Comparison
Resource limitvsResource request
Resource limit
the container went one byte over its memory ceiling and the kernel killed it outright, mid-request.
The hard ceiling a container may consume, enforced by the kernel rather than by the scheduler. The two resources behave completely differently at the ceiling: CPU is throttled, so the workload gets slow, and memory is not compressible, so the process is killed. That asymmetry is why a CPU limit shows up as mysterious latency and a memory limit shows up as a restart loop.
Full entry →Resource request
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.
The amount of CPU and memory a workload reserves, used by the scheduler to decide which nodes can hold it. It is a claim on capacity rather than a cap on usage, which is the distinction almost everyone gets wrong first. Requests set far above real usage are how a cluster ends up at thirty percent utilisation and a full bin at the same time, and requests set far below are how a node ends up oversubscribed and thrashing.
Full entry →