drdr.xp

We just shipped llms.txt for Databend — a progressive skill system for your AI agent. Fetch docs.databend.com/llms.txt for Databend context One line. Your agent gets a structured index of everything Databend — SQL, data loading, tuning, troubleshooting. From there, it progressively fetches deeper docs (sql.txt, guides.txt, tutorials.txt) only when needed. No token waste. Pair it with our MCP server and your agent goes from knowing Databend to acting on your data — query, explore, build — all safely sandboxed. → docs.databend.com/llms.txt → databend.com/mcp/

@silsrc 建议放入部门共享候选池

最近被相亲介绍了十来个女生，结果都是对方有意看得上我，我看不上对方🤦

@plantegg 回归教育本质, 学习只负责向内求, 做买卖负责向外求

职高火爆 我表姐今年做了一个让全家炸锅的决定：她儿子中考考了全区前30%，她放弃了普高，直接送去读职高。 不是没考上，是主动选的。 她的逻辑是这样的：普高三年卷到死，考个普通二本，毕业出来月薪五六千，还得啃老两年才能站稳。但她儿子去的那个职高，跟本地几家德企有定向合作，学的是工业机器人方向，实习期就有六千底薪，转正八千起，还包五险一金。 我舅舅气得摔了碗："读职高丢不丢人？"我表姐回了一句："丢人能当饭吃吗？" 说真的，我现在越来越觉得她是对的。我身边985毕业的朋友，有三个在送外卖，两个在考公第四年。反倒是当年被我们看不起的那些读技校的同学，好几个已经在老家买了房开了店。 中国家长最大的执念就是"学历体面"。但体面不能还房贷，体面不能养孩子。当一个社会的本科生开始过剩，蓝领技术工反而成了稀缺资源。 职高火爆不是教育的倒退，恰恰是一部分人终于想明白了：路不止一条，而最挤的那条往往不是最好的。

@Nicebabycat 亲历的此梗还是之前公司老板逼着大家都用svn备份数据, 包括财务, 全部门一个大repo, 没有权限控制, 我不小心更新了一下根目录.

公司里工资高的和工资低的都沉默了

@BohuTANG 奥迪积分

目前最费功夫的是 AI 返回的 markdown 流式渲染。没什么好用的库，因为 AI 吐出来的大多不是标准 markdown。Claude Code 这块做得不错，但不支持数学公式渲染

@BohuTANG 人类, 你还有价值

自己从头造一个 coding agent 可能是当下最有情绪价值的事。关键是让它每一步都透明可见 —— 你会发现即使最顶尖的模型(Opus 4.7/GPT-5.5)也挺弱智的。比如直觉上上下文越大应该越强，但实际上越强的模型越容易在长上下文里迷路，必须强约束才能把它拉回来：比如它探索超过 N 轮还没收敛，就注入一条 system reminder，逼它停下来明确目标、给出阶段性结论。造完你会真正理解：AI 的智能全靠人类兜底...

New in #Databend: Geospatial goes deeper 🌍 - Geometry aggregate funcs in SQL - Refreshable Spatial Indexes - ST_DWITHIN-powered index pruning - Geo values encoded properly in Arrow results Less full-scan geometry. More pruning. Faster spatial queries. Built in Rust, running on your object storage. github.com/databendlabs/d…

OpenRaft v0.10.0-alpha.18: crates.io/crates/openraf… - Log-stage latency tracking across 6 Raft lifecycle stages - Customizable Batch<T> for log entry collections - Deterministic simulation testing with turmoil github.com/databendlabs/o…

@_4ever00 这么多海胆！

出来约会吃海胆饭 😼

Deterministic Simulation Testing (DST) is powerful. But what’s more powerful? It’s been a bit of a secret (till now) but TigerBeetle has actually been doing “Protocol-Aware” DST all this time. The “Protocol-Aware” part is key. Like PAR for consensus. But now applied to DST.

Built a benchmark suite comparing Rust histogram crates: Covers recording throughput, P99 query latency, memory, and accuracy across various distributions. github.com/drmingdrmer/ru…

base2histogram v0.2 — a 2 KB Rust histogram. Tiny memory. Fast recording. Low error. Built for latency tracking on hot paths. crates.io/crates/base2hi…

@tison1096 tom最后还是回来了不是吗

昨天重新刷旧番的时候正好看到这个了，哈哈。

Now base2histogram supports rescaling a WIDTH=2 histogram to one with WIDTH=4 and vice versa: #L302" target="_blank" rel="nofollow noopener">github.com/drmingdrmer/ba…

@v1gnesh @kellabyte I don't know if there is a standard approach. but what you described is a very commonly used one. :)

252 buckets. 2KB. Sub-0.2% percentile error. We built a histogram that encodes bucket boundaries like floating-point numbers — just bit shifts, no floats. The real win is trapezoid interpolation: accurate, zero extra storage. blog.openacid.com/algo/histogram/

@dotey Exactly. git-worktree is a just cognitive burden

没必要worktree，可以 clone 几份放在固定的目录，轮着用就够了，每次pull最新然后checkout一个新的branch，完成后提PR合并到main

base2histogram's key difference: trapezoidal interpolation. Instead of bucket midpoints, we estimate density gradient from neighbors and solve the inverse CDF. Same 2KB, O(1) — but P99 error drops from ~5% to <0.1% on real latency data. github.com/drmingdrmer/ba…

@BradfordToney Here's a detailed comparison: github.com/drmingdrmer/ba… TL;DR: DDSketch gives formal error guarantees; base2histogram achieves better practical accuracy (<0.3%) with fixed 2KB memory and pure integer ops — no log() on the hot path.

@drmingdrmer How does this compare with the sketch approach? Eg ddsketch?