# Workload Predictability

Workload predictability is the degree to which an application's future resource demand, usage duration, and operating schedule can be anticipated.

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

> Predictability helps engineers decide how much compute usage they can safely commit to and how much should remain flexible. Matching purchasing options to measured demand reduces costs without confusing predictable usage with interruption tolerance or guaranteed capacity.

Practice exams covering Workload Predictability: [CLF-C02 AWS Certified Cloud Practitioner](https://blowstack.com/practice-exams/aws-certified-cloud-practitioner), [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 Workload Predictability

Effective purchasing decisions combine demand forecasts, conservative commitments, interruption requirements, and capacity needs rather than relying on a single traffic pattern.

1. **Demand patterns and time horizon**: Assess resource demand, operating schedules, and expected workload lifetime together. A repeatable schedule provides useful predictability, but limited daily runtime may not justify a continuous hourly commitment.
2. **Baseline and variable capacity**: Separate sustained usage from bursts and uncertain growth. Apply commitments to a defensible baseline and retain flexible purchasing for the remainder.
3. **Commitment utilization**: Evaluate whether eligible usage will consistently consume a proposed commitment. Rightsize first and account for migrations, shutdown schedules, and architectural changes before committing.
4. **Independent interruption tolerance**: Determine whether work can checkpoint, retry, or wait for capacity independently of how predictable its demand is. Use interruption-tolerant execution to evaluate Spot suitability, not predictability alone.
5. **Price versus capacity assurance**: Distinguish discounted billing from reserved capacity. A purchasing discount does not necessarily ensure that resources can launch in a particular location.
6. **Continuous reassessment**: Compare forecasts with actual utilization and spending over representative business cycles. Adjust future commitments and scaling policies when the workload's behavior changes.

## Cloud services for Workload Predictability

CloudWatch supplies operational evidence about resource demand, while Cost Explorer shows how that demand translates into spending and commitment utilization. Together, they help identify a rightsized, dependable baseline. EC2 provides purchasing options for that baseline, uncertain additional demand, and interruptible work. EC2 Auto Scaling adjusts the running fleet as requirements change, including combinations of On-Demand and Spot capacity. Scaling policies manage resource quantity; commitment discounts manage eligible usage costs. Neither forecasting nor automatic scaling eliminates the need to assess capacity availability, application resilience, and the financial risk of unused commitments.

- **[Amazon EC2](https://blowstack.com/tools/cloud-services-hub/amazon-ec2)**: EC2 offers On-Demand and Spot Instances, Reserved Instance discounts, and eligibility for Savings Plans. Engineers can match these options to baseline demand, uncertainty, and interruption tolerance while evaluating capacity reservations separately.
- **[Amazon EC2 Auto Scaling](https://blowstack.com/tools/cloud-services-hub/amazon-ec2-auto-scaling)**: EC2 Auto Scaling adjusts instance capacity to changing demand and supports mixed instance policies combining On-Demand and Spot capacity. Scheduled scaling can prepare for known demand changes, while dynamic scaling responds to observed conditions.
- **[Amazon CloudWatch](https://blowstack.com/tools/cloud-services-hub/amazon-cloudwatch)**: CloudWatch provides metrics that reveal utilization patterns, recurring peaks, and idle periods. These measurements help validate whether an apparent baseline represents necessary capacity or oversized resources.
- **[AWS Cost Explorer](https://blowstack.com/tools/cloud-services-hub/aws-cost-explorer)**: Cost Explorer helps analyze historical costs and usage and review Reserved Instance and Savings Plans utilization and coverage. Its recommendations inform commitment sizing, but engineers must also account for future workload changes.

## Workload Predictability exam questions

Real practice exam questions on Workload Predictability, with the correct answer and its explanation.

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

**Question:**

An ML team runs occasional Amazon SageMaker AI training jobs with a custom container. Each job needs approximately 3 hours of compute, but the team allows up to 10 hours of elapsed time, including capacity waits and interruptions. Completion within this window is best effort. The team wants to configure a 4-hour training execution limit and a 10-hour overall job limit while reducing compute charges compared with standard On-Demand training.

The training script periodically writes checkpoints to a local directory and can restore its state from that directory when restarted. No other workloads can consume a long-term purchasing commitment.

Which solution best meets these requirements?

**Correct answer:** Use SageMaker Managed Spot Training. Configure MaxRuntimeInSeconds as 14,400 and MaxWaitTimeInSeconds as 36,000. Configure the script's checkpoint directory as the local checkpoint path with an Amazon S3 checkpoint location, and restore checkpoints when the container starts.

**Explanation:** Managed Spot Training reduces training compute charges by using interruptible capacity without a long-term commitment. The 4-hour runtime limit and 10-hour wait limit distinguish training execution from total elapsed time, including capacity waits. SageMaker synchronizes the configured checkpoint directory with S3 so the script can restore progress after an interruption.

### Question 2 (SAP-C03)

**Question:**

A company operates a rightsized fleet of Amazon RDS for PostgreSQL databases. During the next 12 months, teams will migrate some workloads to Aurora Serverless v2 and move others between AWS Regions. The migration schedule and final allocation of workloads are uncertain, but the aggregate database demand is predictable.

A pricing analysis that accounts for all planned migrations establishes the following usage throughout the next year:
• For 18 hours each day, eligible database usage costs $60 per hour at On-Demand rates or $40 per hour at the applicable discounted commitment rates.
• For the remaining 6 hours each day, eligible usage costs $120 per hour at On-Demand rates or $80 per hour at the applicable discounted commitment rates.

All planned database configurations are eligible for the applicable database commitment offering. The company has no existing reservations or spending commitments. Finance requires any new hourly commitment to be fully consumed in every hour, including during migrations. The company must preserve flexibility to change database engines, capacity configurations, and Regions.

Which purchasing strategy provides the greatest commitment-based savings while meeting these requirements?

**Correct answer:** Purchase a one-year Database Savings Plan with a $40-per-hour commitment. Allow eligible usage above the commitment to be billed at On-Demand rates.

**Explanation:** Database Savings Plans provide flexibility across eligible database services, engines, capacity configurations, and Regions. The commitment is consumed at Savings Plans rates, so $40 per hour is the largest commitment that the established baseline will fully consume every hour. Usage beyond the commitment remains On-Demand.

### Question 3 (SAA-C03)

**Question:**

A company uses an Amazon DynamoDB table for warehouse operations. For the past 6 months, the table has required 500 read capacity units (RCUs) and 200 write capacity units (WCUs), except during a daily import from 01:00 to 03:00 UTC that requires 5,000 RCUs and 2,000 WCUs. The import starts immediately at full throughput. This schedule and capacity requirement will remain unchanged for the next year, and requests are evenly distributed across partition keys.

Testing shows that reactive scaling from the lower capacity settings causes throttling when the import begins. The application cannot tolerate this throttling. A pricing analysis confirms that provisioning capacity for the measured low-demand and import periods costs less than using on-demand capacity mode.

Which solution meets the requirements at the LOWEST ongoing throughput cost?

**Correct answer:** Use provisioned capacity with scheduled scaling. Raise capacity to 5,000 RCUs and 2,000 WCUs sufficiently before 01:00 UTC, and return it to 500 RCUs and 200 WCUs after the import completes.

**Explanation:** The repeatable schedule allows capacity increases to complete before demand rises. Scheduled reductions eliminate unnecessary peak capacity during the rest of the day. This meets the no-throttling requirement and uses the lower-cost provisioned approach identified by the pricing analysis.

### Question 4 (AIP-C01)

**Question:**

A media company uses a customized foundation model to generate detailed reports from uploaded transcripts. The model requires a GPU and a custom inference container. Individual inference requests take 9–12 minutes, and loading the model onto a new instance takes 4 minutes. Requests arrive unpredictably, with several hours of no activity most days.

The application stores inputs in Amazon S3 and can return a job identifier so clients can retrieve results later. Representative tests confirm that the selected GPU instances can meet the processing deadline, including startup and queueing time. The company requires automatic processing when requests arrive and no charges for running endpoint instances after the queue has drained and the endpoint has scaled in.

Which deployment configuration best meets these requirements?

**Correct answer:** Deploy a SageMaker AI asynchronous inference endpoint on GPU instances. Configure S3 output storage and a 1,200-second invocation timeout. Set the minimum instance count to zero, use backlog-per-instance target tracking, and add a scale-out policy based on HasBacklogWithoutCapacity to start capacity when requests arrive at zero instances.

**Explanation:** Asynchronous inference supports GPU-backed custom containers, long-running requests, queued inputs, and S3 results. A 1,200-second invocation timeout accommodates the observed inference duration. Scaling to zero removes running endpoint instance charges during idle periods. The HasBacklogWithoutCapacity policy explicitly handles requests arriving when no instances are available, while backlog-per-instance tracking manages capacity as demand changes.

## Practice Workload Predictability in a full exam

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

- [MLA-C02 (beta)](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate): 8 questions
- [SAP-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional): 7 questions
- [AIP-C01](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional): 5 questions
- [SAA-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate): 3 questions
- [CLF-C02](https://blowstack.com/practice-exams/aws-certified-cloud-practitioner): 1 question

## What is Workload Predictability

Workload predictability describes how confidently engineers can estimate future compute requirements. Relevant factors include the minimum capacity needed, how demand changes over time, when resources run, and how long the application is expected to operate. Historical utilization, business schedules, growth forecasts, and planned architecture changes all inform that estimate. Predictability is not simply whether traffic remains constant: a daily processing job can be highly predictable even though it runs for only a few hours.

In cloud purchasing decisions, separate the dependable baseline from variable demand. An application might continuously require enough compute for eight instances but need twenty during business hours. The baseline may justify a commitment-based discount, while additional capacity can use flexible purchasing. Commit conservatively after rightsizing resources; purchasing discounts for oversized infrastructure preserves unnecessary spending.

On AWS, On-Demand Instances are suitable for short-lived workloads, experiments, uncertain demand, or applications whose future requirements are not yet understood. They do not require a long-term usage commitment. Their higher unit price relative to discounted options buys flexibility, but choosing On-Demand does not itself guarantee that capacity will be available when an instance launch is requested.

Reserved Instances and Savings Plans can reduce the cost of sustained, predictable usage in exchange for one-year or three-year commitments. EC2 Reserved Instances provide a billing discount for matching instance usage, with flexibility depending on the reservation type and attributes. Savings Plans commit to a consistent amount of eligible compute usage, measured in dollars per hour; their scope depends on the plan type. Unused commitment still costs money, so neither option should automatically cover forecast peaks or speculative growth. Predictable aggregate usage across multiple applications may support a commitment even when individual applications fluctuate.

Spot Instances address a different question: can the workload tolerate interruption? They use spare EC2 capacity at discounted prices, but capacity can be reclaimed. A predictable nightly batch job may suit Spot if it can checkpoint, retry, and meet its deadline despite interruptions. Conversely, an unpredictable customer-facing application may require On-Demand capacity because it cannot safely depend on Spot availability. Predictability and interruption tolerance must therefore be evaluated separately.

Another distinction is between a billing commitment and a capacity reservation. Savings Plans do not reserve capacity, and regional EC2 Reserved Instances do not provide a capacity reservation. Zonal EC2 Reserved Instances include a capacity reservation in the specified Availability Zone; On-Demand Capacity Reservations can reserve matching capacity without a long-term commitment. Dedicated Hosts primarily address physical-server isolation and certain licensing requirements, rather than unpredictable demand.

For example, an online retailer can use commitment discounts for its established year-round baseline, On-Demand capacity for uncertain promotional traffic, and Spot capacity for restartable image processing. Review utilization, commitment coverage, and business plans regularly. A workload that was predictable before a migration, seasonal change, or product retirement may no longer justify the same purchasing strategy.

## Learn Workload Predictability in lessons

6 short lessons on Workload Predictability in 5 certification courses.

- CLF-C02: Workload Predictability: finding a dependable compute baseline (Premium), https://app.blowstack.com/lessons/workload-predictability-finding-a-dependable-compute-baseline-clf-c02
- MLA-C02 (beta): Workload Predictability: dependable demand and flexible capacity (Premium), https://app.blowstack.com/lessons/workload-predictability-dependable-demand-and-flexible-capacity-mla-c02
- SAP-C03: Workload Predictability: the dependable usage baseline (free), https://app.blowstack.com/lessons/workload-predictability-the-dependable-usage-baseline-sap-c03
- SAA-C03: Workload Predictability: choosing a safe compute commitment (Premium), https://app.blowstack.com/lessons/workload-predictability-choosing-a-safe-compute-commitment-saa-c03
- AIP-C01: Workload Predictability: inference demand baselines (Premium), https://app.blowstack.com/lessons/workload-predictability-inference-demand-baselines-aip-c01
- CLF-C02: Workload Predictability vs. Fault Tolerance (Premium), https://app.blowstack.com/lessons/workload-predictability-vs-fault-tolerance-clf-c02
