Kubernetes: Watch Costs and Pod Group Preemption

API server watch and caching saw a coordinated efficiency push around allocations and progress checks, while scheduler work advanced Pod Group preemption and Job controllers fixed status accounting. The pattern is lower overhead at scale and more reliable batch scheduling.

Duration: PT2M37S

Episode overview

This episode is a short developer briefing from Kubernetes.

It explains recent repository work in plain language.

  • Show: Kubernetes
  • Published: 2026-10-12T13:02:21Z
  • Audio duration: PT2M37S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

Good morning, it's Monday, October 12th, 2026, and this is your Kubernetes developer briefing.

The lead today is efficiency at scale. A cluster of API server changes targets memory and lock contention in the watch path, while scheduling work makes gang preemption cost-aware.

First, watch cache costs. Pull request 142935 replaces full deep copies with shallow copies when converting added and modified watch events, matching the strategy already used for list responses. That directly cuts allocations on hot paths that fan out to many watchers. Related cleanup in 142904 avoids checking…

Second, workload-aware scheduling. Pull request 142932 implements pod group preemption with victim-aware scoring, fixing prior behavior where potential victims were stripped and all nodes looked equally empty, scattering preemptors onto high-cost victims. Cleanup in 142928 unexports the children field of pod group…

Finally, Job reliability. Pull requests 142923 and 142922 fix over-counting: creation metrics no longer count a failed create as both failed and succeeded, and ready status is no longer decremented when a pod delete fails.

What's next: watch for a dedicated periodic to run the new allocation…

Nearby episodes from Kubernetes

  1. Weekly Recap - Go 1.27 Hardening, Scheduling and Storage Correctness
  2. Storage Efficiency and Scheduler Resilience
  3. API Stabilization and Controller Reliability Fixes
  4. Pod Group Hardening and Node Reliability Fixes
  5. Storage Correctness and Kubelet Admission Fixes
  6. DRA Hardening and Validation Push
  7. Scheduler Preemption, Storage Consistency, and Exec Fixes
  8. Storage Consistency and Kubelet Edge Fixes