What Is the FileHippo Safety Guarantee?
# Bugs Fixed
* InnoDB: A regression fix resulted a memory leak for memcached insert operations.
* InnoDB: The trx->error_key_num field was not initialized in the error injection code found in storage/innobase/handler/handler0alter.cc. The error_key_num field is usually 0 but can be a non zero value if the memory buffer of a DDL transaction object is reused.
* InnoDB: Fault-tolerant code found in the log apply code for InnoDB ALTER TABLE ... IN PLACE could result in data corruption.
* InnoDB: Databases names beginning with a digit would cause a full-text search (FTS) parser error.
* InnoDB: An ALTER TABLE ... CHANGE [COLUMN] operation would result in an rbt_empty(index_cache->words) assertion.
* InnoDB: Running SHOW ENGINE INNODB STATUS on one connection thread and killing that thread by running a KILL CONNECTION statement from a different connection thread would result in a severe error.
* InnoDB: An excessive amount of memory would be consumed when querying INFORMATION_SCHEMA.INNODB_FT_INDEX_TABLE. The problem would occur for very large full-text search indexes.
* InnoDB: CHECK TABLE would ignore the QUICK option.
* InnoDB: The information_schema.innodb_metrics index_merge counter was not incremented in btr0btr.cc. This patch also introduces new counters (index_page_reorg_attempts, index_page_reorg_successful and index_page_discards) and renames the index_merges counter to “index_page_merge_attempts” to distinguish it from the index_page_merge_successful counter.
* InnoDB: Setting the O_DIRECT flag on a file on tmpfs on some operating systems would result in an error printed to the error log. Creating multiple temporary tables on tmpfs would cause the error to be printed repeatedly. The error message has been changed to a warning that is only printed once when running CREATE TABLE multiple times.
* InnoDB: In debug builds, test case failures would occur due to ibuf_contract_ext performing merges and dict_stats_update returning evicted pages back into the buffer pool while ibuf_change_buffering_debug is enabled.
* InnoDB: InnoDB would fail to return an error when attempting to run a query after discarding the tablespace.
* InnoDB: A severe error would occur after discarding a tablespace.
* InnoDB: During a TRUNCATE TABLE operation, InnoDB: Trying to TRUNCATE a missing index of table ... warnings would be printed to the error log. These warnings should not be printed when the index is a full-text search (FTS) index.
* InnoDB: Full-text index creation on a large table would fail due to insufficient temporary table space and result in a misleading “incorrect key file” error.
* InnoDB: The InnoDB mecached plugin would update a record before inserting to the binlog, which would cause slave server replication to stop. The insert should occur before the update.
* InnoDB: During parallel full-text search (FTS) index creation, a scanner thread reads in documents and passes them to the tokenizer. The tokenizer frees documents from memory when tokenization is complete. When tokenizing documents with a large amount of text, the tokenizer thread would not keep pace with the scanner thread. As a result, memory would not be freed fast enough and the “tokenization pending list” would grow in size.
* InnoDB: A full-text search (FTS) BOOLEAN MODE query with an invalid character in the query string could result in a memory access violation failure.
* InnoDB: trx_create and trx_free would be called for every memcached get request.
* InnoDB: The hardcoded size for the srv_max_n_threads variable was insufficient. The variable setting is now configured based on the maximum number of connection threads and InnoDB background threads.
* InnoDB: In btr_validate_level there are checks to ensure that all B-tree pages are marked when allocated. The checks would fail on the change buffer because the allocation of change buffer pages is handled differently than other B-tree pages.
* InnoDB: When the change buffer is enabled, InnoDB would fail to write a transaction log record when merging a record from the insert buffer to a secondary index page if the insert was performed as an “update-in-place”.
* InnoDB: A SELECT COUNT (*) query would run very slowly when run concurrently with a LOAD DATA operation.
* InnoDB: An existing full-text index would become invalid after running ALTER TABLE ADD FULLTEXT due to an unsynchronized full-text cache.
* InnoDB: Due to a regression in MySQL 5.6, creating or dropping tables with innodb_force_recovery set to 3 (SRV_FORCE_NO_TRX_UNDO) would fail. Additionally, this bug fix includes a code modification that sets InnoDB to read-only when innodb_force_recovery is set to a value greater than 3 (SRV_FORCE_NO_TRX_UNDO).
* InnoDB: An InnoDB memcached configuration error message contained an incorrect file name. The error message stated, Please create config table containers in database innodb_memcache by running innodb_config.sql. error 31. The correct file name is innodb_memcached_config.sql. Also, the “error 31” portion of the error message has been translated to its text equivalent, which is “Table not found”.
* InnoDB: In mutex_spin_wait(), the sync_array_reserve_cell function could fail to find an empty slot on systems with sync wait arrays that are small in size, resulting in an error.
* InnoDB: When index_read_map is called for an exact search and fails to return a record due to non-matching search criteria, the cursor would be positioned on the next record after the searched key. A subsequent call to index_next would return the next record instead of returning the previous non-matching row, thereby skipping a record.
* InnoDB: Full-text search (FTS) index savepoint information would not be set resulting in a severe error when attempting to rollback to the savepoint.
* InnoDB: Converting a table with a large number of columns from MyISAM to InnoDB would cause an assertion due to insufficient log buffer space. Instead of asserting, InnoDB now attempts to increase log buffer size automatically if the redo log size is too large.
* Partitioning: The storage engine was set incorrectly during a rebuild of a partition; the table storage engine was ignored and the default storage engine used instead. Thus, in MySQL 5.1, it was possible for REBUILD PARTITION to change the partition storage engine from InnoDB to MyISAM, and for the reverse (rebuilding partitions of MyISAM tables causing the partitions to use InnoDB) to occurin MySQL 5.5 and later. Now, when rebuilding partitions, the storage engine actually used by the table is checked and used by the handler for the rebuild operation, so that the partition storage engine is not inadvertently changed.
* Partitioning: After disabling the parent table's indexes with ALTER TABLE ... DISABLE KEYS, rebuilding any of its partitions enabled the indexes on those partitions, leading MyISAM to fail with an error when the optimizer tried to use one of the affected indexes.
* Replication: A replication master did not handle correctly the disabling of the semisync plugin on the master and the slave, with a subsequent stopping of the slave.
* Replication: The final argument in the SET clause of a LOAD DATA ... SET statement was repeated in the binary log.
* Replication: When an error encountered by the dump thread while reading events from the active binary log file was a temporary error, so that the dump thread tried to read the event, it was possible for the dump thread to seek the wrong position, which could cause one or more events to be resent. To prevent this, the thread's position is obtained after each correct read of an event.
* Replication: When stopping the I/O thread, it was possible with a very large transaction (equivalent to a binary log size greater than 100MB) that the thread did not receive the transaction to the end. When reconnecting with MASTER_AUTO_POSITION=1 it then tried to fetch changes from the next transaction, which could lead to loss of the incomplete transaction and its data.
* Replication: Setting rpl_semi_sync_master_enabled while the master was waiting for a reply from the slave could in some cases cause the master to fail.
* Replication: Trying to set CHANGE MASTER TO ... MASTER_AUTO_POSITION = 0 failed with error 1777 (ER_AUTO_POSITION_REQUIRES_GTID_MODE_ON).
* Replication: The value of LAST_INSERT_ID() was not correctly replicated when filtering rules were used on the slave.
* Replication: An internal function used for storing GTID values could sometimes try to handle them as strings of the wrong length.
* Replication: During row-based replication with binlog_row_image set to MINIMAL, updating only some columns of a table having 9 or more columns caused mysqlbinlog to fail when it was used with the --verbose option.
* Replication: Issuing a GRANT statement with invalid parameters caused the master to write LOST_EVENTS events into its binary logs, causing replication to stop. Now such cases, if one or more grants or revocations of privileges are successful, an incident is written to the log; otherwise, only a warning is logged.
* libmysqlclient version 18 files were removed from MySQL-shared-compat RPM packages to avoid a conflict between the MySQL-shared and MySQL-shared-compat RPM packages.
* Enabling Index Merge optimizer switches and setting a small sort_buffer_size value could lead to a server exit.
* Semi-join materialization strategy was not used for VARCHAR columns longer than 512 bytes, resulting in use of a less-efficient strategy and worse query performance. (The limit in characters rather than bytes depends on the column character set; 170 characters for utf8, for example.)
* Some license and documentation files were missing from Windows MSI packages.
* Selecting from the Performance Schema session_connect_attrs table under high load could cause a server exit.
* Compilation failures under Visual Studio 2012 were corrected.
* The CLIENT_CONNECT_WITH_DB flag was improperly handled in the C client library. This could lead to a malformed packet sent to the server.
* An assertion was raised if SET PASSWORD was used for an account that has been manually deleted from the mysql.user table but still present in memory.
* The mysql_options() C API function could leak memory if called more than once with the MYSQL_SET_CLIENT_IP option.
* The CONV() function could call abs(INT_MIN), which is undefined, and cause a server exit.
* An error array in the SSL code was missing a comma, leading to implicit concatenation of adjacent messages and a resulting off-by-one error in the relationship between error numbers and messages.
* The filesort implementation sometimes failed to allocate enough buffer space, leading to a server exit.
* GRANT without an IDENTIFIED BY clause resulted in an error even for existing users.
* GROUP_CONCAT() with an invalid separator could cause a server exit.
* An internal InnoDB string routine could write past the end of a buffer.
* GIS intersection-related code was missing a return value check, leading to a loop in nondebug builds and a raised assertion in debug builds.
* Using the binary client/server protocol, the second execution of a prepared statement for a query with parameters in the LIMIT clause raised an assertion.
* For upgrades using the Windows MSI package installer, the upgrade dialog message was missing the “from” version.
* Very long database names in queries could cause the server to exit.
* Standalone Windows MSI packages did not have the ALLUSERS property set. They now set ALLUSERS=1. For earlier MSI packages in this MySQL series, a workaround is to use the following command:
- C:\> msiexec /i msi_installer_name ALLUSERS=1
* Recursion by functions called as arguments to XPath expressions was not handled correctly, sometimes causing such expressions to fail.
* For an ALTER TABLE statement that renamed or changed the default value of a BINARY column, the alteration was done using a table copy and not in place.
* Some possible cases of memory use after being freed were fixed. Thanks to Jan Staněk for the patch.
* In debug builds, static initialization code could call DBUG functions before the DBUG subsystem was initialized.
* With the thread pool plugin enabled, the the PROCESSLIST_USER and PROCESSLIST_HOST columns of the Performance Schema threads table were always NULL for client sessions. Also, for the main thread, those columns were not NULL but set to a user account.
* COUNT(DISTINCT) should not count NULL values, but they were counted when the optimizer used Loose Index Scan.
* For queries of the form UPDATE ... WHERE unique_key ORDER BY ... LIMIT ..., incorrect rows could be updated. Unique keys permit multiple NULL values, but the optimizer did not always consider all of them.
* If asked to upgrade a server that was running without InnoDB enabled, mysql_upgrade issued complaints about InnoDB tables not existing (tables that will not exist unless InnoDB is available).
* In some cases, range conditions over indexes defined on column prefixes returned incomplete result sets. (For example, SELECT ... WHERE 'abcdef1' < col_name AND col_name < 'abcdef9', where the index on col_name indexed only the first 6 characters.)
* Performance Schema instrumentation overhead was reduced for frequent connect/disconnect operations.
* Setting host_cache_size at startup had no effect.
* InnoDB full-text searches failed to find records within transactions that included savepoints.
* Some INSERT INTO ... SELECT ... FROM statements were slow unless the tmp_table_size and max_heap_table_size system variables were set large enough to permit the temporary table used for query processing to be stored in the MEMORY storage engine.
* Incorrect reference counting in the range optimizer module resulted in potential for missing or duplicate rows in the query result set.
* Host names in grant tables are stored in lowercase, but mysql_install_db could fail to observe this convention, leading to accounts that could not be dropped with DROP USER.
* The server uses the ethernet hardware address for UUID generation, but made assumptions about the names of ethernet devices rather than querying the system for their names.
* Several issues identified by the Coverity static analysis tool were fixed. Thanks to Jan Staněk and Honza Horak for the patches.
* Missing va_end() calls were added to logging and UCS2 code. Thanks to Jan Staněk for the patch.
* MySQL did not compile on Mac OS X 10.9 (Mavericks).
* Killing a query that is performing a filesort operation resulted in an ER_SERVER_SHUTDOWN (Server shutdown in progess) error.