River wubx retweeted
New to Databend Cloud 🤔 ?
Databend Cloud Fundamentals is now on YouTube, with short lessons on data ingestion, SQL, and lakehouse operations.
Start with Episode 1: A Lakehouse for Modern Apps & AI Watch and subscribe: youtu.be/LsbmYRjz1Us
River wubx retweeted
Cloud Agent is in Preview 🚀 — your AI assistant for Databend Cloud.
Ask in natural language 💬.
Analyze data and workloads 📊.
Take action with your approval ✅.
One example: “Where can I reduce warehouse costs?”
bit.ly/4xXfQix
River wubx retweeted
On a controlled 100M-row test:
62.9% → 13.6% of blocks scanned.
63.5M → 13.6M rows read.
The difference: a workload-driven Databend Cluster Key.
Learn how to choose columns, key order & granularity—and verify the result 👉 bit.ly/4y3VF2k
River wubx retweeted
社区都在猜 Ox Alpha 是谁家的，我直接做了组对照实验（ox-alpha vs glm-5.3 vs mimo）：
🔍 Tokenizer 指纹：4 种探针文本（英文/中文/代码/随机串）下，ox 与 glm 的 token 增量逐项一致（-8/+10/+75），且差值恒为 75 token——内容 tokenize 完全相同，只差一个隐藏 system prompt。MiMo 完全不同。
🔍 CoT 指纹：中文提问，ox 和 glm 都用英文思考，开场都是 "The user is asking in Chinese..."，连 "classic question that trips up many people (and even some LLMs)" 这种句式都一样。MiMo 用中文思考，人设痕迹明显。
🔍 BIP39 采样：两者词池互相纠缠（harbor/velvet/canyon/mirror/meadow...），MiMo 零重叠。glm 还打出了社区原帖同款 "harbor velvet" 开局。
结论：基本确定是智谱未发布的 GLM-5.3 多模态变体（GLM-5.3 是纯文本，ox 支持图片/视频）。跟 @teortaxesTex 的判断一致
River wubx retweeted
Recent Snowflake threads repeat the warning: clustering helps pruning, but can burn credits when keys miss the workload.
Cluster keys aren’t indexes. Databend pairs automatic clustering with block metrics + SQL RECLUSTER controls.
Deep dive → bit.ly/4xSfQjz
River wubx retweeted
DeepSeek Harness puts Agent Trace in focus: long contexts, tool calls, and retries create a data problem.
Production proof: a leading AI company uses Databend Cloud for TB-scale ingestion, turning spans into eval-ready data—with less OSS overhead.
Reed 👉 bit.ly/4z19ujk
River wubx retweeted
Agent Trace pipelines fail at the boundaries: schema drift, small writes, and early offset commits.
bend-ingest-kafka batches events, commits after COPY succeeds, then Databend Stream + MERGE processes new data incrementally.
Engineering deep dive 👉 bit.ly/4g6pDLH
AI 会写代码，不等于 AI 会造软件。明天晚上我邀请 OpenLogos 作者黄湘龙，从百亿级生产系统的经验出发，聊聊如何用规格、场景、测试和验收，把不可控的 Vibe Coding 变成可交付的软件工程。 如果你有兴趣参与，欢迎一起交流。 #阿炳数记
River wubx retweeted
When agent traces reach TB scale, is “observability” still the right mental model?
They’re also eval datasets, replay records, and training feedback.
Should teams build the underlying data infrastructure themselves, or use a platform like Databend?
Here’s a deep dive into the engineering considerations👉 bit.ly/4v7Ddno
最近通过分学习 Cube 和 Snowflake 的 semantic 层建设，做了一个小的 demo : github.com/wubx/databend-sem… 同时也被朋友安利知识库，调研了一圈，目前还没真正能打开的 AI 时代的 wiki 。 感觉也是一个苦活。
🤖 Made with AI
River wubx retweeted
Immersive Translate needed product analytics without a heavy big data stack.
So they kept it simple:
📥 NDJSON on S3
⚙️ Databend Cloud scheduled Tasks
📊 SQL analytics + BI dashboards
💰 elastic warehouses with auto-suspend
🚀 POC completed in one afternoon.
Read the customer story: bit.ly/4bw6EZn
#DataEngineering #CloudDataWarehouse
参考 Cube 的SQL 语义编译做一个基于语义建模的 demo
github.com/wubx/databend-sem…
如果你对智能问数敢用还存疑虑，那么这个就是你要找的解决方案。 我们可以一起交流一下。
River wubx retweeted
AI can generate an AST Visitor migration fast.
Proving it didn’t change SQL semantics is the hard part.
Inside Databend: source-derived AST metadata, runtime traces, independent comparison, real SQL corpora, and coverage feedback.
Read the blog👉 bit.ly/4wGL6RX
#Database #AI
River wubx retweeted
Small bitmaps are cheap.
Until you aggregate millions of them.
In Databend, HybridBitmap keeps small sets lightweight and only switches to RoaringTreemap when the set actually grows.
The result: up to 80% faster bitmap aggregation in small-bitmap-heavy workloads.
Read the blog 📝bit.ly/4aH0cyt
#DataEngineering
River wubx retweeted
抽空做了个 Eval,评估下最近比较火的 GLM-5.2 能力
先说结论: GLM-5.2 的模型能力(理解 + 写代码)不输 Opus 4.6,真正的差距在"如何在真实环境里高效干活"——而这恰恰是最吃 harness / RL 训练的部分，也是 Cursor Composer 重度使用的方式。harness 调好了,这模型的威力还能再放出不少。
方法:同一道 Rust bug(serde_json #979),给 evot / claude-code / pi 三个 agent ,分别换 GLM-5.2 和 Opus 4.6 两个模型跑,对比它们的执行轨迹。
结果:
两个模型、6 个 session 全部 PASS,GLM-5.2 在"做对"这件事上没问题,bug 理解和最终代码质量都过关。
但代价差了一个量级:
• Opus 4.6:三个 agent 齐刷刷 18 轮收工,~80s
• GLM-5.2:38 / 43 / 61 轮,慢约 7 倍
拆 trace 看清楚了,差距不在懂不懂 bug,而在 agentic 执行这层:
1. 环境处理是软肋——GLM 有十几轮卡在 cargo 上反复试错(CARGO_HOME、cargo metadata、翻 serde 源码),Opus 一轮就绕过去了
2. 不收敛、爱反复验证——写完还手搓临时测试文件建了又删,thinking 字符数是 Opus 的几十倍
3. 但代码品味在线——GLM 最终的 fix 反而更克制地道,一个 match 合并三个 case,注释把设计意图讲清楚了,比 Opus 那版还干净
