Ok, I can see how I didn't manage to fully explain myself - by replying to your "collisions" argument I actually meant something more than that and while it seemed obvious in my head what I wanted to say, I can agree that my wording was imprecise. Sorry about that.

What I pictured in my head is that the ingestion workloads in contrast to the read-only workloads are different in the fact that collisions in former can trigger hash-map resizing/rehashing and which is why I say that these are two distinctive workloads. Moving bunch of data to and from the buckets.

So, my hypothesis is the following: if I PGO the implementation using the dataset (volume, data_distribution) that it triggers a handful of such resizing/rehashing operations, and we observe the positive improvements in wall-time runtime performance, what are the chances that using the same binary but now over a completely different dataset will preserve the same runtime performance improvements? What you're saying, if I understand correctly, is that the improvements will be the same regardless of the dataset (and even type of workload) and I am saying I am not so sure about that.