Dora Noda
Software Engineer
2331 posts · View all authors
IPVS Is Deprecated in Kubernetes 1.35: What Actually Changes When kube-proxy Moves to nftables
Kubernetes 1.35 deprecates IPVS proxy mode in favor of nftables. Here's the concrete before/after — the load-balancing algorithm you lose, the conntrack behavior shift, and a node-by-node rollout sequence — for migrating a self-hosted fleet's kube-proxy before the next cluster upgrade.
Dynamic Resource Allocation Comes of Age in Kubernetes 1.36: Scheduling GPUs Declaratively Instead of Renting a Managed Tier
Kubernetes 1.36 matures Dynamic Resource Allocation's GPU-scheduling features right as Fly.io kills its GPU line — here's what DeviceClass and ResourceClaim actually replace, and the real cost math between owning a Hetzner GPU and renting one by the second.
Kubernetes Just Deprecated externalIPs. Here's the Exact Migration Path for Bare Metal
Kubernetes 1.36 deprecated Service externalIPs over a security hole the project never patched. Here's the exact four-stage removal timeline and the commands to migrate a bare-metal fleet to MetalLB before the default flips in 1.40.
Kubernetes 1.36 Ships Gang Scheduling In-Tree: What It Actually Buys an AI-Agent Batch Fan-Out
Kubernetes 1.36 ships PodGroup and Workload APIs in-tree, letting the scheduler gang-schedule a multi-pod fan-out atomically. Here's exactly what the five alpha feature gates do, what they replace Volcano/Kueue for, and how far alpha still is from safe for tenant workloads.
Your Kubernetes Install Doesn't Have a Finish Line — And More Docs Won't Give It One
Why self-hosted Kubernetes installs for AI platforms stall for weeks, and why owning the node fleet with Cluster API fixes what better documentation can't.
Kubernetes 1.35 Killed the CPU Restart Tax. Memory Still Pays It.
Kubernetes 1.35 made CPU resizes truly restart-free, but memory resizes still restart the container by default. Here's the resizePolicy mechanics, the real constraints, and how a self-hosted PaaS should wire this into its tenant plan-upgrade flow.
Lagoon Outlived Its Mirantis Acquisition: What a Drupal-Born Kubernetes PaaS Still Gets Right
Lagoon, the Drupal-born Kubernetes PaaS Mirantis acquired in 2022, still automates per-branch config-imports and database provisioning that Coolify, Dokploy, Kubero, and Canine haven't fully matched — a concrete look at the mechanism and what a general-purpose platform has to build to catch up.
LiteLLM's Agent Platform Fixes Session Amnesia With One Postgres Table, Not a New Kubernetes Primitive
BerriAI's alpha-stage LiteLLM Agent Platform keeps an AI agent's session alive through a pod restart by storing state in Postgres instead of the sandbox itself — here's the actual mechanism, how it stacks up against E2B, Daytona, and Modal, and what it means for any MCP server with real deploy authority.
Manufact Raised $6.3M to Be 'Vercel for MCP Servers' — Here's Exactly What a Generic Git-Push PaaS Doesn't Give You
Manufact raised $6.3M to build a hosted cloud purpose-built for MCP servers. Here's the actual pricing, the actual feature list, and what deploying the same MCP server on a generic git-push PaaS gets you instead.