On Sat, Apr 13, 2024 at 5:28 AM Alexander Korotkov <aekorotkov@gmail.com> wrote: > Yes, I think so. Table AM API deals with TIDs and block numbers, but > doesn't force on what they actually mean. For example, in ZedStore > [1], data is stored on per-column B-trees, where TID used in table AM > is just a logical key of that B-trees. Similarly, blockNumber is a > range for B-trees. > > c6fc50cb4028 and 041b96802ef are putting to acquire_sample_rows() an > assumption that we are sampling physical blocks as they are stored in > data files. That couldn't anymore be some "logical" block numbers > with meaning only table AM implementation knows. That was pointed out > by Andres [2]. I'm not sure if ZedStore is alive, but there could be > other table AM implementations like this, or other implementations in > development, etc. Anyway, I don't feel good about narrowing the API, > which is there from pg12.
I spent some time looking at this. I think it's valid to complain about the tighter coupling, but c6fc50cb4028 is there starting in v14, so I don't think I understand why the situation after 041b96802ef is materially worse than what we've had for the last few releases. I think it is worse in the sense that, before, you could dodge the problem without defining USE_PREFETCH, and now you can't, but I don't think we can regard nonphysical block numbers as a supported scenario on that basis.
But maybe I'm not correctly understanding the situation?
Hi, Robert!
In my understanding, the downside of 041b96802ef is bringing read_stream* things from being heap-only-related up to the level of acquire_sample_rows() that is not supposed to be tied to heap. And changing *_analyze_next_block() function signature to use ReadStream explicitly in the signature.