combines demand forecasts, workload-specific measurements, cost-aware headroom, scaling controls, and tested recovery capacity.

1. **Demand forecasting and representative benchmarks**: Forecast baseline demand, peaks, growth, and workload variability, then benchmark realistic operations against latency and throughput objectives. Use measured sustainable capacity rather than theoretical maximum specifications.
2. **Token-aware model sizing**: Estimate input tokens, output tokens, concurrency, and context-length distributions separately for foundation model workloads. Validate processing rates and memory requirements using the actual model and serving configuration.
3. **Utilization and scaling feedback**: Monitor saturation, queue depth, tail latency, and resource utilization to adjust allocation over time. Set scaling limits and signals that reflect workload pressure, while accounting for startup delays and predictable traffic peaks.
4. **Throughput and cost trade-offs**: Balance idle headroom against the operational cost of throttling, slow responses, and failed requests. Tune batching and provisioned throughput against measured demand rather than assuming higher utilization always produces a better result.
5. **End-to-end resource constraints**: Plan network addresses, connection capacity, storage space, IOPS, throughput, and compute together. Increasing one resource does not improve performance when another component remains the bottleneck.
6. **Failure capacity and quota readiness**: Size surviving infrastructure and standby environments for the load they must handle during recovery. Verify relevant quotas in each target Region and distinguish quota approval from actual resource availability.

## Cloud services for Capacity Planning

CloudWatch measurements provide evidence for sizing compute, model inference, storage, and network resources. Bedrock capacity choices or EC2 Auto Scaling policies then address processing demand, while VPC address planning and EBS performance settings prevent infrastructure bottlenecks from limiting that capacity. Service Quotas helps identify administrative ceilings that could block growth or recovery. Together, these services support a feedback loop: forecast demand, benchmark realistic workloads, configure capacity and headroom, monitor results, and revise the plan. Validate the same dependencies in standby Regions because scaling policies and sufficient quotas alone do not guarantee a successful failover.

- **[Amazon Bedrock](https://blowstack.com/tools/cloud-services-hub/amazon-bedrock)**: Amazon Bedrock requires planning around model-specific request and token quotas, supported inference options, and application latency requirements. Where Provisioned Throughput is supported, benchmark the selected model and review utilization before choosing capacity and commitment terms; do not assume application autoscaling changes provisioned model throughput.
- **[Amazon CloudWatch](https://blowstack.com/tools/cloud-services-hub/amazon-cloudwatch)**: Amazon CloudWatch provides metrics, dashboards, and alarms for detecting saturation and validating capacity assumptions. Combine service metrics with application metrics such as token volumes, queue depth, and latency distributions to identify scaling signals and bottlenecks.
- **[Service Quotas](https://blowstack.com/tools/cloud-services-hub/service-quotas)**: AWS Service Quotas helps engineers inspect supported service quotas and request increases for adjustable quotas. Check relevant quotas in both active and standby Regions before growth or failover, remembering that quota increases do not reserve physical resources.
- **[Amazon EC2 Auto Scaling](https://blowstack.com/tools/cloud-services-hub/amazon-ec2-auto-scaling)**: Amazon EC2 Auto Scaling adjusts instance counts within configured minimum and maximum bounds using supported scaling policies. For self-hosted model inference, account for model loading and accelerator initialization, and use suitable custom workload metrics when CPU utilization does not represent demand.
- **[Amazon VPC](https://blowstack.com/tools/cloud-services-hub/amazon-vpc)**: Amazon VPC makes address allocation and subnet design part of capacity planning. Reserve enough address space for application growth, network interfaces, deployment overlap, and failover, and evaluate quotas and performance limits for the network components on the traffic path.

## Capacity Planning exam questions

Real practice exam questions on Capacity Planning, with the correct answer and its explanation.

### Question 1 (MLA-C02 (beta))

**Question:**

A company ingests training observations into an Amazon DynamoDB table in provisioned capacity mode. Each observation creates one new item of 2.4 KiB by using a nontransactional write. The table has no secondary indexes, and writes are evenly distributed across partition keys.

An upcoming scheduled data collection event will increase ingestion from 200 to 2,000 observations per second almost immediately and sustain that rate for 30 minutes. The table currently has 1,000 provisioned write capacity units (WCUs). The ML engineer must prepare capacity before the event so that peak ingestion consumes no more than 75% of provisioned write capacity, without relying on burst capacity or reactive scaling. Applicable service quotas allow the required increase.

Which capacity plan meets these requirements with the least provisioned write capacity during the event?

**Correct answer:** Increase the table to 8,000 WCUs before the event and return it to its normal capacity afterward.

**Explanation:** DynamoDB rounds each 2.4-KiB item up to 3 KiB for write capacity calculations, so peak ingestion requires 2,000 × 3 = 6,000 WCUs. To keep utilization at or below 75%, the table needs at least 6,000 / 0.75 = 8,000 provisioned WCUs. Increasing capacity before the event avoids dependence on reactive scaling.

### Question 2 (SAP-C03)

**Question:**

A company is preparing a stateless application for a scheduled sales event. The application runs on identical 8-vCPU EC2 instances distributed evenly across three Availability Zones in one AWS Region. Representative load tests show that each instance can sustain 600 requests per second while meeting the application's latency objective. Tests confirm that downstream services can support the planned load.

Peak demand is forecast to reach 18,000 requests per second. The company requires ready application capacity sufficient for 120% of forecast peak demand, even immediately after the complete loss of any one Availability Zone. Starting or resuming an instance and initializing the application takes 12 minutes. No additional instances can be counted toward the capacity requirement during that interval.

Other workloads continuously consume 160 vCPUs from the same Regional EC2 On-Demand vCPU quota. The applied quota is currently 576 vCPUs. All application instances count against this quota. Sufficient EC2 infrastructure capacity is available for the planned launches.

Which capacity plan meets all the requirements?

**Correct answer:** Obtain and verify an applied vCPU quota of at least 592. Before the event, launch and initialize 54 application instances, with 18 in each Availability Zone.

**Explanation:** The required surviving capacity is 18,000 × 1.20 = 21,600 requests per second, which requires 36 ready instances. With 18 instances in each of three Availability Zones, losing one zone leaves 36. The full application fleet consumes 54 × 8 = 432 vCPUs. Including the other workloads requires an applied quota of at least 432 + 160 = 592 vCPUs. Initializing all instances before the event avoids dependence on launches during the failure.

### Question 3 (AIP-C01)

**Question:**

A company is adding an incident-investigation agent to an application that uses Amazon Bedrock. Each investigation makes between 2 and 12 model invocations, including parallel calls and additional calls after tool results. Prompt lengths grow during an investigation. Historical traces show that major incidents trigger many investigations simultaneously, and new investigations continue arriving even when earlier investigations are slow.

The agent will share a model's Regional inference quotas with an existing production application. The company must determine before launch whether the forecast workload can meet a p95 investigation-completion objective without degrading the existing application. Historical investigation traces and the existing application's peak traffic profile are available.

Which capacity-validation approach provides the MOST reliable basis for the launch decision?

**Correct answer:** Replay representative investigation traces at forecast event-arrival times while generating the existing application's peak load. Preserve invocation dependencies, parallel calls, and token-length patterns; validate end-to-end latency and model-specific quota consumption, and secure any required quota increases before launch.

**Explanation:** This translates forecast business demand into realistic model demand across the shared capacity boundary. Arrival-driven replay preserves incident bursts without letting slow responses suppress new arrivals. Testing the actual invocation structure and token patterns exposes quota contention and its effect on both applications, allowing capacity requirements to be addressed before launch.

### Question 4 (SAA-C03)

**Question:**

A company is selecting an AWS Direct Connect hosted connection for a new analytics workload. Measurements over a complete business cycle show average traffic of 200 Mbps toward AWS and 150 Mbps toward the data center. Peak traffic reaches 700 Mbps toward AWS and 500 Mbps toward the data center, with both peaks occurring simultaneously.

The company forecasts 40% traffic growth over the next year. Its network policy requires projected peak utilization to remain at or below 75% of the connection's capacity in each direction. All measurements include protocol overhead. The provider confirms that each connection supports its rated bandwidth simultaneously in both directions and offers 500 Mbps, 1 Gbps, 2 Gbps, and 5 Gbps connections. Recurring charges increase with connection speed. Additional connection redundancy is outside the scope of this decision.

Which connection speed meets the capacity requirements at the LOWEST recurring cost?

**Correct answer:** 2 Gbps

**Explanation:** Projected peaks are 980 Mbps toward AWS and 700 Mbps toward the data center. The higher peak requires at least 980 / 0.75, or approximately 1,307 Mbps, of rated capacity. A 2 Gbps connection is the smallest offered speed that satisfies the utilization policy in both directions.

### Question 5 (ANS-C01)

**Question:**

A company forecasts IPv4 address demand for an application deployed across three Availability Zones. Each zone has one application subnet, and application instances must launch in their assigned zone. The subnet in one zone currently has 300 available IPv4 addresses. Forecast growth will consume exactly 20 additional addresses per day in that subnet. The other two application subnets each have more than 1,000 available addresses.

The constrained subnet must retain at least 100 available addresses for an unplanned replacement deployment until additional capacity is ready. The VPC's existing IPv4 CIDR block is fully allocated to subnets. An approved, nonoverlapping secondary IPv4 CIDR block is available. Associating this block, creating and configuring a new subnet in the constrained zone, and updating application placement takes seven days from initiation. The new subnet will provide sufficient capacity for subsequent growth and replacement deployments. Existing resources must remain operational.

Assume the forecast is accurate, service quotas are sufficient, and no addresses can be reclaimed. Which capacity plan meets the requirements while allowing the latest safe initiation of the expansion?

**Correct answer:** Initiate expansion no later than three days from now, when the constrained subnet has 240 available addresses. Associate the secondary CIDR block, create the new subnet in that zone, and update application placement.

**Explanation:** Seven days of growth consumes 140 addresses. The expansion must therefore begin with at least 240 available addresses: 140 for growth plus the required 100-address reserve. Starting after three days leaves 240 addresses, and completing seven days later leaves exactly 100. The new subnet then provides capacity in the required Availability Zone.

## Practice Capacity Planning in a full exam

Questions on Capacity Planning in each practice exam, with every answer option and an explanation for each:

- [SAA-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate): 12 questions
- [ANS-C01](https://blowstack.com/practice-exams/aws-certified-advanced-networking-specialty): 12 questions
- [AIP-C01](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional): 6 questions
- [SAP-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional): 3 questions
- [MLA-C02 (beta)](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate): 1 question
- [DVA-C02](https://blowstack.com/practice-exams/aws-certified-developer-associate): 1 question

## What is Capacity Planning

Capacity planning starts with a workload’s service-level objectives, such as response latency, throughput, availability, and recovery time. Engineers forecast normal demand, seasonal peaks, unexpected bursts, and future growth, then determine the resources needed to meet those objectives. The goal is not maximum utilization at all times: resources running close to saturation can develop long queues and sharply increasing latency. Effective plans provide deliberate headroom while avoiding unnecessary idle capacity.

Translate business demand into technical units before selecting resources. An application serving 1,000 requests per second might be constrained by database connections, storage IOPS, network bandwidth, or CPU rather than the number of application instances. Benchmark representative requests and datasets, measure sustained performance, and identify the bottleneck across the complete request path. Average utilization alone is insufficient; examine peak demand, tail latency, queue depth, and uneven load distribution.

For foundation model workloads, requests per second are not enough to predict capacity. Track input and output token distributions, concurrent requests, context lengths, time to first token, and output generation rate. For example, 10 requests per second averaging 2,000 input tokens and 500 output tokens imply approximately 20,000 input tokens and 5,000 output tokens per second. These are separate demand estimates, not interchangeable units of compute: prompt processing and token generation behave differently, and service quota accounting depends on the provider and model. Long contexts and concurrent sequences also increase memory requirements, including memory used for attention caches in self-hosted inference.

Benchmark the selected model, hardware, and serving configuration under realistic prompt and completion patterns. Dynamic batching can improve accelerator utilization and aggregate throughput, but waiting to assemble batches can increase latency. Large batches can also exhaust memory. Tune batch limits and waiting time against latency objectives, and use asynchronous batch processing when workloads do not require interactive responses. For provisioned model throughput, compare sustained demand with measured usable capacity and contractual costs; consistently underused provisioned capacity can cost more than an on-demand approach.

Autoscaling implements part of a capacity plan rather than replacing it. Configure minimum capacity, maximum capacity, scaling metrics, and cooldown or stabilization behavior around workload characteristics. For model serving, queued work, active sequences, or estimated outstanding tokens may reveal pressure better than CPU utilization alone. Account for instance launch time, model loading, accelerator availability, and metric delays. Scheduled scaling can prepare for predictable peaks, while bounded queues, admission control, and backpressure protect the application during bursts that scaling cannot absorb immediately.

Network planning must consider future address consumption, connection counts, bandwidth, packet rates, and intermediary limits. A subnet that fits the initial deployment may run out of IP addresses during scale-out or rolling replacement. Allocate address space for growth and failover, and verify that load balancers, NAT gateways, private connectivity, and instance network limits can support projected traffic. Adding compute does not resolve an exhausted address pool or a constrained network path.

Storage planning separates capacity in bytes from performance in IOPS, throughput, and latency. Estimate data growth, retention, replication, temporary files, and recovery copies. A volume with enough free space can still throttle because its provisioned performance or the attached instance’s storage bandwidth is insufficient. Managed storage may expand automatically, but access patterns, request limits, partition distribution, and downstream consumers still require planning.

Finally, validate capacity during failures, not just normal operation. If two zones normally share traffic equally, the surviving resources must accommodate the required workload when one zone fails. A standby Region needs appropriate quotas, dependencies, data, and a tested scale-up path before failover. Quota increases do not necessarily propagate across Regions, and an approved quota is permission to consume resources, not a guarantee of available physical capacity. Request increases early, use capacity reservations where supported and necessary, and load-test recovery scenarios. Revisit the plan whenever traffic, models, architecture, or service limits change.

## Learn Capacity Planning in lessons

8 short lessons on Capacity Planning in 4 certification courses.

- MLA-C02 (beta): Capacity Planning: sizing data ingestion pipelines (Premium), https://app.blowstack.com/lessons/capacity-planning-sizing-data-ingestion-pipelines-mla-c02
- SAP-C03: Capacity Planning: sizing baseline and burst capacity (Premium), https://app.blowstack.com/lessons/capacity-planning-sizing-baseline-and-burst-capacity-sap-c03
- AIP-C01: Capacity Planning vs Elasticity (Premium), https://app.blowstack.com/lessons/capacity-planning-vs-elasticity-aip-c01
- SAA-C03: Capacity Planning: storage growth and performance (Premium), https://app.blowstack.com/lessons/capacity-planning-storage-growth-and-performance-saa-c03
- AIP-C01: Capacity Planning: sizing foundation model serving capacity (Premium), https://app.blowstack.com/lessons/capacity-planning-sizing-foundation-model-serving-capacity-aip-c01
- SAA-C03: Capacity Planning: network growth headroom (Premium), https://app.blowstack.com/lessons/capacity-planning-network-growth-headroom-saa-c03
