Day 48 of 100
The Pod Has No Resources
π₯ Problem Statement
During a routine review of production workloads, someone notices a deployment running without any CPU or memory requests or limits configured at all. It's been running fine β no incidents, no visible problems. But "no problems yet" and "no risk" aren't the same claim, and this configuration gap raises a real question before it turns into an actual incident.
ποΈ Environment
- Deployment: no CPU or memory requests/limits configured
- Cluster: shared, multiple workloads across several teams
- QoS class: BestEffort by default, as a result
- Symptom under discussion: hypothetical review before a node-pressure incident occurs
π€ Your Challenge
What operational risks does running a pod with no resource requests or limits actually create, even if nothing has gone wrong yet?
- What Quality of Service class does Kubernetes assign a pod with no requests or limits, and what does that mean under node pressure?
- How does the scheduler make placement decisions for a pod with no resource requests specified?
- What happens to this pod first if a node runs low on memory and something has to be evicted?
- How would you plan cluster capacity for workloads whose resource needs aren't declared anywhere?
Solution Hidden
Think through the problem yourself before looking at the answer.
π‘Solution
Step 1 β Understand the symptoms
Kubernetes assigns pods to Quality of Service classes based on their resource configuration, and a pod with no requests or limits at all lands in the lowest tier: BestEffort. That classification matters most under pressure β when a node runs low on memory, Kubernetes evicts BestEffort pods first, before pods with any guaranteed resources, regardless of how important that workload actually is to the business. "No problems yet" simply means this hasn't been tested under real node pressure β the risk is latent, not absent.
Step 2 β Identify the likely bottleneck
Beyond eviction priority, a pod with no CPU request gives the scheduler no information to make a good placement decision β it can be packed onto a node alongside other workloads with no accounting for how much CPU it might actually need, contributing to exactly the kind of contention and noisy-neighbor risk seen elsewhere. Without any request or limit, capacity planning for the cluster as a whole is also compromised β nobody can reason accurately about whether the cluster has enough capacity for its actual workloads if some of them declare no resource needs at all.
Step 3 β Investigation
- Determine this workload's actual resource usage under normal and peak conditions β CPU and memory, observed over a representative period, to establish a real baseline rather than guessing.
- Assess how critical this workload is β if it were evicted under node memory pressure, what's the actual business impact? This determines how much QoS priority it genuinely deserves.
- Check whether other workloads in the cluster have the same gap β an audit across all deployments for missing resource configuration, since this is rarely an isolated case once found once.
- Check current cluster capacity planning practices β is total capacity being reasoned about based on declared requests, and if so, how much of the cluster's real usage is currently invisible to that planning because of gaps like this one?
Step 4 β Recommended action
Set explicit CPU and memory requests based on observed real usage, with reasonable headroom, and limits appropriate to prevent this workload from consuming unbounded resources (directly connecting to the noisy-neighbor risk). This moves the pod to at least the Burstable QoS class, giving it real eviction priority over genuinely BestEffort workloads, and gives the scheduler accurate information to place it well.
For workloads where resource needs are highly variable or hard to characterize precisely, requests set conservatively with limits set higher (allowing bursting) is often a reasonable middle ground between under-provisioning and over-provisioning.
At the platform level, this is a good justification for enforcing LimitRanges or admission policies that reject deployments with no resource configuration at all, so this gap can't be introduced again for a future workload without at least a deliberate, explicit decision to allow it.
Step 5 β Engineering lesson
Resource requests and limits aren't just a performance-tuning detail β they directly determine a workload's priority when the cluster is under real pressure, and its absence isn't neutral, it's a default to the lowest possible priority tier. A workload running fine today with no resource configuration is running on borrowed reliability: it works exactly until the one day the cluster is under pressure and something has to be evicted first.
π§ Todayβs Takeaway
Resource configuration is part of reliability engineering.
π Donβt Miss Tomorrowβs Challenge
A new real-world engineering challenge is released every day.
100 Days β 100 Challenges β 100 New Things Learned.
Get the useful stuff, not the noise.
Occasional notes on engineering, Platform Engineering, AI, cloud and things Iβm learning along the way.

