Predictable Demands Pattern

Introduction

• Kubernetes containers contain logic written in different languages—each with different resource requirements
• The developer is best placed to know these requirements—from testing

Solution

• Knowing resource requirements up-front allows:
  • Kubernetes to intelligently make decisions on Pod placement—more efficient hardware utilization
  • Capacity planning—more cost-effective environments

Runtime Dependencies

• File storage
  • Container storage is ephemeral
  • Kubernetes offers Pod-level storage which can survive restarts
• Kubernetes Volumes:
  • emptyDir—ephemeral to Pod
  • PersistentVolume—long-lived
  • Type of required volume affects Pod placement—Pod might not be scheduled if no node available
• Configuration e.g. ConfigMaps, Secrets

Resource Profiles

• Compute resource requirements
• Types:
  • Compressible—can be throttled if container goes over limit, e.g. CPU, network bandwidth
  • Incompressible—can’t be throttled (container killed if goes over limit), e.g. memory
• Profiles—soft/hard limits:
  • requests—minimum amount of resource required
  • limits—maximum amount of resource permitted
• Scheduler uses requests for Pod placement on nodes
• Quality of Service (QoS):
  • Best effort—no requests, no limits—killed first when node under incompressible resource pressure
  • Burstable—requests != limits—killed if no best effort Pods available
  • Guranteed—requests == limits—killed last

Pod Priority

• Importance of Pods relative to other Pods
• Specify PriorityClass object
• Influences scheduling order
• If no Pods with enough capacity, scheduler can pre-empt other Pods with lower priority—QoS not considered
• Use with caution

Resource Quotas

• Limit resources in namespaces
  • CPU, mem, storage consumed
  • Number of objects (ConfigMaps, Secrets, Pods, Services)

Limit Ranges

• Min/max permitted amounts for requests/limits
• Also set request to limit ratio—overcommit level
  • Only requests used for Pod placement
  • High overcommit level increases chance for overcommitting of resources on node—degrades performance, Pods evicted

References

• Ibryam, B and Huß, R (2019), Kubernetes Patterns, O’Reilly