Mellanox / nic-configuration-operator / 38883510816-1

60%

Build Type

push

github

Committed by web-flow

Commit Message

fix: scope daemon NicDevice and Node caches to the local node (#546)

Every configuration daemon cached all NicDevice and Node objects in
the cluster, so per-pod memory grew with cluster size and the total
across the cluster grew with its square. On large clusters the daemon
can exceed its memory limit during the initial list.

Declare .status.node as a selectable field on the NicDevice CRD and
restrict the daemon's NicDevice informer to its own node. Restrict the
Node informer to the daemon's own Node with a metadata.name selector.

At startup the daemon checks that the API server accepts the
status.node field selector. When the installed CRD predates the
selectable field, it logs this and keeps the cluster-wide NicDevice
cache, so an upgrade that leaves the old CRD in place keeps working.

A NicDevice without status.node is outside the scoped cache. Device
discovery therefore re-reads an existing CR through the uncached API
reader when Create reports AlreadyExists.

Downstream report: https://github.com/NVIDIA/aicr/issues/3149

Signed-off-by: Mike Cook <micook@nvidia.com>

Coverage Stats

20 of 61 new or added lines in 3 files covered. (32.79%)

9 existing lines in 2 files now uncovered.

6556 of 11016 relevant lines covered (59.51%)

0.6 hits per line