NVIDIA / nodewright / 37381006125

83%

Build Type

push

github

Committed by web-flow

Commit Message

fix(operator): raise interrupt container memory to 256Mi (#771)

The interrupt container ran under a hard-coded 64Mi request and limit.
On 64K-page arm64 nodes the Python agent (v6.x) peaked at that limit
while running reboot: two CPython interpreters (the agent and its
chroot_exec helper) plus reboot itself, with page tables and slab
inflated by the large page size. The memcg OOM killer took out the agent
after it had written the interrupt flag but before the reboot took
effect, so every retry skipped the interrupt and the node stayed
cordoned without ever rebooting.

Raise the request and limit to 256Mi through a named constant. The
container also runs systemctl for service restarts and package interrupt
scripts, so the headroom matters regardless of agent version, and the
node is drained when the interrupt runs, so the larger reservation costs
nothing in practice. Interrupt Jobs already in flight are not recreated:
podMatchesPackage does not compare resources on interrupt pods.

Document that interrupt pods use fixed resources that neither the
LimitRange nor a per-package override changes.

The flag-before-reboot half of #770 is fixed separately by agent v7.0.0,
which confirms node_restart by a changed boot ID.

Refs #770

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

Coverage Stats

2 of 2 new or added lines in 1 file covered. (100.0%)

16 existing lines in 3 files now uncovered.

9716 of 11716 relevant lines covered (82.93%)

7.64 hits per line