Bigtable at Twenty: What Stayed and What Got Rebuilt
SIGMOD Companion '26 has a 13-page paper by 50+ Google authors revisiting Bigtable two decades on. 10 EB, 7 billion QPS, and an architecture that barely moved. This is my reading of why.
Think and Act as One
CEO & Founder of Greptime. Writing about systems, distributed computing, AI & Programming.
Latest posts
Why Now: The Eight-Year Wait for Unified Observability
An idea someone had fully worked out back in 2018, yet it took eight years before it could actually be built. This piece is about one thing: why now.
When Agents Go to Production: What Actually Breaks
When agents go to production, they break the four assumptions today's infrastructure was built on: deterministic, human-driven, request-response, stateless. A first-principles walk through what actually breaks.
I Burned 23 Billion Tokens to "Rewrite" Langfuse
A migration retro: forking Langfuse to swap its analytics store from ClickHouse to GreptimeDB, run mostly by two AI agents. Most of the work was paying down the debt of one decision — make a single database both the source of truth and the analytics layer.
The Three Pillars of Observability: The Unification That Never Quite Arrived (Part 2 of 2)
Part 1 was about how the three pillars split apart. This part is about how many smart people, starting in 2018, tried to put them back together — what they actually built, and why none of it quite reached the finish line.
The Three Pillars of Observability: A History No One Planned (Part 1 of 2)
Today we treat metrics, logging, and tracing as the natural structure of observability. But it wasn't designed; it grew. Part 1, on how it split apart (2010–2017).
When Systems Turn Uncertain: How Datadog Sees Observability in the AI Era
Datadog's Investor Day deck quietly redefines observability: the object of observation turns probabilistic, AI agents become operators, and the reader of the data shifts from human to model.
TMA1 v2: Making the Agent Loop actually loop
TMA1 v2 adds an MCP server, enhanced hooks that auto-inject build/session/anomaly context, and cross-agent context sharing between Claude Code and Codex.
Observing my coding agents without leaving localhost
TMA1: a local observability tool for AI coding agents with full session trace, tool decision breakdown, latency tracking, and SQL-queryable storage.
Six Ways to Make Your Infrastructure Agent-Friendly
Have we been building infrastructure for humans this whole time, without ever treating agents as users? Six principles for agent-friendly infrastructure, rooted in Unix philosophy.
Is Database MCP Just Unnecessary Overhead?
My colleague called Database MCP 'pulling your pants down to fart.' After thinking about it, I mostly agree—but we built one anyway.
My AI Coding Experience: Three Principles
AI Coding doesn't fundamentally make you a better programmer. It's an amplifier—amplifying your strengths and your weaknesses.
Using AI ≠ AI Native: What It Really Means to Transform How Teams Work
The difference between using AI as a tool you pick up occasionally versus treating AI as infrastructure you don't know how to work without.
Rethinking Software Design Principles in the Age of Vibe Coding
When AI can rewrite an entire module in seconds, do the design principles built for 'humans maintaining code' still apply?
The AI Productivity Divide: From 10x to 100x Engineers
AI's real value isn't in producing the final output—it's in creating Builders. Whether you can create Builders is what separates the 2x from the 100x.