log raw benchmark measures
In addition to aggregating performance indicators such as the number of objects, their size and the time, log the data from which these performances indicators can be recalculated at a later time. The raw data can be logged every 30 seconds to keep it manageable. Logging every action would not only make a huge log, it would also likely have a measurable impact on the performance indicators themselves.
For instance:
https://forge.softwareheritage.org/T3149#65900
histogram of response times of random reads, in buckets of 5 or 10 ms wide ranges
could be calculated later if the size and response time of all objects are logged.