Eviction tests incorrectly evict pods when stats API fails, leading to unexpected behavior.
Eviction tests incorrectly evict pods when stats API fails, leading to unexpected behavior.
Problem statement normalized from the linked evidence
CANDIDATE VIEW · NOT A PUBLISHED OPPORTUNITY
01
Qualification checklist
Gate decision: Aug 15, 2026
×
At least 3 independent users2 / 3
NOT MET
×
At least 2 independent discussions1 / 2
NOT MET
✓
At least 1 supporting signalPersistence
MET
02
Observed signals
Users2
Discussions1
Mentions2
Active days2
Source coverageGitHub
TrendUNKNOWN
First seenFeb 29, 2024
Last seenAug 09, 2026
Why the gate is holding
Fewer than 3 independent users
Fewer than 2 independent discussions
Competition research, commercial scoring, and Build Gap analysis do not run until this gate passes.
03
Source evidence
2 evidence records
github · HIGH · Aug 09, 2026
kubernetes/kubernetes
I reproduced this behavior against the current `master` branch. Before working on a fix, I'd like maintainer input on the expected behavior and scope. The manager-level test has two active pods under nodefs pressure (`750Mi` available against a `1Gi` threshold). The stats summary succeeds but includes only the lower-priority pod, which is also using more disk. The pod without stats has higher priority, but the eviction manager ranks and selects it first. When I add its stats, the measured high-usage pod is selected…
Eviction manager: When stats calls failed, we evict the pod that we couldn't get stats for.
### What happened? As I was investigating the eviction tests, I discovered an interesting problem. We are seeing a wide range of flakes where the one pod that should not be evicted gets evicted in the eviction test. Eviction tests work by assigning one pod that it should never be evicted and then it ranks these pods and evicts the other pods. If there is a stats API failure, we evict the pod that we couldn't get stats for. This seems to be by design (cc @derekwaynecarr) ### What did you expect to happen? We should…