# Cost Efficiency

# Choosing between Goldilocks or Prometheus

Fairwinds Insights provides workload-level resource request and limit recommendations. This is sometimes referred to as "app right-sizing".

You can choose one of two reports to generate resource request and limit recommendations:

Prometheus (Recommended): Uses historical workload usage metrics to generate finer-grained recommendations. This report unlocks additional features in the Insights platform, such as usage graphs.
Fairwinds Goldilocks: Uses Vertical Pod Autoscaler (VPA) to generate resources recommendations, but lacks a historical timeseries of workload usage metrics

Check out a summary of the differences between each tool below:

Feature	Resource Type	Prometheus Installed	Goldilocks Installed	None Installed
Recommendation Engine		Best - Uses Prometheus to generate and store a historical timeseries of workload usage metrics, driving finer-grained resource recommendations	Good - Uses Vertical Pod Autoscaler (VPA) to generate resources recommendations. Lacks a historical timeseries of workload usage metrics
Generates Action Items		X	X
Average Total Cost	Workloads	Best - Cost is calculated using the max(requests or usage)	Better - Cost is calculated using the max(requests or VPA recommendation)	Good - Estimates are based on the average of requests and limits
Total Cost with Recommendations	Workloads	X	X
Cost Difference with Recommendations	Workloads	X	X
Request/Limit Recommendations	Workloads	X	X
Quality of Service (QoS) Recommendations	Workloads	X
Visualize Historical Usage	Workloads	X
Rolling 30 days of Cluster Usage	Cluster	X
Historical Cluster Utilization	Cluster	X
Works with AWS Billing Integration	Cluster	X

# Configuring Cost Settings

Cost attribution and resource tuning can be one of the most difficult and tedious parts of running Kubernetes at scale. Fairwinds Insights provides some features to help make this process simpler and more automated.

The first step is to start mapping CPU and memory usage back to dollars. This is a very difficult problem (opens new window) and inevitably somewhat subjective: how should we rank a CPU intensive application against a memory intensive application? In order to accurately attribute cost, we have to find ways of comparing apples to oranges.

To help us best estimate workload costs in your cluster, we ask for a few pieces of information the first time you visit the Efficiency > Workloads page:

The average node size
The average node cost
The strategy you'd like to use for workload estimation (more on that below)

We've pre-populated a list of instance types from AWS and GCP and can sync data from your bill when AWS Costs is enabled. You can also set custom numbers if you're running on a different cloud provider or if you're using spot instances. If you have multiple node types in your cluster, use the most representative type.

These numbers don't have to be perfectly accurate. They simply give us a baseline for converting memory and CPU to dollars. We will take the Cost per Node, attribute half that cost to memory and half to CPU. By dividing those numbers by the amount of memory and CPU in a single node, we can come to per-CPU and per-GB-memory costs.

# Viewing Workload Costs

On the Efficiency > Workloads page, you can see a list of all the workloads in your cluster. By default, they'll be sorted by their Average Total Cost. Costs are computed using actual workload usage or configured memory and CPU settings.

If you have the prometheus-metrics or goldilocks report installed, Fairwinds Insights will use the maximum of requests and actual usage (max(requests, usage)) in order to compute the Average Total Cost of this workload
If neither report is installed, Fairwinds Insights will use the average of requests and limits. (This can be less precise since some workloads may not have requests and limits configured, or the spread between requests and limits can be large.)

In the next column you'll see Total Costs with Recommendations, followed by Cost Difference with Recommendations. If you're not seeing values in these fields, make sure the prometheus-metrics or goldilocks report is installed and operating properly.

When goldilocks or prometheus-metrics is installed, Insights will analyze actual resource usage for your workloads and make recommendations for how much memory and CPU you should be setting for your requests and limits. While it may recommend moving resources up or down, we typically find that teams have set resources too high since workloads with resources that are too low will experience noticeable performance issues.

If you notice a workload with substantial savings available, click into it to see what Insights recommends setting your resource requests and limits to:

Here, Insights has recommended that we change our memory requests and limits from 1Gi to 263M (a savings of around 75%) and our CPU requests and limits from 500m to 25m, for a savings of 95%.

Note that these recommendations should be sanity checked by the user. If your application experiences periodic bursts in traffic, you may want to keep your limits relatively high. For mission critical applications, it's wise to make any reduction in resources gradually, monitoring your application for any degradation in performance along the way.

It's also good to let Insights gather usage data from either tool for 1-7 days before taking its recommendations. Without a good, representative baseline for actual resource usage, Insights won't be able to make confident recommendations.

# Quality of Service (QoS) Recommendations

When you install the prometheus-metrics report, Fairwinds Insights allows you to generate different resource request and limit recommendations based on your workload's behavior.

The resource recommendation calculations are different depending on your QoS target. See below for additional detail.

QoS	Description	Requests recommendation	Limits recommendation
Critical	Used for mission-critical workloads that should be over-provisioned for reliability	max usage observed over last 2 weeks	max usage observed over last 2 weeks
Guaranteed (default)	Production workloads that can withstand some variability	p95 usage observed over last 2 weeks	p95 usage observed over last 2 weeks
Burstable	Workloads that should prioritize cost efficiency over maximum reliability	mean usage observed over last 2 weeks	p95 usage observed over last 2 weeks
Limited	Workloads that should be given as little resources as needed to operate	mean usage observed over last 2 weeks	mean usage observed over last 2 weeks

← Container Security Policy Enforcement →