Microsoft Fabric

Episode 323 – Microsoft Fabric March 2026 Feature Summary (Part 1)

With the FabCon Atlanta keynote announcements already covered in Episode 322, this one picks up where that left off and gets into the platform, OneLake, and data engineering sections of the March summary. Power BI, data science, and the rest of Data Factory will land in a future episode. For now, buckle up — the pair keep things moving, only pausing to circle back on the parts that actually change how the day-to-day work gets done.

Fabric Platform: Governance, CI/CD, and a pile of MCP

The platform section opens with a run of GA announcements. Govern in the OneLake Catalog for admins is now generally available — Jason’s take is that it’s cleaner than what was there before but not deep enough to dwell on. Same energy for workspace tags going GA, which caps out at ten tags per workspace. “Why just ten?” — a fair question without a clear answer. Jason wonders whether the tagging metadata is quietly laying groundwork for Fabric’s ontologies story, which is the kind of read-between-the-lines observation that makes these episodes worth the price of admission.

The meatier news: the OneLake Catalog Search API is now available as an MCP tool in preview, and it’s baked into the Fabric Core MCP server. As John points out, this is the pattern going forward — “where we haven’t had APIs before, we have APIs, because you need the APIs to be able to get the AI on top of it.” Cross-workspace discovery finally shows up at the code level, which is a big unlock for anyone building agents against Fabric.

DLP, IRM, and the purview of Purview

Data Loss Prevention policies now extend down into structured data in OneLake — warehouses, KQL databases, SQL databases, the whole neighborhood. Jason likes that the DLP capability itself lives at the Fabric level without requiring Purview, which matters for shops that haven’t gone all-in on Purview yet. The adjacent Purview Insider Risk Management (IRM) story is a separate layer, and Jason drops a war story from a client the day before: his dev team put IRM on a UAT validation doc that prevented the UAT team from copying snippets into their team sources. A good reminder that IRM and DLP solve different problems, and throwing IRM at everything isn’t the move.

Worth noting: “If you’ve ever had anybody at Microsoft send you a document, you’ve probably bumped into the fact that you couldn’t read it because of IRM, because they always forget to turn that off. Always.” Quick policy for data theft is now GA, and IRM pay-as-you-go usage reporting is available for teams who want to dip a toe in without committing to a full license. Fabric Copilot security posture management rounds out the Purview-adjacent updates.

Branched workspaces, selective branching, and the connection reference variable

Now we’re in their wheelhouse. Branched workspace with Git integration is in preview, and it brings with it selective branching — the ability to push just the items you need from a source workspace to a target, rather than cloning the whole thing. Jason walked his team through it the day before recording and had good things to say. Worth noting that these rollouts are often staggered by data center and tenant — list mirroring in particular still hadn’t shown up in some tenants at recording time, a feature many have been eager to put to work since FabCon.

The connection reference item type in variable library is the one Jason is most fired up about. Rather than embedding static connection strings in code, the new variable type lets you reference existing connections like Azure SQL or Snowflake by storing a connection ID. “That’s huge,” Jason says. “Being able to switch between my dev, my UAT and my production Azure SQL databases, that’s an important one.” Anyone who’s ever had to untangle hard-coded connection strings across environments knows exactly why.

Bulk export and import APIs for item definitions quietly showed up this month too. Up until now, the way to get a snapshot of item definitions was to use Git integration and read them out of the repo. The new APIs open up scenarios like offline backup and programmatic recovery, which John’s team has been chasing for a while.

Fabric CLI v1.5 and the MCP parade

Fabric CLI v1.5 shipped with CI/CD deployments, Power BI scenario support, and DX improvements. The headliner is one-command workspace deployments — hand the CLI a workspace definition, and it deploys. John’s read on why the CLI is moving so fast: “AI likes CLIs, right? Your AI agents love to work with CLIs. Ideally they talk to the CLI, they don’t talk directly to an API.” Jason sums it up with an acronym stew: “Your CLI talks to your API in behind and you have a UX for your MCP server that all talks to an AI. How many acronyms can we string together into a soup here?”

Speaking of MCP, the Fabric Remote MCP server is now available. If you’ve been running the local Fabric MCP up to now, the remote version lets you point your AI tool of choice at it without wiring anything up in Visual Studio Code. Fabric MCP AI code assistance went GA alongside it — though the code assistance side still requires the local MCP.

Fabric extensibility reached GA this month as well, CI/CD and remote support for custom workloads is in preview, and Variable Library support landed for workloads. There’s also a new Fabric Scheduler for remote jobs — John’s read is that this pertains to custom workloads and gives them access to the Fabric scheduler for remote execution. As Jason notes, “our good friend Mike Carlo will be all over this stuff” — a not-so-subtle suggestion to bring Mike back on the show to unpack custom workloads in depth.

OneLake: Third-party security and File Explorer GA

Third-party support for OneLake security is a real unlock — it means OneLake security is no longer just for Fabric engines. Third-party tools can now read OneLake data with the security posture respected, which is exactly what customers with mixed ecosystems have been asking for.

The OneLake File Explorer is now generally available at v1.1. It’s stable, supported, and lives directly in Windows File Explorer. A few rough edges remain — it’s still Windows-only, it prompts at startup (OneDrive-style silent background behavior would be the ideal), and it still doesn’t support multiple tenants simultaneously. The word from the OneLake team is that multi-tenant is seen as a niche ask — though anyone working across customer or dev tenants regularly would beg to differ. Microsoft clearly pulled from the OneDrive team’s playbook — both hosts agree that’s a smart move, even if they spent years bashing on OneDrive before it got good.

Data Engineering: Runtime 2.0, custom live pools, and a whole notebook refresh

Fabric Runtime 2.0 (Preview)

Fabric Runtime 2.0 is in preview, and it’s a big jump. Java 21, Scala 2.13, Python 3.12, Spark 4.0, and the headline — Delta Lake 4.0. Runtime 1.3 topped out at Spark 3.5 and Delta 3.2, so this is a material step forward. You can swap the runtime on a workspace-by-workspace basis in workspace settings. John’s betting it’s “faster, more efficient” and both hosts agree that Mr. Finland (Chris Finlan) would have excellent insight on what’s under the hood here — which turns into a note to get him back on the show.

Custom live pools and workspace monitoring

Custom live pools for Fabric Data Engineering address the long startup times that hit teams running custom pools rather than the default starter pool. Instead of a multi-minute cold start, custom live pools stay warmed up and ready, with a roughly three-to-five-second handoff. They address the noisy neighbor problem on starter pools and give teams predictable, isolated compute — at a cost. Jason is honest that he’s still on starter pools for most of what he does because he hasn’t needed the investment yet.

Workspace monitoring picked up new UI for watching what your live pools are doing. The sharp editorial on the whole category: it’s great that Fabric can be used to monitor Fabric, but it probably shouldn’t be. The reasoning lands: if Fabric is having a serious problem, the monitoring built on Fabric is having the same problem. “It’s like sending out an email to tell everyone that the email server is down.” The wish list: a dedicated monitoring capacity, or better yet, a monitoring approach that doesn’t burn CUs to tell you how many CUs you’re burning. Notably, Azure monitoring is now part of the Fabric organization inside Microsoft — so the hosts expect real movement here down the road.

Notebook updates, end to end

There’s a lot in the notebook section, so here’s the quick tour. Quick mode for library installation is in preview, giving teams a fast on-demand install path alongside the full snapshot-based mode for production. High concurrency mode now supports up to 50 notebooks in a shared Spark session (up from ten), which matters because the number of Spark sessions you can spin up is capped. Data export settings give admins tenant-level control over data leaving a notebook — download buttons go dark, grid exports are blocked, and transformed data stays inside Fabric. Session start insights expose how long sessions are taking to warm up, so performance drift becomes visible.

Z-order and liquid clustering support in the Native Execution Engine is the one the hosts punt on. Jason plans to go do some reading: “I usually just refer to it as a flight of beers. That’s liquid clustering in my world.” John, for his part, knows V-ordering cold (it’s the columnar compression that powers Power BI’s performance story) but is honest that liquid clustering isn’t in his wheelhouse yet. Expect a follow-up on a future episode.

Copilot for data engineering and data science in notebooks picks up context awareness, including a Fix with Copilot button inside the notebook. Outside the service, there’s a new Microsoft-published Fabric custom agent for VS Code — notebooks and all — which Jason appreciates mostly because “otherwise you’re looking at custom ones from Bob Smith and Sally Ride and you never know if they’re real people or if they died in a shuttle explosion.” Local notebook execution against a Fabric kernel in VS Code is also here, and the extension now supports tenant switching. “Tenant switching should be a standard capability across all tools,” Jason says — and as someone who hops between tenants constantly, he’s not wrong.

Materialized Lake Views now support multiple schedules in a single lakehouse, and PySpark authoring is available in preview. Both are welcome: MLVs went GA this month with real flexibility for teams running medallion patterns. Copy jobs got easier access from the Lakehouse’s “Get Data” experience. Notebook default lakehouse binding now travels through Git, as does the notebook resource folder, and you can create files in the resource folder directly from the UI. Fabric notebook public APIs reached GA with full CRUD support, and Copilot notebook completion got the usual bigger-better-faster refresh.

John wraps the theme neatly: “Are we detecting a theme? APIs, AI, APIs, AI.” With that, Jason calls time — data science, AI, and Data Factory are coming up in Part 2 — and heads off to go figure out what liquid clustering actually is.

Links

Microsoft Blog Posts

Previous Episodes

Subscribe: SoundCloud | iTunes | Spotify | TuneIn | Amazon Music

Tags: , , , , , , , , , , , , , , , , , , , , , ,

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

AvePoint