Adapt or fall behind: a review of the Databricks executive survey on AI and its use
Understand how 600 tech leaders think about AI and find out how to future-proof your business before adopting the emerging tech.
Lakebase, Agent Bricks, and the Future of the Lakehouse: Highlights from Databricks Data+AI Summit 2025 Day 1
13 Jun 2025 · 8 min read
Budapest, June 13, 2025
Day 1 of the Databricks Data + AI Summit 2025 brought a wave of announcements that reshape how we build data systems and AI applications. As European practitioners, we’ve distilled the key innovations, what they do, why they matter, and how to start applying them. Here’s an enhanced, technically detailed breakdown of the announcements, grounded in Databricks’ official releases and documentation.
Lakebase (Public Preview): Real-Time OLTP + Analytics in One
What it is:
A new fully managed, Postgres-compatible database built directly into the Databricks Lakehouse, combining real-time transactions with analytics in one governed environment.
Technical Deep Dive:
- Architecture: Lakebase is a fully managed, Postgres-compatible operational database, powered by Neon technology. It separates compute and storage, allowing each to scale independently for cost efficiency and performance. Compute autoscaling is continuous, supporting bursty agent workloads without manual intervention.
- Integration: Native integration with Unity Catalog ensures centralized governance, security, and lineage tracking across both operational and analytical data.
- Transactional Guarantees: Supports full ACID-compliant transactions, branching (for safe experimentation and rollback), and point-in-time recovery, making it suitable for mission-critical workloads.
- API Compatibility: Exposes standard Postgres APIs, enabling seamless migration of existing apps and tools.
- Lakehouse Convergence: Data is stored in low-cost object storage, unifying operational and analytical workloads. No more ETL lag between OLTP and analytics.
Why It Matters:
Lakebase eliminates the classic separation between OLTP and analytics, letting developers build real-time, AI-powered apps directly on the lakehouse. This reduces architectural complexity and latency and supports next-gen use cases like AI agents that require fresh, reliable data.
Practical tip:
Start by migrating lightweight operational tables—like user preferences or feature flags—into Lakebase. You can write through Postgres APIs, then query live in Delta with full governance. This avoids duplicated pipelines and simplifies the architecture.
Sources:
https://www.databricks.com/product/lakebase
Agent Bricks (Beta): Build, Benchmark & Deploy AI Agents in One Step
What it is:
A low-effort, high-trust way to build AI agents: just describe the task, and Databricks handles data, tuning, evaluation, and deployment automatically.
Technical Deep Dive:
- Declarative Agent Framework: Users define agent tasks at a high level (e.g., “extract invoice fields”), and Agent Bricks automates the rest: data generation, evaluation, tuning, and deployment.
- Synthetic Data & Benchmarking: Leverages Mosaic AI research to generate domain-specific synthetic data and task-aware benchmarks, enabling robust, repeatable agent evaluation.
- Optimization: Automatically selects and tunes models using advanced techniques, optimizing for both cost and quality without manual trial-and-error.
- Observability & Governance: Built-in enterprise controls, cost/quality metrics, and MLflow 3.0-based observability for end-to-end traceability and compliance.
Why It Matters:
Agent Bricks addresses the bottleneck of moving AI agents from prototype to production by automating evaluation and optimization, ensuring trustworthy, cost-effective, and governable deployments on enterprise data.
Practical tip:
For Q2 or Q3, choose a pilot task, like a support chatbot or document extractor, and try Agent Bricks. From day one, you’ll get concrete outputs with performance metrics and automated tuning.
Sources:
https://www.databricks.com/product/artificial-intelligence/agent-bricks
LakeFlow GA + Designer: Self-Service ETL You Can Actually Trust
What it is:
Databricks’ native orchestration tool for building reliable data pipelines, now with a visual drag-and-drop interface that makes self-service ETL fast and intuitive.
Technical Deep Dive:
- LakeFlow Designer (Preview): A no-code, drag-and-drop pipeline builder with a GenAI-powered assistant for natural language pipeline creation, enabling analysts to build production-grade pipelines visually.
- Unified ETL Platform: Engineers can extend pipelines with full IDE support, adding custom logic in Python or SQL.
- Connectors: Out-of-the-box connectors for major enterprise systems (Google Analytics, ServiceNow, SQL Server, SFTP).
- Governance: All pipelines are governed and lineage-tracked via Unity Catalog, eliminating “shadow ETL” and ensuring compliance.
Why It Matters:
LakeFlow bridges the gap between agility and governance, letting analysts and engineers collaborate seamlessly on reliable, scalable data pipelines.
Practical tip:
Try setting up a LakeFlow pipeline for weekly marketing metrics. Let analysts assemble data flows, then let engineers optimize and scale components with Python or SQL, all within the same system.
Sources:
https://www.databricks.com/blog/announcing-lakeflow-designer-no-code-etl
MLflow 3.0 + Serverless GPU Compute (Beta): GenAI Done Right
What it is:
A major upgrade for managing GenAI workflows, with prompt tracking, agent monitoring, and serverless GPU power for easy scaling, and no infrastructure headaches.
Technical Deep Dive:
- MLflow 3.0: Introduces prompt registries, versioning, and comprehensive agent observability, tracking the full lifecycle of GenAI agents, including prompt changes, model versions, and cost metrics.
- Serverless GPU Compute: Provides instant access to A10g GPUs (with H100s coming soon) for training and inference, fully managed and autoscaled. Integrated with Notebooks, Unity Catalog, and MLflow, it eliminates the need for manual cluster management.
- Environment: CUDA 12.4, torch 2.6.0, torchvision 0.21.0, with stable environment versions for compatibility and security.
- Use Cases: Deep-learning-based forecasting, fine-tuning, computer vision/audio, and recommender systems.
Why It Matters:
MLflow 3.0 and serverless GPUs bring enterprise-level observability and scalability to GenAI, making it easier to manage, track, and scale LLM and agent deployments without DevOps overhead.
Practical tip:
Deploy an LLM or RAG agent in staging, then use MLflow to track prompt changes, model accuracy, and cost. Run inference via serverless GPUs to test real-time performance without cluster tuning.
Sources:
https://www.databricks.com/blog/mosaic-ai-announcements-data-ai-summit-2025
https://docs.databricks.com/aws/en/compute/serverless/gpu
AI SQL Functions + Storage-Optimized Vector Search
What it is:
AI SQL Functions: Built-in SQL functions let you process documents, generate embeddings, and tap into GenAI, right from a simple query.
Storage-Optimized Vector Search: Scalable, cost-efficient vector search for semantic and RAG-based applications, fully managed and deeply integrated with the Lakehouse.
Technical Deep Dive:
- AI SQL Functions: Functions like ai_parse_document are now up to 3× faster and 4× cheaper, enabling scalable document parsing and information extraction directly in SQL.
- Storage-Optimized Vector Search (Public Preview): Supports billions of embeddings at ~7× lower cost, making large-scale semantic search feasible within the lakehouse.
- Integration: Both features are governed by Unity Catalog, ensuring security and compliance for sensitive data.
Why it matters:
Parsing documents and semantic search are critical in legal, support, and R&D contexts. These tools embed multimodal/embedding power directly into SQL, governed, scalable, and cost-effective.
Practical tip:
Run ai_parse_document(…) on sample contract sets and index document embeddings via vector search. Then compare structured output versus the original PDFs. This can add instant value to search and understanding processes.
Sources:
https://www.databricks.com/blog/mosaic-ai-announcements-data-ai-summit-2025
High-Scale LLM Model Serving
What it is:
An enterprise-ready engine for serving LLMs at massive scale, supporting real-time AI apps with high throughput and low latency.
Technical Deep Dive:
- LLM Serving Engine: New proprietary inference engine supports 250,000+ QPS, leveraging custom kernels and private optimizations for Meta Llama and other open-source LLMs1.
- Performance: Up to 1.5× faster than open-source vLLM-v1 on standard workloads.
- Deployment: Available in all regions, fully integrated with Databricks workspace and infrastructure, with autoscaling and cost efficiency1.
Why it matters:
High-volume AI applications (customer service bots, query copilots) need robust serving infrastructure. Databricks now supports deployment directly in your workspace at a real scale.
Practical tip:
For scalable AI apps, benchmark a live LLM endpoint against your expected user load. Use Databricks serving to test 10k–100k QPS, exposing end-to-end latency and autoscaling behavior.
Sources:
https://www.databricks.com/blog/mosaic-ai-announcements-data-ai-summit-2025
Practical First Steps to Try
- Load a lightweight table into Lakebase: query it live alongside Delta tables.
- Run an Agent Bricks pilot: choose a business task and generate your first agent.
- Build a scheduled pipeline using LakeFlow Designer: ingesting external data with governance.
- Track prompt versions with MLflow 3.0: deploy small-scale using serverless GPUs.
- Experiment with AI SQL and vector search: turn PDFs or support logs into searchable datasets.
- Benchmark LLM serving: simulate real traffic and measure performance.
Final Thoughts
Day 1 at the Summit reinforced Databricks’ commitment to unifying data processing, AI generation, and governance. The platform now supports real-time apps, AI-grounded pipelines, and scalable AI at production quality, all under one umbrella.
If you’re building modern data systems in Europe or beyond, these tools mark turning points, not experiments. They make the Lakehouse not just a storage lake for analytics, but a fully functional platform for real-time systems, intelligent agents, and governed AI.
To explore how these innovations align with your architecture, use cases, or roadmap, connect with tools, test small pilots, and see what real value emerges.
About DATAPAO
DATAPAO is a Data Engineering and Data Science consulting firm that supports the entire data journey, including strategy, implementation, training, and innovation.
As a trusted Databricks partner since 2016, DATAPAO is the one-stop solution to understand and leverage data better building on Databricks. DATAPAO tackles the most challenging data problems and helps organizations become truly data-driven.
, DATAPAO
