CUs calculation

[Bendik]

 Our team is quite confused by how CU's are computed. We have this case from last week, where we went over our limit, and now how to wait for the burn-down. We had a heavy process running, and we got rapported 1 426 222 CU(s), for 61 302 seconds. At the time, we ran on an F8 capacity. In my head, dividing CU / seconds should be equeal to how many compute units we on average used for that process, but given that thats equal to 23, I dont understand how this is computed? This was a synapse notebook that ran. 

[v-shex-msft]

Hi @Bendik ,

Did the above suggestions help with your scenario? if that is the case, you can consider Kudo or Accept the helpful suggestions to help others who faced similar requirements.

If these also don't help, please share more detailed information and description to help us clarify your scenario to test.

How to Get Your Question Answered Quickly 

Regards,

Xiaoxin Sheng

[frithjof_v · Accepted solution]

Fabric is allowed to burst. It means it can consume more than the CU capacity for a time period, and then smooth that CU (s) usage over 24 hours (for background operations).

The duration of 61 300 seconds is roughly 17 hours. In this period, it has on average consumed resources as if it has 23 CUs, but it only has 8 CUs.

Now it will try to smooth this out over a future 24 hours window.

However this has probably been too much (too high CU over too long time), so it enters throttling. Consuming on average 23 CUs for 17 hours in a day will not be sustainable on an F8.

But using 23 CUs on average for 8 hours a day could in theory be sustainable on an F8, due to the 24 hour smoothing:

23CU/8CU * 8h/24h < 100%

But 23 CUs for 17 hours becomes too much for it:

23 CU / 8 CU * 17 h / 24 h = 203.6 %

https://blog.fabric.microsoft.com/en/blog/fabric-capacities-everything-you-need-to-know-about-whats-...

Here's the best video I have seen about the Fabric Capacity Metrics App:

https://youtu.be/EuBA5iK1BiA?si=9AaIyvSvih1i4mm3