NVIDIA / nodewright / 41511736622

84%

push

github

feat(operator): set a soft memory limit from the container memory limit (#932)

Nothing set GOMEMLIMIT, so with the default GOGC=100 the heap grew to
about twice the live heap before a collection and the container limit had
to cover that, with an OOM kill as the only signal of a limit set too low.

At startup the manager now reads the container's memory limit from cgroup
v2 (memory.max) or v1 (memory/memory.limit_in_bytes) and sets the Go soft
memory limit to GOMEMLIMIT_FRACTION of it, 0.9 by default. An explicit
GOMEMLIMIT wins, a container without a limit is left alone, and a fraction
outside (0, 1] fails startup. The chart exposes the fraction as
controllerManager.manager.env.memoryLimitFraction, and
resources-at-scale.md covers sizing, including the GC CPU cost of a soft
limit near the live heap.

Closes #904

Signed-off-by: Brian Lockwood <lockwobr@gmail.com>

53 of 60 new or added lines in 2 files covered. (88.33%)

10447 of 12447 relevant lines covered (83.93%)

7.68 hits per line