• 31 May 2026

  Building a GPU SaaS Platform - Worker Sidecar

  Part 15: add a serverless worker sidecar, define a UDS framework contract, and return results and metrics through NATS.
• 17 May 2026

  Building a GPU SaaS Platform - Activator Dispatch

  Part 14: add a dedicated activator that consumes ingress invocations, selects or creates GPUUnit workers, and publishes worker-targeted dispatch messages.
• 10 May 2026

  Building a GPU SaaS Platform - Queue-First Ingress

  Part 13: record the runtime-side serverless contract on GPU units and enqueue invocations durably through NATS JetStream before any worker executes them.
• 3 May 2026

  Building a GPU SaaS Platform - OverlayBD Cold Start

  Part 12: explain how OverlayBD converts OCI images into a lazy-pullable block format and why that changes cold start for large GPU images.
• 26 Apr 2026

  Building a GPU SaaS Platform - Security and Metrics

  Part 11: harden runtime workloads with capability drops, shared-memory sizing, egress isolation, and expose runtime and Nvidia GPU metrics.
• 12 Apr 2026

  Building a GPU SaaS Platform - Shared Proxy and SSH Access

  Part 10: add a shared storage proxy, switch the accessor to dufs, and expose GPUUnit SSH through sidecars and frp.
• 4 Apr 2026

  Building a GPU SaaS Platform - Storage Data Jobs

  Part 9: add storage prepare jobs, the first accessor path, and recovery state to GPUStorage.
• 28 Mar 2026

  Building a GPU SaaS Platform - Storage Lifecycle

  Part 8: introduce GPUStorage, mount persistent data into GPUUnit, and separate data lifecycle from runtime lifecycle.
• 26 Mar 2026

  关于 AI Coding 的吐槽

  从开源质量、架构演进、生产可靠性和工程责任的角度，聊聊我为什么反对在缺少 review 与验证时把 AI Coding 全权交给 agent。
• 21 Mar 2026

  Building a GPU SaaS Platform - One Unit, One Controller

  Part 7: collapse stock and runtime into one GPUUnit resource, seed stock explicitly, and hand off warm units into active runtime.