percona server 5.7 super_read_only and gtid_executed table

[Maxim Fedotov]

Suppose that you have percona server 5.7 slave with super_read_only option enabled.
thread/sql/compress_gtid_table will be stuck always rollbacking on commit.

This is due to following change in handler.cc file.
In percona-server 5.7:
  if (rw_trans && !ignore_global_read_lock)
1780 {
1781 /*
1782 Acquire a metadata lock which will ensure that COMMIT is blocked
1783 by an active FLUSH TABLES WITH READ LOCK (and vice versa:
1784 COMMIT in progress blocks FTWRL).
1785
1786 We allow the owner of FTWRL to COMMIT; we assume that it knows
1787 what it does.
1788 */
1789 MDL_REQUEST_INIT(&mdl_request,
1790 MDL_key::COMMIT, "", "", MDL_INTENTION_EXCLUSIVE,
1791 MDL_EXPLICIT);
1792

1793 DBUG_PRINT("debug", ("Acquire MDL commit lock"));
1794 if (thd->mdl_context.acquire_lock(&mdl_request,
1795 thd->variables.lock_wait_timeout))
1796 {
1797 ha_rollback_trans(thd, all);
1798 DBUG_RETURN(1);
1799 }
1800 release_mdl= true;
1801

1802 DEBUG_SYNC(thd, "ha_commit_trans_after_acquire_commit_lock");
1803 }
1804

1805 if (rw_trans && stmt_has_updated_trans_table(ha_info)
1806 && check_readonly(thd, true))
1807 {
1808 ha_rollback_trans(thd, all);
1809 error= 1;
1810 goto end;
1811 }

in mysql 5.7:
 if (rw_trans && !ignore_global_read_lock)
1747 {
1748 /*
1749 Acquire a metadata lock which will ensure that COMMIT is blocked
1750 by an active FLUSH TABLES WITH READ LOCK (and vice versa:
1751 COMMIT in progress blocks FTWRL).
1752
1753 We allow the owner of FTWRL to COMMIT; we assume that it knows
1754 what it does.
1755 */
1756 MDL_REQUEST_INIT(&mdl_request,
1757 MDL_key::COMMIT, "", "", MDL_INTENTION_EXCLUSIVE,
1758 MDL_EXPLICIT);
1759

1760 DBUG_PRINT("debug", ("Acquire MDL commit lock"));
1761 if (thd->mdl_context.acquire_lock(&mdl_request,
1762 thd->variables.lock_wait_timeout))
1763 {
1764 ha_rollback_trans(thd, all);
1765 DBUG_RETURN(1);
1766 }
1767 release_mdl= true;
1768

1769 DEBUG_SYNC(thd, "ha_commit_trans_after_acquire_commit_lock");
1770

1771 if (stmt_has_updated_trans_table(ha_info) && check_readonly(thd, true))
1772 {
1773 ha_rollback_trans(thd, all);
1774 error= 1;
1775 goto end;
1776 }
1777 }

The problem is in this:
   if (rw_trans && stmt_has_updated_trans_table(ha_info)
1806 && check_readonly(thd, true))
1807 {
1808 ha_rollback_trans(thd, all);
1809 error= 1;
1810 goto end;
1811 }

it should be inside if (rw_trans && !ignore_global_read_lock) condition.

This bug causes huge growth of gtid_executed tables and severe IO, because due to stuck thread a lot of IO is generated to iblog files.

[Laurynas Biveinis (laurynas-biveinis)]

This is bug 1654682 (http://bugs.mysql.com/bug.php?id=84437), already fixed in Percona Server 5.7.17-12. Closing as duplicate - please re-open if needed