[MDEV-40804] Interrupted CREATE TABLE with a FULLTEXT index leaves use_stopword empty and stopwords off - Jira

XML

Word

Printable

Details

Type: Bug
Status: Open (View Workflow)
Priority: Major
Resolution: Unresolved
Affects Version/s: 10.11, 11.4, 11.8, 13.1
Fix Version/s: 10.11, 11.4, 11.8, 13.1
Component/s: Full-text Search, Storage Engine - InnoDB
Labels:
None

Description

--source include/have_innodb.inc

# An interrupted CREATE TABLE with a FULLTEXT index must keep the table stopword setting.

# The stopword 'the' must not be indexed, so each count is 0.

SET SESSION max_statement_time=0.001;

CREATE TABLE ct1 (id INT PRIMARY KEY, txt TEXT, FULLTEXT(txt)) ENGINE=InnoDB;

SET SESSION max_statement_time=0.005;

CREATE TABLE ct2 (id INT PRIMARY KEY, txt TEXT, FULLTEXT(txt)) ENGINE=InnoDB;

SET SESSION max_statement_time=0.02;

CREATE TABLE ct3 (id INT PRIMARY KEY, txt TEXT, FULLTEXT(txt)) ENGINE=InnoDB;

SET SESSION max_statement_time=0;

INSERT INTO ct1 VALUES (1,'the quick brown fox');

INSERT INTO ct2 VALUES (1,'the quick brown fox');

INSERT INTO ct3 VALUES (1,'the quick brown fox');

SELECT COUNT(*) FROM ct1 WHERE MATCH(txt) AGAINST('the');

SELECT COUNT(*) FROM ct2 WHERE MATCH(txt) AGAINST('the');

SELECT COUNT(*) FROM ct3 WHERE MATCH(txt) AGAINST('the');

DROP TABLE ct1,ct2,ct3;

Leads to:

CS 13.1.0 0ee34e07bdc10abf32a535d7234c6f5e0bb9ed55 (Debug, Clang 22.1.8-20260622) Build 14/08/2026
2026-08-18 8:48:56 4 [ERROR] InnoDB: (Operation interrupted) writing `use_stopword'

The effect, comparing an interrupted table against a clean one. the is a default stopword, so it must not be indexed:

USE test;

SET SESSION max_statement_time=0.001;

CREATE TABLE ct1 (id INT PRIMARY KEY, txt TEXT, FULLTEXT(txt)) ENGINE=InnoDB;

SET SESSION max_statement_time=0;

CREATE TABLE ct2 (id INT PRIMARY KEY, txt TEXT, FULLTEXT(txt)) ENGINE=InnoDB;

INSERT INTO ct1 VALUES (1,'the quick brown fox');

INSERT INTO ct2 VALUES (1,'the quick brown fox');

SELECT COUNT(*) AS ct1_the FROM ct1 WHERE MATCH(txt) AGAINST('the');

SELECT COUNT(*) AS ct2_the FROM ct2 WHERE MATCH(txt) AGAINST('the');

SELECT COUNT(*) AS ct1_fox FROM ct1 WHERE MATCH(txt) AGAINST('fox');

SELECT COUNT(*) AS ct2_fox FROM ct2 WHERE MATCH(txt) AGAINST('fox');

SET GLOBAL innodb_ft_aux_table='test/ct1';

SELECT * FROM information_schema.INNODB_FT_CONFIG;

SET GLOBAL innodb_ft_aux_table='test/ct2';

SELECT * FROM information_schema.INNODB_FT_CONFIG;

ct1_the

ct2_the

ct1_fox

ct2_fox

KEY	VALUE

optimize_checkpoint_limit	180

synced_doc_id	0

stopword_table_name

use_stopword

KEY	VALUE

optimize_checkpoint_limit	180

synced_doc_id	0

stopword_table_name

use_stopword	1

ct1 indexed the stopword, ct2 did not. Both match fox, so the index otherwise works. use_stopword is empty for ct1 and 1 for ct2.

The write path, and where the interrupt enters it:

create_table_info_t::create_table_update_dict()      ha_innodb.cc:13499

  innobase_fts_load_stopword()                       ha_innodb.cc:11724

    innobase_fts_update_stopword_config()            ha_innodb.cc:11653

      fts_config_set_ulint()                         fts0config.cc:171   <- logs the error

        fts_config_set_value()                       fts0config.cc:125

          FTSQueryExecutor::update_config_record()   fts0exec.cc:351

            QueryExecutor::replace_record()          row0query.cc:495

              QueryExecutor::select_for_update()

                row_search_mvcc()                    row0sel.cc:5022     <- returns DB_INTERRUPTED

innobase_fts_update_stopword_config() returns false on that error, so innobase_fts_load_stopword() fails and fts_optimize_add_table() is skipped. The CREATE TABLE itself still commits and reports success.

Nothing later repairs it. The reader in fts0fts.cc:5052 takes the empty value through strtoul(), which gives 0, sets cache->stopword_info.status to STOPWORD_OFF and returns success. So an empty value reads as "stopwords are off" and raises no error. The value is on disk, so a restart does not clear it: after a restart MATCH(txt) AGAINST('the') still returns the row and use_stopword is still empty.

trx_is_interrupted() in ha_innodb.cc:3621 is true for any statement whose kill level is set, so KILL QUERY and a lost client connection reach the same path. max_statement_time is only the easiest way to show it.

On 13.0 this config value is written through the InnoDB internal parser, fts_parse_sql() plus fts_eval_sql(), which runs the statement in the que layer. That layer holds no interrupt check, so the write finishes even after the statement timer has expired. ~~MDEV-28730~~, commit 81c3836bb2ba0823fe69fd6953c21fccf6476b92, replaced the internal parser with FTSQueryExecutor, and the new path runs through row_search_mvcc(), which does check. The same CREATE TABLE takes 26.8 ms on CS 13.0.2 and 29.5 ms on CS 13.1.0, so a 1 ms timer expires on both, yet only CS 13.1.0 loses the value.

Bug Detection Matrix
Rel o/d Build Commit UniqueID observed
CS 10.6 dbg 220726 5520c9aac5b6a5a9ea60d78582ddffa5349e2d0f No bug found
CS 10.6 opt 220726 5520c9aac5b6a5a9ea60d78582ddffa5349e2d0f No bug found
CS 10.11 dbg 220726 6268f6023fe3e8dc8ec81869dea711e14cbb28c3 No bug found
CS 10.11 opt 220726 6268f6023fe3e8dc8ec81869dea711e14cbb28c3 No bug found
CS 11.4 dbg 220726 2cbce592c39dd39902c8f03db2b35069832fd10b No bug found
CS 11.4 opt 220726 2cbce592c39dd39902c8f03db2b35069832fd10b No bug found
CS 11.8 dbg 220726 1d23deff797ab448f091e39756434903d81a98c2 No bug found
CS 11.8 opt 220726 1d23deff797ab448f091e39756434903d81a98c2 No bug found
CS 12.3 dbg 220726 9b075b2cb53338f67e7230cc585e02d2d34514ed No bug found
CS 12.3 opt 220726 9b075b2cb53338f67e7230cc585e02d2d34514ed No bug found
CS 13.0 dbg 220726 84c246ca5387c0611f75097136ff4f4bea092aa3 No bug found
CS 13.0 opt 220726 84c246ca5387c0611f75097136ff4f4bea092aa3 No bug found
CS 13.1 dbg 140826 0ee34e07bdc10abf32a535d7234c6f5e0bb9ed55 INNODB_ERROR\|(Operation interrupted) writing `use_stopword'
CS 13.1 opt 140826 0ee34e07bdc10abf32a535d7234c6f5e0bb9ed55 INNODB_ERROR\|(Operation interrupted) writing `use_stopword'
ES 10.6 dbg 220726 fcecb2620f25965723d640decede7c018bcb1dcc No bug found
ES 10.6 opt 220726 fcecb2620f25965723d640decede7c018bcb1dcc No bug found
ES 11.4 dbg 220726 3b34189bfe675c18c4ced3ef531d016ea74c76f4 No bug found
ES 11.4 opt 220726 3b34189bfe675c18c4ced3ef531d016ea74c76f4 No bug found
ES 11.8 dbg 220726 4694e931d10fecf733c34f83ea2146d31b708eb3 No bug found
ES 11.8 opt 220726 4694e931d10fecf733c34f83ea2146d31b708eb3 No bug found
ES 12.3 dbg 220726 9d8abb61e913bec023cd8caeccad4b42717151cb No bug found
ES 12.3 opt 220726 9d8abb61e913bec023cd8caeccad4b42717151cb No bug found

Attachments

Issue Links

is caused by

MDEV-28730 Remove internal parser usage from InnoDB fts

Closed

relates to

MDEV-39621 FTSQueryExecutor::read_config_with_lock crashes on NULL thd during OPTIMIZE TABLE

Closed

MDEV-39642 InnoDB FTS: Assertion `doc_id > node->last_doc_id' failed in fts_optimize_encode_node during OPTIMIZE TABLE

Closed

MDEV-40806 A handled user interrupt on an InnoDB FULLTEXT table writes [ERROR] to the error log

Open

Activity

People

Assignee:: Thirunarayanan Balathandayuthapani

Reporter:: Roel Van de Paar

Votes:: 0 Vote for this issue

Watchers:: 2 Start watching this issue

Dates

Created:: 2026-08-18 03:14

Updated:: 4 days ago 17:04

Time Tracking

Estimated:

Not Specified

Remaining:

Logged:

3h 35m

Git Integration

Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.