While reviewing #62282, kv95 benchmarks were shown to regress somewhat. This still holds after landing on master, with e.g. kv95/enc=false/nodes=1/cpu=32 giving 66304 ops/s before the patch and 58862 ops/s after the patch. When reverting to using 1 as the minimum capacity instead of estimatedRowCount in the ResetMaybeReallocate call the performance goes back to the baseline:

While debugging, it was found that estimatedRowCount would often be 0, which would cause ResetMaybeReallocate to allocate a batch of coldata.BatchSize() (i.e. 1024) every time, instead of 1 which was the old behavior. This probably explains the performance drop.