Forum Discussion
Fabric Performance Extraction
We are currently extracting Fabric Capacity Metrics to generate downscale recommendations for our Fabric capacities. During this process, we observed an issue where the same operation is being recorded across multiple timepoints, even though the operation ran only once within a specific time window.
Because of this behavior, the same CU usage appears in several hourly timepoints, which leads to incorrect utilization calculations when we evaluate hourly capacity usage.
Our concern is whether these repeated entries across timepoints represent actual utilization for each timepoint or if they are simply the way the Capacity Metrics model distributes operations across timepoints.
This impacts our downscale recommendation logic, since it becomes unclear whether all these timepoints should be considered as utilized.
Observed Example
Capacity Name: cadifabric
Operation: Notebook Run
Operation ID: 95f04ce5-8163-4974-b89b-93a6971bbc38
Actual Operation Duration:
Start Time:
2026-03-09 15:53:16 +05:30
End Time:
2026-03-09 16:24:10 +05:30
However, the same operation appears in multiple timepoints in the Capacity Metrics dataset.
Timepoint Status Duration (s) Total CU (s) % Base Capacity
| 2026-03-09 16:00:00 | InProgress | 358 | 1434.104 | 0.0083 |
| 2026-03-09 17:00:00 | Stopped | 1854 | 7419.486 | 0.0429 |
| 2026-03-09 18:00:00 | Stopped | 1854 | 7419.486 | 0.0429 |
| 2026-03-09 19:00:00 | Stopped | 1854 | 7419.486 | 0.0429 |
Even though the notebook execution ended at 16:24, the same operation appears in later timepoints as well.
Questions
Is this expected behavior in the Fabric Capacity Metrics dataset?
Does the model intentionally repeat operations across multiple timepoints for aggregation purposes?
For capacity utilization analysis, should we treat all these timepoints as actual utilization, or should we only consider the timepoints that fall within the operation start and end time?
Hi Rajkumar106,
This is because of the smoothing that Fabric does to CU usage. Background operations are smoothed over a 24 hour period.
Understand your Fabric capacity throttling - Microsoft Fabric | Microsoft Learn
So even though the process stopped, the compute it used is smoothed over a longer period.
To directly answer your questions:
1: yes
2: No
3: you should use all timepoints as actual utilization.
Hi Rajkumar106,
For background operations, that is correct. All background operations are smoothed over a 24 hour window.
Interactive operations are a little more fun, the smoothing window for them is at least 5 minutes, but can be longer depending on what type of operation and how intensive it is.
4 Replies
Hi Rajkumar106,
This is because of the smoothing that Fabric does to CU usage. Background operations are smoothed over a 24 hour period.
Understand your Fabric capacity throttling - Microsoft Fabric | Microsoft Learn
So even though the process stopped, the compute it used is smoothed over a longer period.
To directly answer your questions:
1: yes
2: No
3: you should use all timepoints as actual utilization.
Just to confirm our understanding: since Fabric smooths CU usage over 24 hours, the same operation appears across multiple timepoints even after it stops. Based on this, should we consider all these timepoints as actual utilization?
For our downscale recommendation logic, should these smoothed timepoints also be treated as utilized periods, meaning they are not safe for downscaling?
Hi Rajkumar106,
For background operations, that is correct. All background operations are smoothed over a 24 hour window.
Interactive operations are a little more fun, the smoothing window for them is at least 5 minutes, but can be longer depending on what type of operation and how intensive it is.
Just to confirm our understanding: since Fabric smooths CU usage over 24 hours, the same operation appears across multiple timepoints even after it stops. Based on this, should we consider all these timepoints as actual utilization?
For our downscale recommendation logic, should these smoothed timepoints also be treated as utilized periods, meaning they are not safe for downscaling?
