This is a placeholder issue for such problems, since I was looking into this multiple times, I will add more info later

There are 3 types of issues:

• when RSS is higher the MemoryTracking metric
  • Don't remember anything existing, but I've seen such cases now
• when MemoryTracking is higher then RSS
  • Example of previous fixes/problems
    • Fix inaccurate MemoryTracking metric in case of allocation failures #72603
    • Fix possible overestimate memory tracking #73081
• non-jemalloc builds (memory usage under sanitizers)
  • Example of previous fixes/problems
    • ci: fix MEMORY_LIMIT_EXCEEDED for UBsan jobs (due to 00980_alter_settings_race) #76143
  • Examples of reports that are still exist:
    • https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=81960&sha=6bac7d1bf40b30eed5cc0e01b74253bdb681c913&name_0=PR&name_1=Stateless%20tests%20%28amd_msan%2C%203%2F4%29
  • Todo
    • Resubmit tests: limit memory usage for tests #76388 - ci: limit stateless tests to 5G of RAM and fix tests that exceeds this limit #82361

For the (2) case we can gather jemalloc profiles from CI - #82035

And we have both

If you have some detailed reports, please post them here.