Kubernetes Rightsizing

Kubernetes workloads are often given more CPU and memory than they actually need.

Rightsizing helps reduce that waste by adjusting resource requests and limits based on observed workload behavior.

Consize helps identify these opportunities and turn them into safer, actionable changes.

What Consize looks at

Consize analyzes workload resource usage, including:

CPU usage
Memory usage
CPU requests
Memory requests
CPU limits
Memory limits
Workload configuration
Historical usage patterns

The goal is to compare what a workload has been allocated with what it actually uses.

Why percentiles, not averages

Consize sizes requests from the p95 of usage and limits from the p99, both computed over a 14-day window of 15-minute buckets, with explicit headroom multipliers on top. It never uses a plain average.

Averages hide bursts, a workload that spikes to 2Gi once a day but averages 500Mi would get under-provisioned and start OOM-killing under an average-based policy. Maxes overcorrect the other way and recommend sizes so generous there's no real savings. Percentiles sit in between: robust to spikes and weekly patterns, without chasing a single outlier.

This means Consize's recommendations are sometimes slightly larger than a naive average-based tool would suggest. That's intentional, the goal is safe savings, not maximal savings.

A simple example

Imagine a deployment has:

resources:
  requests:
    cpu: "1000m"
    memory: "2Gi"

But the workload consistently uses much less than that.

Consize can identify the difference as a potential optimization opportunity.

Instead of immediately changing the workload, the recommendation can go through the configured review and safety process.

The rightsizing workflow

Consize follows the same core loop described on the homepage:

Observe → Analyze → Recommend → Review → Apply → Verify

Recommendations

A recommendation should answer three questions:

What should change?

For example:

CPU request: 1000m → 500m
Memory request: 2Gi → 1Gi

Why should it change?

The recommendation is based on observed resource usage (p95/p99 over a 14-day window, see above) and the configured optimization strategy.

Is the change safe?

Consize evaluates the proposed change against its configured safety boundaries before it can be applied automatically.

Applying the change

Consize supports two ways to apply a recommendation. You don't have to pick one globally, either can be used per namespace, see Environments.

Review before applying

Keep rightsizing inside your existing Infrastructure-as-Code workflow.

A typical workflow is:

    Consize
       ↓
    Recommendation
       ↓
    Infrastructure-as-Code change
       ↓
    Pull request
       ↓
    Review
       ↓
    Merge
       ↓
    Deploy

This is useful for teams that want optimization recommendations without giving an automated system direct write access to production.

Direct runtime changes

Teams that enable direct application can allow Consize to update supported workloads directly.

Write access should be limited to the namespaces and resources where runtime optimization is allowed.

See The Safety Net for more information about permissions and guardrails.

Start with a small scope

You do not need to optimize every workload at once.

Start with:

A small number of namespaces
A few representative workloads
Recommendations before automatic application
A review of the resulting changes
Gradual expansion once the results are understood

This makes it easier to validate the optimization process before increasing automation.

Monitoring after a change

Rightsizing should not end when the resource configuration changes.

After a change, monitor the workload for:

Increased CPU pressure
Increased memory pressure
Restarts
Failed deployments
Increased latency
Other application-specific signals

Consize's verification workflow helps close this loop.

Next steps

Try it in the Interactive Sandbox
The Safety Net
How Consize Works
Production Installation
Configuration
FAQ for questions about how recommendations are computed

Was this page helpful?

Suggest edits Raise issue