Javi Builds tool · Engineering planning estimate
Microsoft Fabric Capacity Planner
My Fabric capacity is running hot. Do I need a larger SKU, or should I optimize the workload? Enter what the Capacity Metrics app shows (or what you expect to run) and get an engineering read on health, headroom, throttling risk and the next SKU worth testing.
Engineering planning estimate, not an official Microsoft sizing recommendation. Validate recommendations with the Microsoft Fabric Capacity Metrics app and the official Microsoft Fabric SKU Estimator (Preview) before purchasing or resizing capacity.
The Javi Builds Capacity Planner interprets utilization you already have. The Microsoft SKU Estimator sizes a capacity from a workload description. Use both. Everything here runs in your browser; nothing you enter is stored or sent, except the optional public price lookup (region and currency only).
Throttling or rejections happening right now? Work through the Microsoft Fabric Capacity Saturation Runbook first.
Planner
Result
Results appear here.
Optional: estimated cost
Public retail prices from the Azure Retail Prices API, loaded only when you ask. Region and currency are the only values sent.
730 = always on. Paused hours are not billed for pay-as-you-go.
- Not a billing quote. Prices are Microsoft retail list prices; agreements, discounts and taxes are not included. Non-USD prices are for reference.
- OneLake storage, mirroring storage and some other meters are billed separately from capacity.
- Enabling capacity overage or autoscale-style billing can add cost: capacity overage is billed at a higher rate than the base CU price.
- A reservation bills at the reserved size; scaling below it does not reduce the bill.
What is a Capacity Unit?
A Fabric capacity is a budget of compute measured in Capacity Units (CUs), shared by every workspace assigned to it. An F64 has 64 CUs. Fabric accounts for consumption in 30-second timepoints, so the budget per timepoint is CU × 30 CU-seconds: an F64 provides 64 × 30 = 1,920 CU-seconds every 30 seconds. For Spark, one CU maps to two Spark VCores.
Why can Fabric exceed 100%?
Bursting. Operations run as fast as possible, even if that means temporarily using more compute than the SKU provides. The cost is then smoothed over future timepoints. That is why the Metrics app can show 120%, 180% or 250%: utilization above 100% means the capacity is using future capacity, not that it is throttling. Fabric also gives overage protection: up to 10 minutes of future capacity can be consumed with no throttling at all.
What is smoothing?
Smoothing spreads the CU cost of an operation over future timepoints. How far depends on the operation type:
- Interactive operations (report queries, most user-triggered work) are smoothed over a minimum of 5 minutes and up to 64 minutes, depending on their size.
- Background operations (refreshes, pipelines, Spark, almost all Warehouse work) are smoothed over 24 hours.
So one large background job barely moves any single timepoint, while a busy hour of report queries lands almost where it happened. When smoothed usage exceeds the SKU, the excess becomes carryforward; idle capacity in later timepoints pays it back (burndown). Smoothing changes the accounting, not the speed of the job.
When does Fabric throttle?
Throttling depends on how much future capacity smoothed usage has already consumed (Microsoft's current policy, which Microsoft says may change):
| Future capacity consumed | Stage | What happens |
|---|---|---|
| Up to 10 minutes | Overage protection | No throttling. |
| 10 to 60 minutes | Interactive delay | New interactive operations are delayed by 20 seconds at submission. |
| 60 minutes to 24 hours | Interactive rejection | New interactive operations are rejected; background jobs can still start. |
| More than 24 hours | Background rejection | All new requests are rejected until carryforward is paid off. |
Rejected requests return CapacityLimitExceeded. Operations already running when throttling starts are allowed to finish. Source: Understand capacity throttling and smoothing (Microsoft Learn).
How do I know whether I actually need a bigger capacity?
Scale on evidence, not on a utilization number. In the Capacity Metrics app:
- Throttling charts on the Compute page: does smoothed usage cross the 10-minute, 60-minute or 24-hour limit, and how often?
- Overages tab: is carryforward accumulating faster than it burns down?
- System events: when did the capacity enter and leave throttling states?
- CU timepoint analysis: drill through to the worst timepoint and rank interactive and background operations by CU.
- Workload and item drilldown: one refresh, notebook or query is often most of the peak.
Optimize before blindly scaling: full reloads, idle Spark sessions, colliding schedules and refreshes more frequent than the source changes often account for a whole SKU. Scale up when throttling is sustained and the workloads are already efficient; scale out (a second capacity) when one workload type keeps hurting another. A temporary scale-up also burns down carryforward faster. Microsoft documents these options in Evaluate and optimize your Microsoft Fabric capacity. For the deeper walkthrough, read Microsoft Fabric Capacity Explained and Fabric performance troubleshooting.
F SKU reference table
CU-seconds per timepoint = CU × 30. Spark VCores: 1 CU = 2 VCores; the burst column is Microsoft's documented maximum for concurrency.
| SKU | CUs | CU-seconds per 30-second timepoint | CU-hours per day | Spark VCores | Max Spark VCores with burst |
|---|---|---|---|---|---|
| F2 | 2 | 60 | 48 | 4 | 20 |
| F4 | 4 | 120 | 96 | 8 | 24 |
| F8 | 8 | 240 | 192 | 16 | 48 |
| F16 | 16 | 480 | 384 | 32 | 96 |
| F32 | 32 | 960 | 768 | 64 | 192 |
| F64 | 64 | 1,920 | 1,536 | 128 | 384 |
| F128 | 128 | 3,840 | 3,072 | 256 | 768 |
| F256 | 256 | 7,680 | 6,144 | 512 | 1,536 |
| F512 | 512 | 15,360 | 12,288 | 1,024 | 3,072 |
| F1024 | 1,024 | 30,720 | 24,576 | 2,048 | 6,144 |
| F2048 | 2,048 | 61,440 | 49,152 | 4,096 | 12,288 |
| F4096 | 4,096 | 122,880 | 98,304 | 8,192 | 24,576 |
| F8192 | 8,192 | 245,760 | 196,608 | 16,384 | 49,152 |
F64 and larger also let users with a free Fabric licence view Power BI content (viewer role); smaller F SKUs need Pro or PPU licences for viewers. Resizing across the F256 / F512 boundary can briefly interrupt the capacity.
How this planner works: Microsoft facts vs Javi Builds heuristics
From Microsoft Learn
- F SKU → CU table and CU × 30 per timepoint.
- 30-second timepoints; interactive smoothing 5–64 minutes, background 24 hours.
- Throttling stages at 10 minutes, 60 minutes and 24 hours of future capacity.
- Spark VCore and burst limits per SKU.
- F64 free-viewer threshold, the F256/F512 resize boundary, reservation billing.
Javi Builds heuristics (not Microsoft guidance)
- Health bands: Comfortable below 60%, Watch below 80%, Investigate below 100%, High risk above, adjusted by throttling evidence and recurrence.
- A 70% typical-utilization target when choosing the next SKU to test.
- Scale-down bands: comfortable below 50%, watch below 70%, little headroom up to 100%.
- The flat-peak overage estimate: (peak − 100%) × minutes above 100%.
- All Mode B coefficients (CU-hours per run, session size, refresh weight).
SKU comparisons are mathematical normalization: the same CU consumption divided by a different base capacity. Real workloads change with the SKU (concurrency limits, Spark pool sizes, run durations), so a normalized 80% on a larger SKU is a starting expectation to measure, not a prediction. Microsoft does not publish a numeric scale-down threshold; it recommends measuring with trial or pay-as-you-go capacity and the Metrics app, and starting small.
Microsoft facts last checked on against: licences and SKUs, plan capacity, throttling, scale capacity, Spark concurrency, optimize capacity, Azure Retail Prices API. Official sizing: What is the Fabric SKU Estimator (preview)?