Placement, resources and the pod lifecycle

you promise the scheduler the workload needs half a core, and that half core is subtracted from the node whether or not you ever use it.

Resource request

the container went one byte over its memory ceiling and the kernel killed it outright, mid-request.

Resource limit

the sum of every limit on the node is three times the node, and it works fine until the day everything is busy at once.

Overcommit

average CPU use is twelve percent and the p99 latency is terrible, because the container is being stopped for milliseconds at a time.

CPU throttling

the container has no stack trace and no error log, just a restart count going up and exit code 137.

OOM kill

the workload with no requests set was the first one killed when the node ran out of memory, and nobody had chosen that.

Quality of service class

a healthy workload was terminated because the node it happened to be on ran out of disk, and it had done nothing wrong.

Eviction

the cluster is at ninety percent booked and forty percent used, and adding one more workload still needs a whole new node.

Bin packing

the deploy has been sitting at pending for twenty minutes with no error, because no node satisfies every constraint you set.

Unschedulable workload

the expensive GPU nodes sat empty because they push away every workload that has not explicitly said it can live there.

Taint

the workload carries a note saying it can live with the thing that node is marked with, so the scheduler stops excluding it.

Toleration

you say the workload must land on nodes labelled with the fast disk, and the scheduler filters out everything else.

Node affinity

all three replicas were on one node and one reboot took the service down, which is exactly what this rule prevents.

Pod anti-affinity

instead of forbidding two replicas per zone you say the counts must stay within one of each other, and the scheduler balances them.

Topology spread

a node upgrade rolled through the cluster and stopped, because taking one more replica down would have breached what you declared.

Pod disruption budget

you mark the machine unschedulable and evict everything on it politely, so the upgrade takes an hour instead of a page.

Node drain

traffic doubled and the replica count went from four to nine on its own, a minute or two after it needed to.

Horizontal autoscaling

something watched the workload for a week and rewrote its requests, and traditionally had to restart it to apply them.

Vertical autoscaling

a workload could not be placed, so a new machine was bought, booted and joined the cluster — three minutes after you needed it.

Cluster autoscaling

nothing is running and nothing is billed until a request arrives, and that first request waits for a container to start.

Scale to zero

the application takes ninety seconds to warm up, and this is what stops the health checker killing it at thirty.

Startup probe

you get thirty seconds after the signal to finish, and at second thirty-one you are killed with no negotiation.

Termination grace period