A BI developer at a manufacturing client rebuilt their Dynamics 365 Finance and Operations reporting as a Direct Lake semantic model the week Fast Fabric rolled out, expecting the same near-instant refresh everyone had been promised in the release notes. For the first ten days it worked exactly that way: framing updates that used to take an overnight BYOD job finished in seconds, and finance users watched inventory valuation numbers move in something close to real time. Then a month-end load pushed one fact table past a few hundred million rows, a refresh failed outright with no warning beyond a guardrail error code, and the model sat unqueryable until someone tracked down which table had tripped which limit. Nothing about Direct Lake had changed. What changed was that nobody had modeled for its ceilings before the data volume found them.
That gap between “Direct Lake is fast” and “Direct Lake behaves differently enough from Import mode that it needs its own design discipline” is where most F&O teams get tripped up right now. Fast Fabric has genuinely solved the latency problem that made near-real-time F&O reporting impractical for years, cutting sync times that used to run one to two hours down to under fifteen minutes in most observed cases, according to Microsoft’s own Dataverse and Fabric integration update. But building a semantic model on top of that faster pipe is a different exercise than plugging Power BI into a data warehouse, and the differences matter most exactly when a model is otherwise working well enough that nobody is looking closely at it.
What actually lands in OneLake now
F&O data doesn’t write to OneLake directly. It flows through Dataverse, which is the layer that transforms F&O table data into Delta Parquet and exposes it to Fabric through shortcuts and mirroring, without the customer managing any Spark compute or Azure infrastructure themselves. That’s the architectural distinction from the older Synapse Link pattern, where a customer’s own Spark jobs did the transformation. Fast Fabric, now in public preview with global rollout completed and an automatic upgrade for existing customers planned for the first quarter of 2026, is what tightened the sync cadence on that same Dataverse-to-OneLake pipeline. Anyone still comparing today’s F&O-to-Fabric numbers against the old BYOD or Data Export Service benchmarks is working from a stale mental model of what the pipeline can do.
The practical consequence for a solution architect is that F&O tables show up in a Fabric lakehouse as Delta tables with a Dataverse-flavored schema underneath them: GUID-based keys, some binary columns, and complex types that don’t map cleanly onto every Direct Lake data type. That schema inheritance is the first thing worth checking before any modeling work starts, not after a refresh fails.
Choosing a Direct Lake semantic model: OneLake or the SQL analytics endpoint
Fabric currently gives two ways to put a semantic model directly on top of those Delta tables, and the choice isn’t cosmetic. Direct Lake on the SQL analytics endpoint reads through the lakehouse or warehouse’s SQL layer, supports non-materialized views, and, critically, falls back to DirectQuery automatically when a table can’t load in Direct Lake mode, for example when SQL-based row-level security is enforced. Direct Lake on OneLake skips that SQL layer and reads Delta tables directly from any Fabric source, which unlocks composite models that mix Direct Lake with Import or DirectQuery tables, calculated tables, and calculated columns (both still in preview and, for calculated columns, limited to user-context evaluation). The tradeoff is that Direct Lake on OneLake has no fallback path at all. If a query exceeds a guardrail, it fails. There is no quietly-slower DirectQuery safety net to catch it.
For most F&O reporting models, that argues for a specific default: use Direct Lake on OneLake when the model genuinely needs composite modeling or calculated tables over F&O data, and reach for Direct Lake on the SQL analytics endpoint when the model is closer to a straightforward reporting layer where a DirectQuery fallback is worth more than the extra flexibility. Architects who pick OneLake mode by default because it’s the newer, more capable option often find out about the missing fallback during a production incident rather than during design review.
Framing, transcoding, and what a guardrail failure actually costs
Direct Lake’s refresh model is fundamentally different from Import, and it’s worth explaining precisely because the difference is what makes the failure mode above so disruptive. A traditional refresh in Import mode copies the underlying data. Direct Lake framing copies only metadata: it analyzes the latest Delta table version, points the model’s references at the current files in OneLake, and finishes in seconds rather than minutes or hours. Data itself loads column by column, and only the columns a given query actually touches, through a process Microsoft calls transcoding. That’s what lets a Direct Lake model handle data volumes larger than a capacity’s memory ceiling; it never has to hold all of it at once.
Capacity guardrails are what set the boundaries on this, and they scale meaningfully by Fabric SKU. An F2 through F8 capacity caps out at 1,000 Parquet files and 1,000 row groups per table, 300 million rows, and a 10 GB model size limit. Step up to F64 or above and those ceilings jump to 5,000 files, 5,000 row groups, and 1.5 billion rows, with model size effectively unlimited. A large F&O fact table, general ledger transactions or inventory movements on an enterprise instance, can realistically approach those lower-tier limits, which means the F-SKU a Fabric workspace runs on isn’t just a cost decision anymore. It’s a modeling constraint that should be checked against actual F&O table row counts before a Direct Lake model design gets finalized, not discovered when a guardrail breach takes the model offline.
The newer Manage Tables capability, now in public preview for existing Fabric links, gives architects a lever here that didn’t exist a year ago: selecting which F&O-sourced tables actually sync to Fabric, rather than replicating an entire dataset and hoping guardrails hold. Narrowing the synced table set to what a given semantic model actually needs is one of the more direct ways to keep a Direct Lake model inside its capacity’s guardrails without upgrading the SKU.
Data type conversions worth doing before, not after, deployment
Direct Lake doesn’t support every Delta data type. Complex Delta types, binary fields, and GUID columns need to be converted, typically to strings, before a Direct Lake model can use them cleanly. Because F&O data reaches OneLake through Dataverse’s schema, GUID-based surrogate keys and occasional binary fields show up more often than they would in a purpose-built warehouse schema, so this conversion step tends to hit F&O-sourced models harder than it hits models built from cleaner, purpose-designed lakehouse tables. Handling these conversions in a lakehouse notebook or a Fabric pipeline before the semantic model is built, rather than discovering the unsupported type at model creation time, avoids a rebuild cycle that’s easy to skip when a project is already behind schedule.
A few other constraints round out the pre-deployment checklist: Direct Lake has no gateway support, so on-premises data can’t participate directly; the semantic model’s workspace has to sit in the same region as the data source workspace; and Direct Lake isn’t available in personal workspaces at all. None of these are exotic edge cases for an enterprise F&O deployment, but each is the kind of thing that’s far cheaper to confirm during architecture review than to discover during a go-live weekend.
Building for the ceiling, not just the average day
Fast Fabric removed the latency excuse that used to justify slower, batch-oriented F&O reporting architectures, and that’s a real and welcome change. But a semantic model designed only against yesterday’s data volumes, without a look at capacity guardrails, fallback behavior, and the schema quirks Dataverse introduces, is a model that works fine until the month it doesn’t. Routeget Technologies has walked several clients through exactly this kind of pre-deployment guardrail check on F&O Direct Lake models, and the pattern holds: the projects that skip it don’t usually fail on day one. They fail on the first month-end close that actually stresses the numbers the design assumed would stay small.
#DirectLake #MicrosoftFabric #DynamicsFinanceOps #PowerBI #DataModeling #FabricGuardrails
No comment yet, add your voice below!