kubernetes
/
website
mirror of https://github.com/kubernetes/website.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Replace Latinisms with their English equivalents. 

Signed-off-by: Milan Plzik <milan.plzik@grafana.com>
This commit is contained in:
Milan Plzik 2023-10-31 14:17:26 +01:00
parent 820efce9f0
commit 24d2b57f68
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
content/en/blog/_posts/the-case-for-kubernetes-limits/index.md
View File

@ -13,7 +13,7 @@ _Milan Plzik is a Site Reliability Engineer on the [Grafana Labs](http://grafana
_Disclaimer: This article reflects the author's personal opinion; feel free to disagree. Also, there are a lot of subjective ideas that have not made it through any kind of rigorous testing. Use at your own risk._
There’s been quite a lot of posts suggesting that not using Kubernetes resource limits might be a fairly useful thing (e.g., [For the Love of God, Stop Using CPU Limits on Kubernetes](https://home.robusta.dev/blog/stop-using-cpu-limits/) or [Kubernetes: Make your services faster by removing CPU limits](https://erickhun.com/posts/kubernetes-faster-services-no-cpu-limits/) ). The points made there are totally valid – it doesn’t make much sense to pay for compute power that will not be used due to limits, nor to artificially increase latency. This post strives to argue that limits have their legitimate use as well.
There’s been quite a lot of posts suggesting that not using Kubernetes resource limits might be a fairly useful thing (for example, [For the Love of God, Stop Using CPU Limits on Kubernetes](https://home.robusta.dev/blog/stop-using-cpu-limits/) or [Kubernetes: Make your services faster by removing CPU limits](https://erickhun.com/posts/kubernetes-faster-services-no-cpu-limits/) ). The points made there are totally valid – it doesn’t make much sense to pay for compute power that will not be used due to limits, nor to artificially increase latency. This post strives to argue that limits have their legitimate use as well.
Let’s flip the problem upside down. Every pod in a Kubernetes cluster has inherent resource limits – the actual CPU, memory, and other resources of the machine it’s running on. If those physical limits are reached by a pod, it will experience throttling similar to what is caused by reaching Kubernetes limits.
@ -41,13 +41,13 @@ When configuring fixed-fraction headroom limits, a proportional amount of this w
{{<figure alt="Chart displaying various requests/limits configurations" width="40%" src="requests-limits-configurations.svg">}}
For requests = limits (3), this does not hold: Unless we’re able to allocate all node’s resources, there’s going to be some inevitably wasted resources. Without any knobs to turn on the requests/limits side, the only suitable approach here is to ensure efficient bin-packing on the nodes by configuring correct machine profiles. This can be done either manually or by using a variety of cloud service provider tooling – e.g., [Karpenter](https://karpenter.sh/) for EKS or [GKE Node auto provisioning](https://cloud.google.com/kubernetes-engine/docs/how-to/node-auto-provisioning).
For requests = limits (3), this does not hold: Unless we’re able to allocate all node’s resources, there’s going to be some inevitably wasted resources. Without any knobs to turn on the requests/limits side, the only suitable approach here is to ensure efficient bin-packing on the nodes by configuring correct machine profiles. This can be done either manually or by using a variety of cloud service provider tooling – for example [Karpenter](https://karpenter.sh/) for EKS or [GKE Node auto provisioning](https://cloud.google.com/kubernetes-engine/docs/how-to/node-auto-provisioning).
### Optimizing actual resource utilization
Free resources also come in the form of unused resources of other pods (reserved vs. actual CPU utilization, etc.), and their availability can’t be predicted in any reasonable way. Configuring limits makes it next to impossible to utilize these. Looking at this from a different perspective, if a workload wastes a significant amount of resources it has requested, re-visiting its own resource requests might be a fair thing to do. Looking at past data and picking more fitting resource requests might help to make the packing more tight (although at the price of worsening its performance, i.e., increasing long tail latencies).
Free resources also come in the form of unused resources of other pods (reserved vs. actual CPU utilization, etc.), and their availability can’t be predicted in any reasonable way. Configuring limits makes it next to impossible to utilize these. Looking at this from a different perspective, if a workload wastes a significant amount of resources it has requested, re-visiting its own resource requests might be a fair thing to do. Looking at past data and picking more fitting resource requests might help to make the packing more tight (although at the price of worsening its performance – for example increasing long tail latencies).
## Conclusion
Optimizing resource requests and limits is hard. Although it’s much easier to break things when setting limits, those breakages might help prevent a catastrophe later by giving more insights into how the workload behaves in bordering conditions. There are cases where setting limits makes less sense: batch workloads (which are not latency-sensitive, e.g., non-live video encoding), best-effort services (don’t need that level of availability and can be preempted), clusters that have a lot of spare resources by design (various cases of specialty workloads, e.g., services that handle spikes by design).
Optimizing resource requests and limits is hard. Although it’s much easier to break things when setting limits, those breakages might help prevent a catastrophe later by giving more insights into how the workload behaves in bordering conditions. There are cases where setting limits makes less sense: batch workloads (which are not latency-sensitive – for example non-live video encoding), best-effort services (don’t need that level of availability and can be preempted), clusters that have a lot of spare resources by design (various cases of specialty workloads – for example services that handle spikes by design).
On the other hand, setting limits shouldn’t be avoided at all costs – even though figuring out the "right” value for limits is harder and configuring a wrong value yields less forgiving situations. Configuring limits helps you learn about a workload’s behavior in corner cases, and there are simple strategies that can help when reasoning about the right value. It’s a tradeoff between efficient resource usage and performance predictability and should be considered as such.