Hello, Matthias! Just wanted to update you with some information about the next steps in work.
> In heapam_index_build_range_scan, it seems like you're popping the > snapshot and registering a new one while holding a tuple from > heap_getnext(), thus while holding a page lock. I'm not so sure that's > OK, expecially when catalogs are also involved (specifically for > expression indexes, where functions could potentially be updated or > dropped if we re-create the visibility snapshot) I have returned to the solution with a dedicated catalog_xmin for backends [1]. Additionally, I have added catalog_xmin to pg_stat_activity [2]. > In heapam_index_build_range_scan, you pop the snapshot before the > returned heaptuple is processed and passed to the index-provided > callback. I think that's incorrect, as it'll change the visibility of > the returned tuple before it's passed to the index's callback. I think > the snapshot manipulation is best added at the end of the loop, if we > add it at all in that function. Now it's fixed, and the snapshot is reset between pages [3]. Additionally, I resolved the issue with potential duplicates in unique indexes. It looks a bit clunky, but it works for now [4]. Single commit from [5] also included, just for stable stress testing. Full diff is available at [6]. Best regards, Mikhail. [1]: https://github.com/michail-nikolaev/postgres/commit/01a47623571592c52c7a367f85b1cff9d8b593c0 [2]: https://github.com/michail-nikolaev/postgres/commit/d3345d60bd51fe2e0e4a73806774b828f34ba7b6 [3]: https://github.com/michail-nikolaev/postgres/commit/7d1dd4f971e8d03f38de95f82b730635ffe09aaf [4]: https://github.com/michail-nikolaev/postgres/commit/4ad56e14dd504d5530657069068c2bdf172e482d [5]: https://commitfest.postgresql.org/49/5160/ [6]: https://github.com/postgres/postgres/compare/master...michail-nikolaev:postgres:new_index_concurrently_approach?diff=split&w=
