Or how to stop paying for CPU cores that spend most of their lives doing absolutely nothing.
Cloud infrastructure has a funny habit: making it bigger is easy, making it smaller is scary.
A developer needs a powerful EC2 instance for a temporary workload. You provision 8 vCPUs and 32 GB of RAM. The job finishes, the workload changes, and the instance stays.
Three months later, it's still running, consuming your cloud budget while wondering why nobody gives it anything interesting to do.
I've seen this more times than I'd like to admit.
That's why cloud rightsizing matters. It's not about buying the cheapest possible infrastructure. It's about giving workloads the resources they actually need without turning production into an experiment.
CPU Is Not the Whole Story
Let's say you have an instance with:
- 8 vCPUs
- 32 GB RAM
- Average CPU utilization: 12%
- Average memory utilization: 38%
The obvious reaction: "Let's move to two vCPUs!"
Not so fast.
Looking at average CPU utilization and immediately resizing an instance is like deciding someone doesn't need a car because you saw them walking yesterday.
Before making a decision, I'd check five things:
- CPU: Average usage, peaks, and sustained load. A server averaging 12% might hit 85% during batch processing.
- Memory: Low CPU doesn't mean low RAM requirements. Your application might be keeping a large cache in memory.
- Network: Some workloads are CPU-light but network-heavy.
- Historical behavior: A quiet Tuesday isn't a capacity planning strategy.
- Expected growth: Today's oversized instance might be tomorrow's perfectly reasonable configuration.
One attractive percentage in a dashboard is not an infrastructure strategy.
Why 90 Days of Data Matter
When investigating an oversized resource, I want historical context.
In our Rightsizing implementation, we use 90-day historical usage data in recommendation reporting, including CPU and memory utilization charts. We also consider network activity and associated costs.
Why 90 days?
Because infrastructure workloads have habits. Scheduled jobs, weekly reports, monthly processing, traffic spikes — all those things that nobody remembers until something breaks.
Of course, 90 days isn't a magic number. Seasonal workloads or recently changed applications may require a different observation period.
The goal is simple: make decisions based on actual behavior, not guesses.
A Practical Rightsizing Workflow
Here's how I'd approach an optimization candidate.
1. Find the right resource
Start with a specific cloud account and workload. Collect usage and cost data over a meaningful period.
In Binadox Rightsizing, you can configure the analysis period and apply filters such as cloud region and operating system.
Don't try to resize half your infrastructure on Monday morning. That's a great way to make your incident management team very popular.
2. Understand future requirements
Historical data tells you what happened, not necessarily what happens next.
Define whether the workload is expected to remain stable, grow, or decrease.
An instance that looks oversized today might be perfectly reasonable if traffic is expected to double next month.
3. Compare configurations
Don't choose an instance based on price alone.
Compare:
- vCPU count and memory.
- Instance family and capabilities.
- Historical utilization.
- Expected monthly cost.
- Workload requirements.
The cheapest instance isn't necessarily the best one.
Reducing RAM from 32 GB to 8 GB might look great on a spreadsheet. Until your application needs 10 GB and starts swapping.
Congratulations, you've optimized your cloud bill and introduced a memory problem.
4. Validate before changing production
Recommendations should be treated as hypotheses, not commandments.
Check for resource spikes, scheduled jobs, application dependencies, and upcoming traffic changes.
Test the proposed configuration outside production where possible. Then monitor latency, error rates, CPU, memory, and network activity.
I'd much rather discover a problem in staging than in an incident report.
From Recommendations to Actual Savings
There's one problem with cloud cost optimization that often gets overlooked.
Finding an oversized instance doesn't mean anyone will resize it.
Someone still needs to review the recommendation, compare configurations, estimate savings, create a ticket, get approval, apply the change, and verify the result.
Meanwhile, engineers have 47 other things to finish this week.
Manually rightsizing cloud infrastructure is tedious. You collect metrics, compare instance types, calculate costs, check historical usage, prepare tickets, and repeat everything for the next resource.
So We Automated the Boring Part
At some point, manually checking every instance starts feeling like a second job.
Open CloudWatch, check CPU, investigate memory, compare instance types, calculate potential savings, repeat for the next resource. And then do it all again next month.
We decided this part could be automated.
That's how we built Rightsizing in Binadox. It analyzes CPU, memory, network activity, and costs, then suggests alternative instance configurations based on historical usage.
You can see what you're running now, what could replace it, how the estimated monthly cost would change, and how the resource behaved over time. Recommendations can also be turned into tickets, with Slack and email notifications to keep everyone in the loop.
The important thing is that it doesn't blindly resize production. Engineers review the suggestions and decide what to apply.
I wouldn't claim that automation eliminates the need for engineering analysis. It doesn't. But it does eliminate a lot of repetitive data collection and manual comparison.
And honestly, I'd rather have our engineers spend that time debugging something interesting than comparing EC2 instance prices in spreadsheets.
That's a much better use of their skills.


