WhatBrokeToday
GTA 6FAQ
Kubernetes

Kubernetes kube-scheduler: DRA Extended Resource Scoring Uses Stale Caches

medium

Is it fixed? Not yet.

Ongoing since Nov 13, 04:32 UTC. First reported Nov 13, 04:32 UTC.

What happened

The kube-scheduler NodeResourcesFit scoring path for DRA-backed extended resources caches per-device CEL selector results and per-node NodeSelector results in sync.Maps that are never invalidated for the scheduler's lifetime. This leads to stale device and node match caches, resulting in incorrect capacity calculations until a scheduler restart.

What caused it?

DRACaches deviceMatchCache and related caches store results in sync.Maps without an invalidation mechanism over the lifetime of the scheduler process.

No verified workaround yet

None of the linked reports describe a mitigation. This page updates when one appears.

Official fix

Timeline

  1. Investigating(current)
  2. Workaround available(upcoming)
  3. Resolved(upcoming)
  1. First report on GitHub issue

    scheduler: DRA extended resource scoring uses stale device-match and node-match caches (never invalidated, wrong capacity until restart)

    Source

Evidence

1 report · 1 GitHubShow
  1. [1]
    GitHub First report
    scheduler: DRA extended resource scoring uses stale device-match and node-match caches (never invalidated, wrong capacity until restart)(opens GitHub issue in a new tab)

    What happened? /kind bug /sig scheduling /sig node /area kube-scheduler /wg device-management The NodeResourcesFit scoring path for DRA-backed extended resources (/pkg/scheduler/framework/plugins/noderesources/resource_allocation.go) caches per-device CEL selector results and per-node NodeSelector results in two sync.Maps that are never invalidated for the lifetime of the scheduler process: DRACaches.deviceMatchCache — keyed by (CEL expression, driver, pool, device name), but the result also…

    by zl86790 0upvotes 4comments

FAQ

What happened in the Kubernetes issue?

The kube-scheduler NodeResourcesFit scoring path for DRA-backed extended resources caches per-device CEL selector results and per-node NodeSelector results in sync.Maps that are never invalidated for the scheduler's lifetime. This leads to stale device and node match caches, resulting in incorrect capacity calculations until a scheduler restart.

What is affected by the Kubernetes issue?

Kubernetes.

When did the Kubernetes issue start and end?

It started Nov 13, 04:32 UTC and is still ongoing (6 d so far). This page updates as it develops.

How do I get alerts for Kubernetes?

Subscribe to the Kubernetes RSS feed for every new Kubernetes incident, or connect an AI assistant to WhatBrokeToday over MCP and ask whether Kubernetes is down. Kubernetes feed

Where does this information come from?

WhatBrokeToday collects public status pages, developer community reports and security advisories, groups the ones about the same problem, and links every report under Evidence on this page. This one has 1 report from 1 independent source.