Elena Stepanova added a comment - 2023-10-21 13:06 It is also reproducible on bb-10.10-all-builders c9e5d725bb8c0d8eb28caf6bc766e946fc0cf8d7

Yuchen Pei added a comment - 2023-10-22 23:55 The server restart seems to be needed for reproduction. If I do without restart then it passes: --disable_query_log --disable_result_log --source ../../t/test_init.inc --enable_result_log --enable_query_log   CREATE TABLE t (a INT ) PARTITION BY HASH (a) PARTITIONS 2; ALTER TABLE t REBUILD PARTITION p0;   --disable_query_log --disable_result_log --source ../../t/test_deinit.inc --enable_result_log --enable_query_log

Yuchen Pei added a comment - 2023-10-23 08:19 - edited A mid-low level explanation of what happened: A low level function for getting a free entry which will be later written to the log: static bool ddl_log_get_free_entry(DDL_LOG_MEMORY_ENTRY **active_entry) { DDL_LOG_MEMORY_ENTRY *used_entry; DDL_LOG_MEMORY_ENTRY *first_used= global_ddl_log.first_used; DBUG_ENTER( "ddl_log_get_free_entry" );   if (global_ddl_log.first_free == NULL) // (A) { // [... 9 lines elided] global_ddl_log.num_entries++; // (B) used_entry->entry_pos= global_ddl_log.num_entries; // (C) } else { used_entry= global_ddl_log.first_free; // (D) global_ddl_log.first_free= used_entry->next_log_entry; // (E) } // [... 12 lines elided] } A higher level function that uses ddl_log_get_free_entry() in many occasions: uint fast_alter_partition_table(THD *thd, TABLE *table, Alter_info *alter_info, Alter_table_ctx *alter_ctx, HA_CREATE_INFO *create_info, TABLE_LIST *table_list) { // [... 411 lines elided] if (write_log_drop_shadow_frm(lpt) || // (1) // [... 3 lines elided] write_log_add_change_partition(lpt) || // (2) // [... 7 lines elided] write_log_final_change_partition(lpt) || // (3) // [... 10 lines elided] mysql_rename_partitions(lpt) || // (4) // [... 24 lines elided] } in (1): first_free (!= NULL (A)) is assigned to part_info->execute_entry (D). it has entry_pos 1 in (2): first_free becomes NULL (E) in (3): part_elem->log_entry is allocated because first_free is NULL (A), with its entry_pos to be the new size of the ddl_log, which is also 1 (B) (C) part_elem->log_entry is written to the log file at 1st block part_info->execute_entry is written to the log file at 1st block (overwrite!) in (4): expected the (part_elem->log_entry->entry_pos)th block to be a log entry, but found an execute entry instead. Assertion failure: static bool ddl_log_increment_phase_no_lock(uint entry_pos) { uchar *file_entry_buf= global_ddl_log.file_entry_buf; DBUG_ENTER( "ddl_log_increment_phase_no_lock" );   mysql_mutex_assert_owner(&LOCK_gdl); if (!read_ddl_log_file_entry(entry_pos)) { ddl_log_entry_code code= ((ddl_log_entry_code) file_entry_buf[DDL_LOG_ENTRY_TYPE_POS]); ddl_log_action_code action= ((ddl_log_action_code) file_entry_buf[DDL_LOG_ACTION_TYPE_POS]);   if (code == DDL_LOG_ENTRY_CODE && action < (uint) DDL_LOG_LAST_ACTION) { // [... 14 lines elided] } else { /* Trying to deativate an execute entry or already deactive entry. This should not happen ,*/ DBUG_ASSERT(0); // [... 8 lines elided] } FYI monty

Michael Widenius added a comment - 2023-10-23 17:04 The reason for the bug is that spider plugin does some ALTER TABLE commands before recovery is done. I don't think that is a good idea. It would be better to first fix old tables before creating new tables. For example, as part of 'ha_signal_ddl_recovery_done(). Anyway, the problem is that 'first_free' was not reset (as this was never assumed to be change before recovery). I have now fixed that for 10.10. 10.6 does not have this problem

Michael Widenius added a comment - 2023-10-24 12:45 Pushed to 10.10