While implementing the HNSW algorithm for my vector database project vex (initial code AI-assisted, debugging and optimization done by hand), I ran my first Recall test — every single metric was 0%. Not 50%, not 30%. Zero. The HNSW results and BruteForce results had absolutely no overlap. This post documents the entire journey from 0% to 98%+ recall, cross-referencing each bug with the original HNSW paper.

Writing business logic with Temporal is pure joy — “code like a single-machine program, get distributed fault tolerance for free.” But after the honeymoon phase, a nagging question remains: Is it really reliable? What if the Server crashes? What about network partitions? What if the rollback itself fails? This article tears apart these questions from the ground up.

Last post, we hand-rolled Flannel in Kind cluster and witnessed the 50-byte per-packet VXLAN encapsulation overhead via tcpdump. This time, we replace Flannel with Cilium, leveraging eBPF to completely eliminate this “network tax”, and use Hubble for traffic visualization.

This weekend, I built a “networkless” cluster from scratch on Manjaro using kind, manually debugged missing kernel modules and CNI plugins, and finally witnessed the VXLAN encapsulation process with tcpdump. This post documents the entire journey.

K8s Controllers and Temporal are both “control plane” technologies, but they represent fundamentally different design philosophies. This post deep-dives into their core differences and explores how to combine them in AI Infra platforms.

While building a trading system backend, I used Temporal to solve the classic distributed transaction consistency problem. This post breaks down Temporal’s core mechanisms and explores its applications in AI task scheduling.