Hello Igor,
I would recommend to filter out fast dependencies before they are sent from AI SDK to AI Backend.
You can use
Telemetry Processors to do that. With this approach you will be in control of which dependencies are of interest to you (with duration slower than a threshold and/or with some certain property) and which are not.
Sampling is also a good approach to collect less data but still have a good statistical representation. You can set some aggressive sampling values to lower the amount of the data sent.
Dmitry Matveev