Making difficult systems easier to understand.

A practical look at how Kthena separates prefill and decode workloads to run LLM inference efficiently on Kubernetes.

A code-first breakdown of how Volcano checks and rewrites workloads before the Kubernetes API server stores them.

Explore Volcano's plugin framework, Dominant Resource Fairness (DRF), queue fairness, and how scheduler plugins influence resource allocation decisions in Kubernetes.

An in-depth walkthrough of Volcano Scheduler's core scheduling cycle, Sessions, Enqueue and Allocate actions, and how the scheduler avoids deadlocks during batch scheduling.

A deep dive into Volcano's source code — how Kubernetes batch scheduling works under the hood, from gang scheduling to state machines.

Most React apps are slow for the same three reasons: too many re-renders, too much in the DOM, and work that blocks the main thread. This article addresses all three.

Volcano is a batch-scheduling system for Kubernetes. Here's what I learned reading its dashboard source while preparing to contribute and apply for LFX mentorship.

What the memory problem in early AI models was, why it mattered, and how the Transformer changed everything.

Apache Fory lets you define data types once in a schema file. Java, Python, Go, Rust, C++, C# all had code generation — JavaScript didn't. Here's how I added it.

Apache Fory supports a cache-friendly row format for analytics workloads. A walkthrough of the format design, memory layout, and how it compares to standard serialization.