Databricks ADBC Connector Reaches GA in Power BI and Fabric

Microsoft has officially pushed the Databricks ADBC connector to general availability for Power BI and Microsoft Fabric. It marks the first major production milestone in a long-term transition that will reshape how Power Query fetches tabular records from external query engines. If your enterprise estate pairs Dataverse transactions with a Databricks lakehouse, the data pipes feeding your executive dashboards are about to run on completely different underlying transport code.

ODBC has carried the burden of external data connectivity in Power BI for well over a decade. It works, but it shows its age on modern analytical workloads. Row-by-row conversions, high memory footprints, and finicky unmanaged memory allocations have caused headaches for teams refreshing massive datasets. The new connector replaces that old bridge with Arrow Database Connectivity (ADBC), using Apache Arrow columnar formats directly in memory. You do not have to move immediately, but ignoring the change is a mistake.

What actually changed under the hood

The headline from the official Databricks ADBC connector announcement is general availability across semantic models, dataflows, and paginated reports. Behind that simple status change sits an overhaul of the Power Query runtime architecture. For years, Power BI relied on embedded, pre-packaged ODBC drivers shipped directly with Desktop and the cloud service. Those embedded drivers are being retired in stages across the platform.

ADBC establishes an open standard interface specifically designed for columnar data formats. When Power BI talks to Databricks over ODBC, the database engine executes a query, turns columnar chunks into row-oriented formats, sends them across the wire, and Power Query translates them back into columnar tables for the VertiPaq engine. That double serialization is sheer wasted compute. ADBC eliminates that round-trip by keeping Arrow buffers intact from the remote cluster all the way into the Power Query ingestion buffers.

Microsoft has kept safety top of mind for this rollout. Existing connections will not break tomorrow. Semantic models, scheduled dataflows, and paginated reports continue to run on the legacy path without intervention. The migration to ADBC is strictly opt-in, and administrators can enable it granularly per connection, per workspace, or across the whole tenant.

There is also an important boundary to understand regarding custom drivers. This change specifically targets the built-in, embedded ODBC drivers provided by Microsoft. If your solution relies on a separately installed ODBC driver, such as a specialized driver installed on an on-premises data gateway VM with custom DSN parameters, your pipeline remains completely untouched.

Why the shift from ODBC to ADBC matters

The practical difference between ODBC and ADBC comes down to two operational factors: memory safety and large dataset throughput. On enterprise reporting projects, scheduled refreshes do not usually fail because the DAX logic is broken. They fail because the refresh process runs out of memory on the mashup container or hits a timeout while pulling forty million rows out of a delta table.

ADBC drivers introduce modern memory safety and clean garbage collection directly into the driver layer. Classic C-based ODBC drivers have spent years accumulating legacy memory management quirks. When multiple heavy refreshes run concurrently on a Fabric capacity, unmanaged driver memory can fragment, causing container restarts that leave consultants guessing. Moving to memory-safe implementations reduces those mysterious, non-reproducible mashup crashes.

Throughput is the other major win. Because Apache Arrow organizes data in memory by columns rather than rows, ADBC fetches analytical payloads dramatically faster. If your team extracts transformed historical snapshots from Databricks to supplement live operational tables from Dynamics 365, the network transfer and decompression overhead drops substantially. In practical terms, that means shorter refresh windows and less strain on your capacity compute units.

The timing fits cleanly into broader changes landing across the analytics stack. As highlighted in the Power BI October 2026 feature summary, Microsoft is tightening governance across semantic models, including more accurate surge protection controls. When your data connectors pull records more efficiently, your models spend less time hovering near resource limits that trigger protective throttling.

What this means across project roles

A change in connector architecture affects different layers of an organization in different ways. You should divide responsibilities clearly rather than leaving migration decisions to individual authors.

For solution architects and lead consultants

Stop designing new ingestion pipelines around custom ODBC definitions where a native ADBC connector exists. If you are architecting a lakehouse integration that pulls transformed telemetry or billing data into Fabric to merge with Dataverse core entities, design for ADBC out of the gate. You should also update your environment runbooks. Do not approve tenant-wide switches until your team has verified workspace-level pilots. The architectural gain is clear, but unintended query fold breaks can derail production reporting.

For report makers and semantic model authors

Your day-to-day modeling workflow does not change drastically. You will still see the Databricks connector in Power Query. However, when you opt a model into ADBC, pay attention to data type translations. Columnar conversions occasionally handle dates, time zones, and high-precision decimals differently than legacy ODBC layers. If you have custom M code that relied on specific driver-level type coercions, test those queries thoroughly.

For Fabric and Power Platform tenant admins

You control the throttle. Because Microsoft provides opt-in switches at the connection, workspace, and tenant levels, you do not have to endure an all-or-nothing cutover. Your primary task this month is communication. Identify which workspaces maintain active connections to Databricks clusters, notify the respective BI product owners, and establish a sandbox workspace where model authors can test the ADBC switch without threatening production service level agreements.

What I would check first: A five-step migration test plan

Do not flip the tenant-level switch and hope for the best. Work through a controlled verification cycle instead.

1. Inventory existing Databricks connections. Query your tenant activity logs or review gateway connection inventories to identify every semantic model and dataflow currently pointing to a Databricks endpoint.
2. Set up an isolated test workspace. Clone two or three of your heaviest production semantic models into a dedicated non-production workspace assigned to identical capacity sizing.
3. Enable ADBC at the workspace level. Turn on the ADBC connector switch within the test workspace settings, leaving your production workspaces on the legacy driver.
4. Run side-by-side refresh benchmarks. Trigger simultaneous refreshes of the cloned model and the production model. Compare the total duration, peak memory consumption, and mashup engine container logs. Look closely for any variations in row counts or date-time values.
5. Validate DirectQuery interactions. If your models use DirectQuery instead of full import, click through complex report pages. Check that visual load times remain steady and that generated SQL passes cleanly to your Databricks SQL warehouses without syntax rejections.

The caveats and what Microsoft has not said yet

While the Databricks ADBC connector is generally available, the broader transition has obvious gaps that require caution. First, running an opt-in model means your organization will operate in a split-brain state for months. Some models will run on ADBC while older models linger on embedded ODBC. Troubleshooting gateway performance or cluster query histories becomes harder when two different client drivers submit queries to the same Databricks endpoint.

Second, Microsoft has not published a hard deprecation date for the embedded ODBC drivers. We know the transition is underway, but the documentation does not state when legacy embedded drivers will be turned off completely. That ambiguity makes it difficult for project managers to prioritize migration tasks against urgent business feature requests.

Finally, the vendor scope remains narrow. Databricks is the flagship implementation for this ADBC transition. Microsoft has stated that other Power Query connectors will follow the same path, but detailed timelines for other major cloud data warehouses have not been disclosed. If your reporting tier pulls from multiple lakehouse platforms simultaneously, you will be managing a hybrid driver setup for the foreseeable future.

Where to go from here

The general availability of the Databricks ADBC connector is a welcome technical progression. Moving away from thirty-year-old ODBC conventions toward memory-safe, Arrow-native columnar data transfer is the right architectural choice for Power BI and Microsoft Fabric. It reduces container memory spikes, speeds up high-volume refreshes, and stabilizes large enterprise models.

Take advantage of the opt-in design. Do not rush to activate the tenant-level toggle this week. Instead, pick your most troublesome Databricks-backed semantic model, copy it into a sandbox workspace, toggle the ADBC setting, and measure the difference in refresh performance. Real benchmark data from your own environment will tell you everything you need to know before rolling the change across your enterprise.

Sources and further reading

• Databricks ADBC connector for Power BI and Microsoft Fabric (Generally Available) — Microsoft Power BI Blog
• Power BI October 2026 Feature Summary — Microsoft Power BI Blog
• Power Platform licensing FAQs - Power Platform — MS Learn - Dataverse docs
• About the Microsoft Dataverse for Teams environment - Power Platform — MS Learn - Dataverse docs

Written for the MaloomatAI editorial desk and checked against Microsoft's own announcements on 17 October 2026. Product names and trademarks belong to their respective owners.

Spotted something that has changed since publication? Leave a comment and it will be corrected.