dbt Fusion: A Restrictive Rewrite of dbt Core

A day late on this one — I was too busy having a great time connecting, learning, and geeking out at #dbtCoalesce to post in real time. Worth it. Yesterday morning’s keynote made one thing clear: dbt Labs and Fivetran are building toward an open, AI-ready data future. The focus? Open Data Infrastructure — with Fivetran’s Iceberg data lakes, dbt Labs’s Fusion Engine, and a clear commitment to open-source innovation. dbt is open-sourcing Metricflow under an Apache 2 license and continuing to push Fusion as a stateful, developer-first engine built for speed, trust, and context. Some standout moments from the keynote and demo: - Fusion Engine (Preview) — state-aware orchestration, tuned configurations, and smarter builds driving up to 60% faster deployments and 40% cost savings. - AI + Metadata — dbt is leaning hard into AI by enriching context and lineage for truly “AI-ready data.” - VS Code Extension — inline CTE previews, real-time SQL feedback, and a “compare changes” feature that shows data diffs before you run anything. - Agents & MCP Servers — from developer and observability agents to an analyst agent (in beta!) that lets you ask questions about your data directly. It all ties back to one theme: dbt Labs is building an intelligent, open ecosystem for developers and partners — one where speed, clarity, and confidence come standard. For me, Day 2 confirmed it: the future of analytics is open, state-aware, and deeply connected. Anyone else geeking out over inline SQL previews and real-time data diffs? 😄

"Is bauplan an alternative to dbt?" Gold medal question of the dbt Labs - Fivetran merge week. As always, the answer is YES, BUT - YES, Bauplan can run any DAG composed of SQL and Python functions + the bauplanic touch: built-in transactions, containerization, contracts, and data versioning; - BUT "Can I run my dbt on Bauplan?" Well, everything expressed in YAML files (e.g. tests) requires minimal refactoring (our testing docs in the first comment). We are prioritizing a native integration to avoid that effort as well. The true question is: should transformations live inside the warehouse, or in an open lakehouse? We're on team lakehouse because we believe in freedom of choice for compute, language, and business model. In real life, people who ask that question are very pragmatic and care about 2 things: - Shifting-left: bringing bronze and silver onto object storage. Everybody needs Python as easily as SQL because of AI workloads, and everybody wants efficiency and flexibility. - Data infra-as-code: dbt is a first step in bringing software engineering to data workloads, but it's not the end of the journey. Moving out of the warehouse now means handling decoupled compute, which brings its own challenges if you don't have declarative, serverless infrastructure. To sum it up, the reason they consider Bauplan is its compatibility with dbt, but the reason they choose Bauplan is that they're looking for a more expressive and Python-friendly ecosystem. Your move.

🔮 Real-time at scale: InfluxDB 3 handles millions of writes/sec with sub-10ms reads. Real-time isn't a feature anymore-it's the base layer. AWS and InfluxData just dropped something that changes everything. InfluxDB 3 on Amazon Timestream marks the first managed service offering of this next-generation time series database. Here's what caught my attention: • Sub-10ms query response times • Millions of writes per second capacity • Built-in Processing Engine with Python VM • Custom logic execution on incoming data • Zero external service dependencies Built in Rust with Apache technologies, InfluxDB 3 delivers the performance AI/ML workflows actually demand. The architecture supports continuous data streams for model training and real-time decision-making. Evan Kaplan, CEO of InfluxData, puts it perfectly: "Real-time isn't a feature. It's the foundation of every intelligent, autonomous system." The enterprise version includes multi-region durability, automatic failover, and enhanced security. Its cloud-native, diskless design simplifies operations while supporting unlimited cardinality. Now available in 14 AWS regions, InfluxDB 3 becomes the default time series database in AWS Console. This positions both companies perfectly in the growing market for real-time AI infrastructure. For developers building intelligent systems, this combination removes infrastructure complexity while delivering the performance modern applications actually require. What's your take on this shift toward real-time as the base layer rather than just another feature? #AWS #InfluxDB #RealTimeData 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/dYsauCK5

A year ago, we launched Drasi to make building change-driven systems easier—applications that react instantly to data changes across distributed environments. Since then, Drasi has grown into a CNCF Sandbox project with an active community and a strong foundation for real-time, event-driven architectures. Today, we’re marking a major milestone: Drasi now supports the ISO-standard Graph Query Language (GQL) alongside openCypher. This dual support gives developers flexibility and a future-proof path for graph-based systems, enabling richer, context-aware applications. Learn more: https://msft.it/6045sfUR7 #OSS #Drasi #Developers

A year ago, we launched Drasi to make building change-driven systems easier—applications that react instantly to data changes across distributed environments. Since then, Drasi has grown into a CNCF Sandbox project with an active community and a strong foundation for real-time, event-driven architectures. Today, we’re marking a major milestone: Drasi now supports the ISO-standard Graph Query Language (GQL) alongside openCypher. This dual support gives developers flexibility and a future-proof path for graph-based systems, enabling richer, context-aware applications. Learn more: https://msft.it/6047sfYFW #OSS #Drasi #Developers

🚀 AI SQL agents are on the rise - and their demand is only going to grow. I just published a new article on Medium exploring two practical ways to build AI agents that can query SQL databases: 👉 as an MCP server (illustrated with Claude Desktop as an MCP client) 👉 and as a LangChain-based tool, now featuring both v0 and v1 examples, since LangChain released v1 right while I was preparing the post! I also looked into providing an n8n-based solution, but their SQL AI Agent node was removed back in February 2025 - and I couldn't find any simple and generic replacement (except for a Supabase-specific one). If you’ve managed to build something similar with n8n, please reach out - I’d love to learn more! If you’re curious about how to connect LLMs directly to your structured data, check out my full guide here 👇 https://lnkd.in/e_RqeuPg

A year ago, we launched Drasi to make building change-driven systems easier—applications that react instantly to data changes across distributed environments. Since then, Drasi has grown into a CNCF Sandbox project with an active community and a strong foundation for real-time, event-driven architectures. Today, we’re marking a major milestone: Drasi now supports the ISO-standard Graph Query Language (GQL) alongside openCypher. This dual support gives developers flexibility and a future-proof path for graph-based systems, enabling richer, context-aware applications. Learn more: https://msft.it/6042sAWZp #OSS #Drasi #Developers #SkilledbyMTT

I’ve been using the Gemini CLI and Model Context Protocol (MCP) lately — and honestly, it feels like the missing layer in modern data engineering. What used to be fragile glue code between models, APIs, and warehouses is now structured, permissioned automation. Gemini gives me a command-line interface to reason with models. MCP defines how those models connect — securely and contextually — to tools like BigQuery, dbt, or GitHub. Together, they turn AI from a sidekick into an orchestrator — one that understands your data topology, documents transformations, and even proposes pipeline configs safely. It’s not about what the model knows, but what it’s allowed to know. That’s the real unlock for bringing AI into the operational core of data systems. Read more in this post #DataEngineering #Gemini #MCP

When George Fraser from Fivetran said dbt was like “Rails for data” at Coalesce this week, it made me smile because that was exactly how I felt when I started with dbt. Ruby on Rails taught me a lot about; - capturing requirements from users - developing in an agile way - infrastructure is code - automated testing - automated deployment - ci/cd and code reviews - convention over configuration - simplicity and having fewer moving parts that can break - and how small teams can have a big impact An enterprise application that we delivered at Amgen back in 2012 is still in use. I was shadow IT doing everything different from the enterprise. No one used Postgres there at the time. No one knew what Ruby was and no one had 90% test coverage. We had just two developers and one product manager, me. Later on when I was part of the team shaping our new Data platform at Amgen I brought a lot of those lessons into what we built, but it wasn’t until I found dbt that I was finally able to do what I initially envisioned for what a good data architecture could be. dbt felt like Rails, but it didn’t go far enough. It still doesn’t go far enough too many decisions are left for users that have no experience and too many options to choose from one thing that rails does well is telling you exactly how to do something even if you don’t know why you’re doing it eventually, you appreciate why that decision was made. Rails encapsulates the many years of experience of the rails community into conventions that help you scale. This is how we think at Datacoves. How can users know what they don’t know when they’re starting out? How do they know that one decision today is gonna have a lasting impact in a year or two? As a community, we can still make things better and help each other mature the state of the art and analytics. At Datacoves we’re not consultants we’re just opinionated because we have learned what’s important at scale and even when working with small teams, we helped them accelerate their processes by not having to learn so many new things.

🌐 A new chapter for the Modern Data Stack 🌐 The rumors of Fivetran acquiring dbt Labs, coming so soon after the sqlmesh deal, are bigger than consolidation news — they’re a sign that the Modern Data Stack has reached a turning point. We’ve seen this cycle before: 1. Open-source breakthroughs ignite innovation (Airflow, BigQuery, dbt). 2. Growth accelerates, venture capital fuels expansion. 3. Consolidation and commercialization follow. 4. A reset begins. What happens next is more important than the deal itself. This could mark the end of the commercialisation era and the start of a new wave of tooling — AI-native, developer-friendly, and affordability-focused. But unless something changes, the same pattern will repeat. Too often, open-source projects evolve into commercial products tuned for margins rather than developer value. Without stronger governance, every cycle risks ending in a rug pull. Models like the Linux Foundation show us that community-first stewardship can safeguard innovation from being consumed by short-term economics. The takeaway: consolidation isn’t an endpoint. It’s a reset. And in that reset, there is opportunity — to rethink not just the tools we use, but the structures that sustain them. 🔮 The most interesting question now isn’t who owns ELT. It’s: what new models of technology and governance will define the next decade of data?