Telling the scheduler what you need.
WHAT A REQUEST IS
The amount guaranteed to the container.
WHAT A LIMIT IS
The maximum it may use.
WHAT THE SCHEDULER USES
Requests, to decide where a pod fits.
WHAT HAPPENS WITH NO REQUESTS
Scheduling is guesswork, and nodes become overcommitted.
WHAT HAPPENS WHEN MEMORY EXCEEDS ITS LIMIT
The container is terminated.
WHAT HAPPENS WHEN PROCESSOR EXCEEDS ITS LIMIT
It is throttled, not terminated.
WHY THAT DIFFERENCE MATTERS
Memory limits kill; processor limits slow.
WHAT THAT ARGUES FOR
Setting memory limits with headroom, and being cautious with processor limits.
WHY CAUTIOUS WITH PROCESSOR
Aggressive throttling causes latency that is difficult to diagnose.
HOW TO CHOOSE VALUES
Measure actual usage under realistic load.
WHAT NOT TO DO
Copy values from an example Set limits far above requests everywhere
WHY THAT SECOND POINT
It produces nodes that appear to have room and do not.
WHAT QUALITY OF SERVICE MEANS
How the system prioritises pods when a node is under pressure.
WHAT IS EVICTED FIRST
Pods with no requests, then those exceeding them.
WHAT TO SET FOR ANYTHING IMPORTANT
Requests equal to limits.
WHY
It gives the strongest guarantee.
WHAT TO REVIEW
Actual usage against configured values, periodically.
