Fabric Capacity Throttling: How to Read the Capacity Metrics App
Smoothing, overage protection, carryforward and the three throttling stages — and what to look at before you buy a bigger SKU.
Core Engineering Takeaways
- •Fabric smooths interactive operations over 5–64 minutes and background operations over 24 hours.
- •Throttling starts when future smoothed usage exceeds 10 minutes of capacity: first 20-second interactive delays, then interactive rejection after 60 minutes, then rejection of everything after 24 hours.
- •Overages above the 10-minute window become carryforward that idle capacity has to burn down before throttling stops.
- •Most throttling is caused by background load (refreshes) or a few expensive queries — not by too many users.
1. Bursting and smoothing: why the numbers look strange
A Fabric capacity has a fixed number of capacity units (CUs), but individual operations are allowed to burst — use more compute than the capacity's size for a short time — so they finish quickly. To avoid penalising you for bursting, Fabric smooths the CU usage of each operation over a longer window:
- Interactive operations (report queries) are smoothed over a minimum of five minutes and up to 64 minutes, depending on how much they consume.
- Background operations (refreshes, scheduled jobs) are smoothed over 24 hours.
That second point explains a lot of confusing charts. A heavy 06:00 refresh does not show up as a spike at 06:00; it adds a slice of load to every timepoint for the next 24 hours. A capacity with lots of background work can therefore sit at high utilisation all day with nobody using reports.
2. The throttling stages
Throttling is based on how much of the capacity's future has already been used up by smoothed operations:
- Up to 10 minutes of future usage — overage protection. No throttling. Short spikes are absorbed.
- 10 to 60 minutes — interactive delay. New interactive operations are delayed by 20 seconds at submission. Users notice reports getting sluggish.
- 60 minutes to 24 hours — interactive rejection. New interactive operations are rejected; background operations can still start and run. Users see errors.
- More than 24 hours — background rejection. All new requests, interactive and background, are rejected. Refreshes fail too.
Overages beyond the 10-minute window become carryforward CUs. Each later timepoint applies the carryforward, and any unused capacity reduces it — that reduction is called burndown. Throttling continues until idle capacity has paid off all the carryforward.
3. A worked example: why a 06:00 refresh slows reports at 15:00
Capacity is easiest to reason about in CU-seconds. An F64 provides 64 CUs, so it delivers 64 CU-seconds of compute every second — 5,529,600 CU-seconds per day.
Suppose the nightly refresh of a large model consumes 1,000,000 CU-seconds. Because background operations are smoothed over 24 hours, that refresh does not show up as a spike at 06:00. Instead it occupies 1,000,000 ÷ 5,529,600 ≈ 18% of the capacity for every timepoint of the next 24 hours.
Now add four more models with similar refreshes scheduled around the same time. Background load alone sits near 90% all day. Interactive report queries have only the remaining ~10% before the capacity starts borrowing from the future — and once more than 10 minutes of future capacity is committed, interactive delays begin.
The users who complain at 15:00 are not causing the problem; they are simply the first to notice it. The fixes follow directly from the arithmetic:
- Make each refresh cheaper (incremental refresh, removing unused columns) — halving refresh CU-seconds halves the all-day background layer.
- Refresh less often where the business does not need it.
- Move heavy batch work to a separate capacity if it genuinely needs to run.
Scaling to F128 would also work — by doubling the bill to absorb load that may be largely avoidable.
4. Reading the Capacity Metrics app
The Fabric Capacity Metrics app is where throttling is diagnosed. A practical sequence:
- Start with the utilisation chart over the last few days. Separate the background layer (smoothed refreshes and jobs) from interactive spikes. If the background layer alone is high, refreshes are your problem.
- Open the throttling view and find when interactive delay, interactive rejection or background rejection occurred. Note the timepoints.
- Check overages and carryforward around those timepoints to see how much future capacity was already committed.
- Drill through a throttled timepoint to the operations that contributed. Sort by CU usage: usually a handful of items account for most of it — one oversized refresh, one dataflow, or a few very expensive queries.
- Map items to owners. The fix is almost always in a specific model, refresh or report, so you need to know whose it is.
Repeat across a few busy days. A pattern (every Monday at 09:00, every month-end) tells you far more than a single incident.
5. The usual causes — and the fixes
Refresh collisions. Several large models refreshing at the same time create a block of background load that is smoothed across the whole day. Stagger them, move non-critical ones off-peak, and use incremental refresh so each run does less.
Oversized models. Memory and refresh cost scale with model size. Removing high-cardinality columns and unused data often halves refresh cost.
Expensive DAX. A few measures that force the single-threaded formula engine to iterate large tables can consume more CUs than hundreds of normal queries. Find them in the drill-through, then in DAX Studio. The DAX Measure Doctor flags common patterns statically.
Unexpected workloads. Spark notebooks, dataflows or Copilot usage sharing the same capacity as production reports. Separate them, or schedule them for quiet periods.
6. Scale up, or fix?
Scaling up is sometimes the right answer — a genuinely growing workload needs more capacity. But it doubles the bill at each step, and if the cause is one runaway refresh, a bigger SKU only moves the throttling threshold.
A useful rule: if a small number of items account for most of the CU usage at throttled timepoints, fix those first. If usage is spread evenly across many well-behaved items and still exceeds the capacity, scale.
Use the capacity estimator to sanity-check the size you need, and the Fabric SKU comparison to see what the next size costs. If you want a second pair of eyes on the metrics, the Model Care retainer includes a monthly capacity check-in.
Throttling? Find the cause before you scale
Estimate CU and memory pressure from your workload, or have an engineer review the model behind the spike.
Get the next Engineering Note
Practical DAX, semantic model and Fabric notes. 1–2 a month, unsubscribe any time.