Patch is in bb-10.6-MDEV-30009

Thank you, this is a step to the right direction. I think that we need to enable the logic also in release builds. Furthermore, I think that we must ensure that the purge of transaction history will not add more entries to the change buffer during a slow shutdown.

In an rr replay trace that was produced with innodb_change_buffering_debug=1, it turns out that we must remove the change buffer records in the if (!bitmap_bits) code path of ibuf_merge_or_delete_for_page(), to avoid causing corruption. We got an rr replay trace where changes had been buffered for a 3-field index, the index was dropped, created as a 2-field index and stored in the same page. During ibuf_read_merge_pages(), we reset the bitmap bits but skipped the call to ibuf_delete_recs(). Finally, during a change buffer merge, we attempted to merge bogus 3-field records to the 2-field index page and hit an assertion in page_validate() when misinterpreting the records of the new 2-field index page as 3-field ones:

mysqld: /data/Server/bb-10.6-MDEV-30009/storage/innobase/rem/rem0rec.cc:443: void rec_init_offsets_comp_ordinary(const rec_t*, const dict_index_t*, rec_offs*, ulint, const dict_col_t::def_t*, rec_leaf_format) [with bool mblob = false; bool redundant_temp = false; rec_t = unsigned char; rec_offs = short unsigned int; ulint = long unsigned int]: Assertion `index->is_primary()' failed.

This corruption occurred on an .ibd file. While ALTER TABLE…DROP/ADD INDEX was involved in this case, I believe that the corruption is possible even for DML, when lots of records are being deleted.

In MDEV-20934 some code to remove change buffer records was added to this code path, but only in the case when shutdown is running. It was refactored on a merge to 10.5 due to MDEV-19514. The code was later revised in MDEV-23839.

To avoid reintroducing the corruption MDEV-25783, we must thread carefully here. I think that outside shutdown, that code path should only delete the change buffer entries if (block) (in which case we are holding an exclusive latch on the index page). MDEV-25783 was actually caused by some code in ibuf_read_merge_pages() and not ibuf_merge_or_delete_for_page().

This bug was featured in https://fosdem.org/2023/schedule/event/innodb_change_buffer/