The Native Planning Engine stores planning data and application metadata as files in OneLake. That one architectural shift, together with optimizations across the product, makes Fabric Plan faster from the first click to the final writeback.
This is Part 3 of our three-part Native Planning Engine series, covering the Fabric Plan performance improvements beyond calculation and writeback. Part 1 introduces the Native Planning Engine, and Part 2 covers writeback to OneLake.
Covered in Parts 1 and 2: 10× more scale (12M cells), 10× faster calculation and up to 13× faster writeback.

Figures come from our testing. Performance varies with workload, model design, data size and capacity.
One Architecture Change, Felt Across the Product
With the Native Planning Engine, planning data and application metadata are stored as files in OneLake, and Plan items no longer need a separate system database. The benefits go well beyond the calculation and scale gains in Part 1 and the writeback gains in Part 2: creating and opening Plan items, adding Planning sheets and returning to views you use often are all faster too.
We also optimized start-up, the catalog, filters and Intelligence Sheets. Some of these improvements reach every Plan item; the full set comes with the Native Planning Engine and OneLake.
In this article:
Fabric Plan Performance Improvements, Before and After
The same everyday tasks, from launch to writeback, before and after the Native Planning Engine and OneLake.
| What you do | Before | After | Change |
|---|---|---|---|
| Launch Fabric Plan | A large application to download first | A compact package that loads quickly, even on slower networks | Less than ⅓ the size |
| Create a Plan item | Set-up took about twice as long | Ready for planning in about half the time | Up to 53% faster |
| Open a Plan item | Slower to open | Opens up to a third sooner | Up to 33% faster |
| Add a Planning sheet | Adding a sheet took up to five times as long | New sheets are ready in a fraction of the time | Up to 79% faster |
| Browse the catalog | The full catalog loaded up front | Each section loads as you open it | Up to 60% faster |
| Open a filter | Long member lists took time to appear | Lists appear with only the members you need | About 50% faster |
| Plan at scale | Up to 1.2 million cells in aggregate | Up to 12 million cells in aggregate (Part 1) | 10× more scale |
| Calculate the plan | Calculations on the previous planning engine | In-memory calculation on the Native Planning Engine (Part 1) | 10× faster |
| Load an Intelligence Sheet | Visuals competed for shared resources while loading | Visuals load side by side without waiting on each other | Up to 80% faster |
| Return to a view you use often | Data was fetched again each time | A fast local copy is used first | Faster repeat visits |
| Write back results | SQL-based writes | File-based Delta Parquet writes directly to OneLake (Part 2) | Up to 13× faster |
What it means for your team: less waiting during planning sessions, more scenarios tested before a deadline, and a tool that more people keep using.
Four Ways to Get the Most from These Improvements
Some of the result depends on how your environment is set up. These four steps make the biggest difference.
1. Migrate to OneLake
The file-based architecture is what unlocks the full set of improvements in this blog. Plan items still on the previous planning engine miss many of them.
Turn on Store new Plan items in OneLake at the tenant and workspace levels, start with new Plan items, then migrate your existing ones gradually. Part 1 walks through each step.
2. Write Back to OneLake
Saving planning results to OneLake makes everything above even more effective. Your plans, forecasts and assumptions sit in one governed place, and reporting across Fabric reads them directly.
File-based Delta Parquet writes to OneLake are about 5–6× faster than SQL-based writes at typical volumes and up to 13× faster at 12M cells. Fabric SQL remains a fully supported writeback destination. Part 2 shows how to set it up.
3. Keep Your Semantic Model in Good Shape
Fabric Plan reads data from your semantic model, and how long that takes depends on how well the model performs.
Outside Fabric Plan’s scope. The improvements in this blog cover what happens inside Fabric Plan. They do not change how quickly your semantic model loads. If you want data to arrive faster, work with your semantic model owner to keep the model well designed and efficient.
4. Bring In Only the Data You Need
The more detail a planning sheet holds, the more work it has to do. Detail that nobody plans against makes the sheet slower without adding value.
- Plan at the level your decisions need. If a team plans by region and month, there is no need to load every product and day.
- Use filters and RLS to focus each user on the part of the plan they own. With these set up properly, most teams won’t need more than 2.4 million cells in a planning grid.
- If your use cases still need more than 12 million cells, consider multidimensional cube modeling, or dimension partitioning with consolidation through InfoBridge or PowerTable–Planning integration (see Part 1).
For scaling techniques such as context filtering, dimension partitioning, cubes and PowerTable, see Key Considerations for Planning at Scale with Microsoft Fabric Plan.
Complete Your OneLake Migration
Move your Plan items to the Native Planning Engine, write back to OneLake, and keep your data model and sheets lean. You get faster planning everywhere, from the first click to the final writeback.
Start with Part 1: turn on the Native Planning Engine, then Part 2: write back to OneLake.
Learn more in the Fabric Plan documentation and on Microsoft Learn.
Appendix: How Each Improvement Works
For readers who want to know what changed behind the scenes at each step.
- Create and open. With application metadata stored as files in OneLake, a new Plan item can be ready in about half the time, opening one takes up to a third less, and a new Planning sheet takes a fraction of what it did.
- Start-up. The application is compressed to less than a third of its earlier size, so there is far less to download before you can begin.
- Catalog. Instead of loading the full catalog up front, each section loads on its own when you open it. Measures from your data model load separately too.
- Filters. Filter lists now fetch only what the filter panel needs to show, so nothing extra is prepared in the background.
- Large data volumes. Big requests are divided into smaller parts that run side by side, and deciding how to divide them now takes far less effort.
- Scale. The Native Planning Engine supports up to 12 million cells in aggregate in the planning grid, up from 1.2 million, with up to 120 million cells on the roadmap. See Part 1.
- Calculations. Totals are worked out across many values at once, linked formulas run in a smarter order, and plans with template rows calculate faster.
- Repeat access. Data you use often is kept on a fast local copy first. If it isn’t there, Fabric Plan checks OneLake, and only then goes back to the source.
- Intelligence Sheets. Visuals on the same sheet now share connections and resources, so they stop competing with each other while they load.
- Embedded planning. Embedded planning runs on the same improved loading, catalog, filter and data handling, so users in your portals get the gains too.
- Writeback. File-based Delta Parquet writes land results in OneLake, where your reports and other Fabric experiences already read data, with no extra copying step. See the writeback benchmarks in Part 2.
Improvements are described as “up to” or approximate figures, and results vary with workload, model design, data size and capacity.

Comments
0Please log in to comment.