oracle / karpenter-provider-oci / 34515475349

83%

Build Type

push

github

Committed by web-flow

Commit Message

fix: model OCI VM memory overhead in instance type capacity (#73)

OCI reserves memory below the guest, so a VM shape advertised as 32 GB
presents roughly 30.89 GiB of MemTotal to Linux. KPO modelled capacity as
the declared figure, so Karpenter over-estimated what a node could hold,
launched a node the pending pod could not fit on, and - having no feedback
loop - repeated that launch indefinitely. #69 reported a NodePool going from
0 to 50+ nodes in a few hours.

Subtract a configurable overhead from each VM shape's modelled capacity:

    overheadMiB = max(base + perGB * declaredGiB, percent * declaredMiB)

Defaults are 600 MiB base and 19 MiB/GiB. These clear every measurement
reported on #7 (A1/E4/E5/Standard3, 8-128 GB) by at least 71 MiB. They are
deliberately not the tightest fit: the tightest lower bound over the same
points clears the closest of them by only ~9 MiB, too little to survive an
image build or shape family nobody has measured. The extra headroom costs
~0.15 GiB on a 128 GB shape.

Bare metal shapes are exempt and keep their declared memory: the overhead
models a hypervisor, and there is none beneath a BM instance. Those shapes
follow a different model, which is not covered here.

A percentage form is also accepted and combined with max(), so setting it
can only make the estimate more conservative. Setting all three values to 0
restores the previous behaviour.

The overhead applies to capacity rather than InstanceType.Overhead because
it is not reserved memory in the kubelet's sense - the guest never sees it.
Core's InstanceTypeOverhead has slots only for KubeReserved, SystemReserved
and EvictionThreshold.

The failure modes are not symmetric: over-estimating causes an unbounded
launch loop, under-estimating wastes a little memory. The requirement is a
safe number, not an exact one, and the regression test enforces that by
requiring headroom rather than mere non-exceedance. Declared memory is
floored rather than rounded, since flex... (continued)

Coverage Stats

74 of 77 new or added lines in 4 files covered. (96.1%)

9 existing lines in 2 files now uncovered.

5151 of 6446 relevant lines covered (79.91%)

0.92 hits per line