I build reliable systems for data and AI workloads. I am a TiDB maintainer and reviewer/approver with 250+ merged pull requests in the TiDB repository.
My distributed-systems work spans transaction correctness, GC safety, initial data distribution, observability, performance, and compatibility. I am now applying those engineering principles to AI infrastructure by building agentd, a durable single-host runtime for personal agents.
- Reliable agent runtimes — built and released agentd v0.1.0-alpha.1 with scoped execution, explicit capability boundaries, inspectable traces, and transactional delivery. Its independent Telegram adapter keeps transport concerns outside the core runtime.
- Transaction correctness — implemented and hardened Read Committed for pessimistic transactions across timestamp ownership, subqueries, executor fast paths, and lock semantics.
- Distributed GC safety — published per-server transaction watermarks and incorporated their global minimum into GC safe-point calculation.
- Scalability and operations — evolved DDL-time region split/scatter into configurable v8.4 defaults.
- Reliable agent runtimes and durable execution
- Capability boundaries, state ownership, and observable failures
- Inference serving, scheduling, and resource management
I write about systems and engineering at blog.minifish.org.