siddontang

Really enjoyed Uber’s write-up on its internal AI Software Factory. I’ve never been a big believer in: AI productivity = buying everyone a Claude/OpenAI seat. That may help individuals, but at company scale you still need to manage models, context, tools, permissions, evaluation, and cost. Uber …

Great write-up. A lot of databases are built on RocksDB, and many performance issues ultimately come from LSM-tree behavior itself. We’ve also done quite a bit of work in TiKV to push RocksDB further — DB mutex, MVCC GC, compaction, flow control, write batch, etc. A few related notes from our si…

最近看了一篇很好玩的 paper：How to Write to SSDs https://t.co/dqCmvzyHPU 论文讲的是 write amplification 和 DB/SSD co-design，却让我重新想起了经典的 Law of Leaky Abstractions： https://t.co/gmIbbKyiTD DBMS 以为自己优雅地写了一个 page，SSD 却在地下室忙着 FTL、GC 和搬家，最后一次 logical write 变成了好几次 NAND write。 Architecture must follow hardware. 抽象可以关上地下…

看完 Coder 的 Agent Relay，越看越觉得系统设计最后都异曲同工。 数据库做：Storage / Compute 分离。 BYOC 做：Control Plane / Data Plane 分离。 现在 Agent 又开始做：Reasoning / Execution 分离。 Cursor 可以继续在云上想，但代码、Credential、真正的执行留在客户自己的环境里。 本质上好像都是一件事： 把变化快、需要弹性的东西拆出去；把真正有状态、有风险、需要自己控制的东西留下来。 系统设计折腾几十年，最后发现最常用的招还是： 先解耦。😂

Lots of discussion around WAL + S3 lately. I think it’s easy to oversimplify this as: just put WAL on S3. S3 can be the source of truth for data. But WAL is on the latency-sensitive commit path. That’s why in TiDB X we still use: Data → S3 Log → Raft + 3 replicas on EBS → S3 Object storage chan…