Weekly ML Roundup: Fabric events, OneLake mirroring, PRISM2

by TechHub

This week in machine learning, Microsoft Fabric focused on making data and events easier to reuse across analytics and ML pipelines, from OneLake mirroring previews (Azure Monitor Log Analytics and AWS Glue-cataloged Iceberg) to more end-to-end streaming options with Change Event Streaming, Eventhouse, and clearer event architecture guidance. On the operations side, Fabric added practical safeguards and runbook improvements with Item Recovery becoming the default, a new VNet gateway evaluation engine preview, and scheduled User Data Functions for managed recurring logic. We also got concrete agent-building guidance with Microsoft IQ MCP endpoints across web, work, Fabric, and Foundry, and Microsoft Research released PRISM2 pathology foundation model weights on Hugging Face for research and benchmarking.

This Week's Overview

Microsoft Fabric doubles down on OneLake interoperability

Building on last week's OneLake mirroring momentum (including Snowflake Iceberg via metadata shortcuts), Fabric's data platform story got more concrete this week with new mirroring previews that bring external datasets into OneLake without copying data. The headline is a metadata-based, zero-copy approach for two very different sources: Azure Monitor Log Analytics tables and AWS Glue-cataloged Apache Iceberg tables. For teams trying to standardize analytics and ML feature pipelines across clouds, this reduces the usual ingestion work and keeps source-of-truth ownership where it already lives.

In practice, this preview is about faster onboarding and fewer moving parts in ELT: you can surface observability data (Azure Monitor) alongside lakehouse data (Iceberg) for unified querying and downstream modeling. The post calls out KQL (Kusto Query Language) as part of the experience for the Log Analytics side, so expect patterns where operational telemetry becomes first-class analytical data in Fabric without building a separate export pipeline.

Streaming and events in Fabric: from database change capture to event analytics

The Fabric updates this week align around event-driven data engineering: capturing changes, choosing an event backbone, and analyzing high-volume streams with less custom plumbing. If you're building near-real-time ML features, alerting, or operational dashboards, these pieces fit together as an end-to-end path from “something changed” to “we can query and act on it quickly.”

Change Event Streaming for SQL database in Fabric (Preview)

Following last week's broader CES preview across SQL Server 2025, Azure SQL, and Fabric SQL, Microsoft introduced a preview for Change Event Streaming for SQL database in Fabric that streams inserts, updates, and deletes into either Fabric Eventstream or Azure Event Hubs. That puts change-data-capture style pipelines closer to “flip it on and wire it up” instead of maintaining custom CDC connectors and retry logic.

For developers, the key decision point is the sink: Fabric Eventstream keeps you inside Fabric's real-time tooling, while Azure Event Hubs helps when you already have consumers outside Fabric. Either way, the emphasis is near-real-time propagation for downstream pipelines that need low-latency updates rather than batch loads.

Picking the right event pillar (and designing schemas that hold up)

This week's event architecture guidance extends the “streaming patterns for real-time ML signals” thread from last week by focusing less on the mechanics of enrichment and more on the upstream contract choices that determine whether those pipelines stay maintainable. Two guides focus on architecture choices and the practical work of making events usable by downstream teams. One breaks down when to use Business Events vs Fabric Events vs Azure Events, explicitly framing the decision around enterprise patterns that can span Fabric and Azure rather than forcing everything into a single product boundary.

The companion guidance goes deeper on Business Events schema design, with an emphasis on making events something “consumers can act on” (not just log lines with ambiguous fields). That matters for ML and analytics because stable, well-modeled event contracts reduce rework in feature engineering, reduce brittle parsing, and make it easier to version pipelines safely.

Eventhouse positioning for near-real-time analytics

After last week's examples of streaming + inline AI enrichment (Eventstreams, Structured Streaming, AI Functions), Eventhouse was highlighted as Fabric's “single engine” for ingesting, storing, and analyzing high-volume event streams in near real time. The pitch is less time spent tuning and stitching together separate streaming components, which is a common pain point when teams scale from prototypes to always-on ingestion.

For data engineers supporting ML, the practical takeaway is that Fabric is nudging workloads toward a more integrated stream-to-query loop. If you already standardize on Kusto-like patterns for event analytics, Eventhouse is positioned as the place to consolidate those use cases inside Fabric's broader platform.

Fabric operational improvements: recovery defaults, gateway evaluation, and scheduled functions

Several updates this week are less about new “features” and more about making Fabric safer to operate and easier to run at scale, continuing last week's theme of reducing the friction that shows up after the model and data flows work. Together they target three common production issues: accidental deletion, performance unpredictability through gateways, and brittle hand-rolled scheduling for recurring business logic.

Item Recovery becomes the default (starting Aug 16, 2026)

Microsoft Fabric will enable Item Recovery by default on August 16, 2026 for tenants that have not explicitly set the configuration. The default retention window is 3 days, and existing tenant configurations are preserved, so the behavioral change mainly affects orgs that never touched the setting.

Admins should plan for both the human workflow (restore experience via the workspace recycle bin) and the cost angle: OneLake storage billing can be impacted while deleted items remain in retention. This is one of those “quietly important” platform changes because it reduces the blast radius of mistakes, but you still want to verify governance settings and communicate expectations to workspace owners.

Modern evaluation engine for VNet data gateways (Preview)

A preview “modern evaluation engine” is now available (opt-in) for Microsoft Fabric VNet data gateways, providing an updated execution path for supported Mashup evaluations. The post is explicit that this is something to test and measure, not blindly roll out, and it suggests watching both performance and capacity metrics during evaluation.

If you run Dataflow Gen2, DirectQuery, or semantic model scenarios through a VNet gateway, this preview is aimed at improving predictability and throughput where the gateway evaluation path becomes a bottleneck. The recommended rollout approach (test, measure, then expand) is worth following since gateway changes can have non-obvious downstream effects on refresh windows and capacity utilization.

Scheduled User Data Functions (Preview)

Scheduled User Data Functions build on last week's push toward more operational, governed Fabric workflows by moving recurring enrichment and feature logic into the same managed monitoring and scheduling surface as other platform jobs. User Data Functions in Fabric now support scheduling in preview, so recurring business logic can run through the Fabric job scheduler. Runs show up in the Fabric Monitoring Hub, which helps teams treat these functions as managed jobs rather than scripts someone runs manually.

This matters when teams start encoding feature calculations, data quality checks, or periodic enrichment into User Data Functions. Central scheduling plus monitoring reduces “tribal knowledge” operations and makes it easier to troubleshoot failures with a shared view of job history.

Building agents with MCP “IQ” endpoints across web, work, Fabric, and Foundry

Following last week's Microsoft IQ overview as a governance-first way to feed agents the right context, Microsoft published a practical set of recordings and resources showing how to use four “IQs” (Web IQ, Work IQ, Fabric IQ, Foundry IQ) and integrate them via Model Context Protocol (MCP) endpoints. The guidance is aimed at Python developers building agents with Microsoft Agent Framework, then deploying them to Foundry Agent Service.

The architectural through-line is multi-source retrieval for agentic workflows: web content, organizational work context, OneLake/Fabric data, and Foundry resources can be queried through a consistent MCP surface. For ML engineers, the useful angle is how this can standardize tool calling and retrieval across environments, especially when an agent needs to ground responses in enterprise data while still pulling public references.

Multimodal pathology foundation models: PRISM2 weights released for research

Microsoft Research and Paige (Tempus) described PRISM2, a multimodal pathology foundation model trained on tissue images and language derived from pathology reports. The goal is to generalize across multiple cancer-related benchmark tasks without training separate models per task, which reflects the broader push toward reusable medical foundation models rather than narrowly tuned single-purpose classifiers.

For researchers and applied teams, the practical news is access: the model weights are released on Hugging Face for research use. That enables reproducibility, benchmarking against local datasets, and experimentation with fine-tuning or retrieval-augmented workflows where image features and report text need to be interpreted together.

Other Machine Learning News

John Savill's Azure update roundup touched several ML-adjacent platform changes worth tracking if you operate data/AI stacks on Azure, which complements last week's emphasis on scaling and operating real workloads (compute portability, governed context) beyond one-off demos. The list includes Azure Databricks improvements (including Unity AI Gateway mentions), plus infrastructure and data-layer updates (like Azure Private Link over IPv6 and Azure Firewall explicit proxy) that often show up indirectly as constraints or enablers for secured ML pipelines.

A separate video focused on the “real pricing vs estimates” idea as an architecture driver, framing Microsoft Fabric as the foundation for unifying data and applying AI. It is more strategic than hands-on, but it reflects the ongoing push to treat data + context as the input to analytics and ML decisioning.