Skip to content

Instantly share code, notes, and snippets.

View jackfrancis's full-sized avatar

Jack Francis jackfrancis

View GitHub Profile
@jackfrancis
jackfrancis / cluster-autoscaler-failure-recovery-plan.md
Created November 14, 2026 00:42
Cluster Autoscaler: incremental GCE/Azure failure recovery E2E and integration coverage plan

Cluster Autoscaler: incremental plan for multi-pool failure recovery coverage

Proposed PR series, prepared 2026-11-13.

Related change: kubernetes/autoscaler#10680, which makes Azure TargetSize report zero without an error for a missing VMSS.

Goal

Add deterministic coverage for this sequence:

@jackfrancis
jackfrancis / kubernetes-native-node-autoscaling-scheduler-integration-summary.md
Created November 10, 2026 22:56
Kubernetes-native node autoscaling: scheduler integration, alpha scope, nominations, consolidation, and API evolution

Kubernetes-Native Node Autoscaling: Scheduler Integration, Nominations, and Consolidation

The clearest direction was to start with a lean alpha built around a PostFilter plugin, minimizing changes to existing scheduling behavior. A major architectural clarification was that sharing scheduling logic and guaranteeing that a simulated placement remains available are two different problems.

This chronological summary is based on an automated transcript of a roughly 53-minute discussion. Timestamps are approximate; proposed designs and scheduling targets are not finalized API decisions or release commitments.

1. Provisioning flow and NodeClaim batching - 00:00-07:55

The proposed flow was to evaluate potential capacity after ordinary scheduling and preemption fail. The scheduler could determine synchronously whether a pod fits a prospective NodeClaim, while the actual infrastructure provisioning remains asynchronous.

@jackfrancis
jackfrancis / kubernetes-native-node-autoscaling-discussion-summary.md
Created November 9, 2026 23:22
Kubernetes-native node autoscaling: MVP scope, scheduler integration, migration, DRA offerings, and capacity buffers

Kubernetes-Native Node Autoscaling: MVP Scope and Scheduler Integration

The central question was how to make node autoscaling a native Kubernetes capability without preserving a second, divergent scheduling implementation. Participants debated how much architectural work belongs in the MVP and which follow-ups would make it useful enough to adopt.

This is a chronological summary from an automated transcript. Timestamps are approximate; proposals and preferences below should not be read as finalized API decisions or release commitments.

1. MVP integration point and batching - 00:12-05:15

The opening question was whether provisioning should initially run through kube-scheduler's PostFilter extension point, or whether the MVP should already introduce broader framework changes. Participants discussed starting work in the 1.39 cycle.

@jackfrancis
jackfrancis / kubernetes-ai-workload-scheduling-summary.md
Created November 9, 2026 22:13
Kubernetes AI workload scheduling discussion: reservations, topology, sequencing, and disruption

Kubernetes AI Workload Scheduling Discussion

The discussion focused on what Kubernetes scheduling primitives workload controllers need for AI workloads. Most topics ended with requirements or possible approaches - not finalized API decisions.

1. Capacity reservations - 00:00-04:45

Controllers need to preserve capacity and placement during operations such as rolling updates, rather than letting other queued workloads take the freed resources. Participants distinguished hard reservations from soft reservations, where other, preemptible workloads could temporarily use the capacity. The immediate need was to establish concrete use cases before choosing the mechanism.

2. Preferred versus required topology - 04:45-10:40

Why do English countries (England, Canada, The United States, Australia) have such a high incidence of inventing sports, both for their own amusement, and those that are exported around the world: the many types of football (association football for export; Canadian Football, American Football, and Australian Rules Football for local amusement), Rugby, Cricket, regulated Horse Racing, Baseball, Tennis, Badminton, Ice Hockey, Basketball, Volleyball. Why don't English colonies of non-"English" (meaning generally Anglo-Saxon-dominant peoples) peoples invent sports, and instead adopt the English ones (for example India with cricket, African nations with football, Hong Kong with horse racing, badminton, and football)? Why didn't non-English colonies adopt their non-English rulers' sports, but instead also adopted the English sport (Argentina and Brazil adopting football, for example)? Is there something unique about English culture that suggests a predeliction for inventing rules-based athletic games that othe

@jackfrancis
jackfrancis / count_provider_commits.py
Created May 22, 2026 21:46
Count commits per cloud provider in kubernetes/autoscaler/cluster-autoscaler/cloudprovider over a time window
#!/usr/bin/env python3
"""
Count commits per cloud provider in kubernetes/autoscaler's
cluster-autoscaler/cloudprovider/ tree over a given time window.
How it works
------------
GitHub's REST Commits API supports two filters we exploit:
- `path` : restrict to commits that touched files under this path
- `since` : ISO-8601 timestamp lower bound

Here's what Cluster Autoscaler will do in your scenario, end‑to‑end. The behavior is entirely driven by the fact that the Kubernetes Node objects still exist (kubelet stops, but the API server retains them) and the Azure cloud provider does not surface VM power state to CA.

How CA classifies the 33 stopped nodes

CA buckets every node every loop in clusterstate.go (updateReadinessStats). The buckets are mutually exclusive:

@jackfrancis
jackfrancis / ray-azure-agent.md
Created March 5, 2026 19:48
Ray Azure Provider plan

Agent Workflow: Reinforce azure support

Auto-generated by providerize on 2026-02-24 22:56 UTC Repository: https://github.com/ray-project/ray

Context

This repository has cloud provider integrations for: aws, gcp, azure.

Current ranking (best → worst):

@jackfrancis
jackfrancis / kuberay-multicluster.md
Created March 4, 2026 19:33
KubeRay Multicluster analysis

KubeRay Federation Proposal — Analysis & Comparative Landscape

Proposal Summary

Issue #4561 by @yuchen-ecnu proposes adding federation capability to KubeRay so that a single logical RayCluster can span multiple Kubernetes clusters. The core motivation:

  • Fragmented GPUs: Organizations procure GPUs across multiple cloud vendors/AZs. Today these are isolated into separate K8s clusters, preventing a unified Ray cluster.
  • Operational pain: Users must split datasets, deploy multiple small RayClusters, and manually manage them — causing long-tail performance issues and complexity.
  • Virtual Kubelet limitations: The common workaround (aggregating via Virtual Kubelet) creates control-plane scalability bottlenecks, especially at scale (e.g., 10K→400K+ cores in an hour).
@jackfrancis
jackfrancis / gist:9ecc5964c1b4f2af10894d7dfac9368c
Created January 28, 2026 16:24
jack-francis-talks-kubecon-eu-2026
Here are my two talks:
The first one will be a joint session w/ GKE (and fellow SIG Autoscaling TL) describing the OSS CA and our plans to make significant changes in 2026.
The tl;dr is that GOOG will be investing significantly in the OSS CA to correct a decade of forking and blackbox dev.
This is directly due to investments that we have made in the SIG to modernize the OSS surface area, release process, E2E tests, etc.
- https://maintainersummiteu2026.sched.com/event/2EWf1/cluster-autoscaler-evolution-kuba-tuznik-google-jack-francis-microsoft