A Practical Guide to Setting Kubernetes Requests and Limits

7 minute read

Setting Kubernetes requests and limits effectively has a major impact on application performance, stability, and cost. And yet working with many teams over the past year has shown us that determining the right values for these parameters is hard. For this reason, we have created this short guide and are launching a new product to help teams more accurately set Kubernetes requests and limits for their applications.

The Basics

Resource requests and limits are optional parameters specified at the container level. Kubernetes computes a Pod’s request and limit as the sum of requests and limits across all of its containers. Kubernetes then uses these parameters for scheduling and resource allocation decisions.

resource recs

Requests

Pods will get the amount of memory they request. If they exceed their memory request, they could be killed if another pod happens to need this memory. Pods are only ever killed when using less memory than requested if critical system or high priority workloads need the memory.

Similarly, each container in a Pod is allocated the amount of CPU it requests, if available. It may be allocated additional CPU cycles if available resources are not needed by other running Pods/Jobs.

Note: if a Pod’s total requests are not available on a single node, then the Pod will remain in a Pending state (i.e. not running) until these resources become available.

Limits

Resource limits help the Kubernetes scheduler better handle resource contention. When a Pod uses more memory than its limit, its processes will be killed by the kernel to protect other applications in the cluster. Pods will be CPU throttled when they exceed their CPU limit.

If no limit is set, then the pods can use excess memory and CPU when available.

Here’s a quick example of how you set requests and limits on a container spec:

apiVersion: v1
kind: Pod
metadata:
  name: hello-app
spec:
  containers:
  - name: wp
    image: wordpress
    resources:
      requests:
        memory: "64Mi"
        cpu: "250m"
      limits:
        memory: "128Mi"
        cpu: "500m"

CPU requests are set in cpu units where 1000 millicpu (“m”) equals 1 vCPU or 1 Core. So 250m CPU equals ¼ of a CPU. Memory can be set with Ti, Gi, Mi, or Ki units. For more advanced technical info on the mechanics of these parameters, we suggest these articles:

The Tradeoffs

Determining the right level for requests and limits is about managing trade-offs, as shown in the following tables. When setting requests, there is inherently a tradeoff between the cost of running an application and the performance/outage risk for this application. Balancing these risks depends on the relative cost of extra CPU/RAM compared to the expected cost of an application throttle or outage event. For example, if allocating another 1 Gb of RAM ($5 cost) reduces the risk of an application outage event ($10,000 cost) by 1% then it would be worth the additional cost of these compute resources.

Request Too low Too high
CPU Starvation – may not get CPU cycles needed Inefficiency – requires extra CPUs to schedule other Pods
Memory Kill risk – may be terminated if other pods need memory Inefficiency – requires extra RAM to schedule other Pods

When setting limits, the tradeoffs are similar but not quite the same. The tradeoff here is the relative performance of individual applications on your shared infrastructure vs the total cost of running these applications. For example, setting the aggregated amount of CPU limits higher than the allocated number of CPUs exposes applications to potential throttling risk. Provisioning additional CPUs (i.e. increase spend) is one potential answer while reducing CPU limits for certain applications (i.e. increase throttling risk) is another.

Limit Too low Too high
CPU CPU throttling Starve other Pods or resource inefficiency
Memory Killed by kernel Starve other Pods or resource inefficiency

In the following section, we present a framework for managing these tradeoffs more effectively.

Determining the right values

When setting requests, start by determining the acceptable probability of a container’s usage exceeding its request in a specific time window, e.g. 24 hours. To predict this in future periods, we can analyze historical resource usage. As an example, allowing usage to be above a request threshold with 0.01 probability (i.e. three nines) means that it will, on average, face increased risk of throttling or being killed 1.44 minutes a day.

You can classify applications into different availability tiers and apply these rules of thumb for targeting the appropriate level of availability:

Tier Request Limit
Critical / Highly Available 99.99th percentile + 100% headroom 2x request or as higher if resources available
Production / Non-critical 99th + 50% headroom 2x request
Dev / Experimental 95th or consider namespace quotas* 1.5x request or consider namespace quotas*

This approach of analyzing historical usage patterns typically provides both a good representation of the future and is easy to understand/introspect. Applying extra headroom allows for fluctuations that may have been missed by your historical sampling. We recommend measuring usage over 1 week at a minimum and setting thresholds based on the specific availability requirements of your pod.

*Developers sharing experimental clusters may rely on broader protection from resource quotas. Quotas set aggregate caps at the namespace level and can help protect tasks like long-running batch or ML jobs from getting killed because someone improperly specified resources in another namespace.

Our Solution

Seeing the difficulty of setting these parameters correctly and managing them over time motivated us to create a solution in the Kubecost product to directly generate recommendations for your applications. Our recommendations are based on a configurable Availability Tiers (e.g. Production or Dev), which is easily tracked by namespace or other concepts directly in the Kubecost product.

resource recs

In addition to providing request recommendations, this solution also proactively detects out of memory and CPU throttle risks. The full Kubecost product is available via a single Helm command (install options) and these recommendations can easily be viewed for each container in the Namespace view. Our commercial product is free for small clusters, comes with a free trial for larger clusters, and is based on the Kubecost open source project.

Also, here are sample Prometheus queries if you want to calculate these metrics yourself!

Memory Request (Production Tier)

We recommend container_memory_working_set_bytes because this metric excludes cached data and is what Kubernetes uses for OOM/scheduling decisions. More info in this article.

1.5 * avg(quantile_over_time(.99,container_memory_working_set_bytes{container_name!="POD",container_name!=""}[7d])) by (container_name,pod_name,namespace)

CPU Request (Production Tier)

First, create a recording rule with this expression. Note we recommend using irate to capture short-term spike is resource needs.

avg( irate(container_cpu_usage_seconds_total{container_name!="POD", container_name!=""}[5m]) ) by (container_name,pod_name,namespace)

Then run this query:

1.5 * quantile_over_time(.99,container_cpu_usage_irate[7d])

Vertical pod autoscaling

The goal of vertical pod autoscaling (VPA) is to remove the need to worry about specifying values for a container’s CPU and memory requests. It can be a great solution in certain situations, often with stateless workloads, but you should note that this tool is still in beta as of September 2019. Here are some of the practical limitations to be aware of:

  • Pods are evicted and need to restart when the VerticalPodAutoscaler needs to change the Pod’s resource requests.
  • VPA can cause performance risk and outages if not configured correctly and adds observability complexity.
  • To appropriately handle scale up events, it’s recommended that Cluster Autoscaler also be enabled to handle the increased resource requirements sizes of your workloads.
  • The VPA requires careful tuning to implement a tier-based solution with different parameters for highly available apps, dev, prod, staging etc.

We advise teams to be cautious when using VPA for critical production workloads. It introduces complexity to your infrastructure and you should adequately ensure that your deployments and VPA itself are configured correctly. Risks aside, it can be a great solution when applied correctly.

More info on VPA is available here:

Conclusion

Setting requests and limits effectively can make or break application performance and reliability in Kubernetes. This set of guidelines and this new Kubecost tool can help you manage the inherent risks and tradeoffs when solution like vertical pod autoscaling are not the right fit. Our recommendations combined cost data and health insights are available in Kubecost today to help you make informed decisions.

About Kubecost

Kubecost provides cost and capacity management tools purpose built for Kubernetes. We help teams do Kubernetes chargeback and more. Reach out via email (team@kubecost), Slack or visit our website if you want to talk shop or learn more!

A big thanks to all that have given us feedback and made contributions!