On Mon, Apr 8, 2024 at 12:40 AM Andres Freund <andres@anarazel.de> wrote: > On 2024-03-30 23:33:04 +0200, Alexander Korotkov wrote: > > I've pushed 0001, 0002 and 0006. > > I briefly looked at 27bc1772fc81 and I don't think the state post this commit > makes sense. Before this commit another block based AM could implement analyze > without much code duplication. Now a large portion of analyze.c has to be > copied, because they can't stop acquire_sample_rows() from calling > heapam_scan_analyze_next_block(). > > I'm quite certain this will break a few out-of-core AMs in a way that can't > easily be fixed.
I was under the impression there are not so many out-of-core table AMs, which have non-dummy analysis implementations. And even if there are some, duplicating acquire_sample_rows() isn't a big deal.
But given your feedback, I'd like to propose to keep both options open. Turn back the block-level API for analyze, but let table-AM implement its own analyze function. Then existing out-of-core AMs wouldn't need to do anything (or probably just set the new API method to NULL).
I think that providing both new and old interface functions for block-based and non-block based custom am is an excellent compromise.
The patch v1-0001-Turn-back.. is mainly an undo of part of the 27bc1772fc81 that had turned off _analyze_next_tuple..analyze_next_block for external callers. If some extensions are already adapted to the old interface functions, they are free to still use it.
> And even for non-block based AMs, the new interface basically requires > reimplementing all of analyze.c. . Non-lock base AM needs to just provide an alternative implementation for what acquire_sample_rows() does. This seems like reasonable effort for me, and surely not reimplementing all of analyze.c.