This is outstanding! Well done. I'm glad to see a strong engineer who understands this widely misunderstood mechinism and its consequences, so clearly and thoroughly explaining (and demonstrating!) how it actually works. Keep spreading the gospel, brother, but stay frosty. If any real fraction of the kubernetes using world catches this clue, Amazon stands to lose tons of unearned hosting $$$.
that was interesting, thank you may i ask you which K8s version and underlying OS were used in the test? in my case i managed to achieve much better performance for latency-critical app(single digit ms) with guaranteed QoS. There weren’t other workloads at the node, iow the node wasn’t fully loaded. For the case when pod required more CPU time instantly, the guaranteed QoS was the right choice.
This is outstanding! Well done. I'm glad to see a strong engineer who understands this widely misunderstood mechinism and its consequences, so clearly and thoroughly explaining (and demonstrating!) how it actually works. Keep spreading the gospel, brother, but stay frosty. If any real fraction of the kubernetes using world catches this clue, Amazon stands to lose tons of unearned hosting $$$.
Has me re-thinking requests and limits, thank you for this!
Let us know how it goes!
So in a way using limits is like driving with the breaks on.
You have enough engine power, but because you want to be safe you won't go fast.
It was interesting to know how k8s calculates CPU usage more often than 1s which metrics show
that was interesting, thank you
may i ask you which K8s version and underlying OS were used in the test?
in my case i managed to achieve much better performance for latency-critical app(single digit ms) with guaranteed QoS. There weren’t other workloads at the node, iow the node wasn’t fully loaded. For the case when pod required more CPU time instantly, the guaranteed QoS was the right choice.