Log Message:
databases/mysql80-{client|server}: Update to 8.0.39

Changelog:
Changes in MySQL 8.0.39 (2024-07-23, General Availability)

Bugs Fixed

  * InnoDB: In some cases, following the creation of a very large number of
    tables (8001 or more), the server could not be restarted successfully. (Bug
    #36808732)

    References: This issue is a regression of: Bug #33398681.

  * InnoDB: Improved tablespace file scan performance at startup. (Bug #110402,
    Bug #35200385)

  * Group Replication: Running a CREATE TABLE ... SELECT statement on a source
    coming from an asynchronous channel to Group Replication led to errors on
    the replica. (Bug #36784284)

Changes in MySQL 8.0.38 (2024-07-01, General Availability)

Important

This release is no longer available for download. It was removed due to a
critical issue that could stop the server from restarting following the
creation of a very large number of tables (8001, or more). Please upgrade to
MySQL 8.0.39 instead.

  * Audit Log Notes

  * C API Notes

  * Compilation Notes

  * Connection Management Notes

  * Data Dictionary Notes

  * Pluggable Authentication

  * Thread Pool Notes

  * Bugs Fixed

Audit Log Notes

  * Audit log pruning did not function after removing or renaming a file from
    the audit log. Now pruning continues in such cases, but with a warning
    printed in the error log stating that it was not possible to delete the
    missing audit log file. (Bug #35902913)

C API Notes

  * C API applications stalled while receiving results for server side prepared
    statements.

Compilation Notes

  * Upgraded the bundled googletest and googlemock sources to version 1.14.0.
    (Bug #36562482)

  * Added a missing dependency on GenError. (Bug #36551721)

  * It is now possible on Linux systems to build MySQL using a bundled tcmalloc
    library that is provided with the source by specifying -DWITH_TCMALLOC=
    BUNDLED. This is supported on Linux only. (Bug #36313839)

  * The bundled tcmalloc() is now used when building MySQL on Enterprise Linux
    8. (Bug #114844, Bug #35674008)

  * Linux aarch64 platform binaries are now built using patchelf --page-size=
    65536 for compatibility with systems using either 4k or 64k for the page
    size. (Bug #114233, Bug #36393794)

Connection Management Notes

  * The conn_delay/Waiting in connection_control plugin stage was not reset
    after a delay introduced by the connection control plugin which resulted in
    incorrect monitoring information. (Bug #35205358)

Data Dictionary Notes

  * Attempting to upgrade a MyISAM table containing a mix of regular columns
    and generated columns from MySQL 5.7 to 8.0 or later led to table
    corruption. (Bug #105301, Bug #33503328)

Pluggable Authentication

  * The deprecation warning issued when authenticating with the
    mysql_native_password plugin is now issued only once. (Bug #35792948)

Thread Pool Notes

  * Connecting to a thread group that had no connection handler threads
    stalled. We fix this by making sure that connection handler threads
    terminate only if there is at least one connection thread left. (Bug #
    36550125)

Bugs Fixed

  * InnoDB: MySQL unexpectedly halted on an UPDATE after an ALTER TABLE
    operation. (Bug #36571091)

    References: This issue is a regression of: Bug #35183686.

  * InnoDB: The log index size calculation now accounts for column order
    changes. (Bug #36526369)

    References: This issue is a regression of: Bug #35183686.

  * InnoDB: File system operations performed by InnoDB now consistently fsync
    the parent directory when performing directory altering tasks. (Bug #
    36174938)

  * InnoDB: In debug builds, setting the innodb_interpreter_output debug
    variable would cause the server to unexpectedly halt. This is now a
    read-only variable. (Bug #36041032)

  * InnoDB: For tables created with an index on a column that was too wide for
    the redundant row format (allowed before MySQL 5.7.35), an in-place upgrade
    silently imported the table but it was not accessible, which interfered
    with making backups. Now all operations that involve using the invalid
    index are rejected with ER_INDEX_CORRUPT until the index is dropped. An
    ER_IB_INDEX_PART_TOO_LONG error is also reported in the error log. (Bug #
    35869747)

    References: See also: Bug #34826861.

  * InnoDB: An InnoDB assertion error referencing an invalid column index was
    triggered when the column index was valid. (Bug #34800754)

  * InnoDB: With an empty XA transaction, shutting the server down after an XA
    START would cause the server to halt unexpectedly. (Bug #32416819)

  * InnoDB: Shutting down the replication applier or binlog applier while
    processing an empty XA transaction caused the system to unexpectedly halt.
    (Bug #32416819)

  * InnoDB: Removed unnecessary heap usage in the Validate_files::check()
    function.

    Our thanks to Huaxiong Song for the contribution. (Bug #115041, Bug #
    36626203)

  * InnoDB: If a partition table was read with innodb_parallel_read_threads=1,
    read performance greatly decreased from any table after 256 reads. InnoDB
    behaved as if it reached the maximum capacity of parallel read threads
    despite not using any.

    Our thanks to Ke Yu for the contribution. (Bug #114154, Bug #36347408)

  * InnoDB: The result from a spatial index containing a column with a spatial
    reference identifier (SRID) attribute was empty. In addition, using FORCE
    INDEX to force a covering index scan on a spatial index led to an
    assertion. (Bug #112676, Bug #114200, Bug #35894664, Bug #36361834)

  * InnoDB: Fixed performance issues related to querying the data_lock and
    data_lock_waits tables when thousands of read-only transactions were
    present. (Bug #109539, Bug #34951273)

  * Replication: If a source contained a stored, generated column populated by
    a JSON function and binlog_row_image was set to MINIMAL, any subsequent
    update or deletion on the underlying column failed with the following
    error:

            Invalid JSON text in argument 1 to function json_extract: 'The \ 
document is empty.'

    The replica attempted to re-evaluate the generated column and failed with
    that error because the underlying column was unavailable. As of this
    release, stored, generated columns are not re-evaluated when the underlying
    columns are unavailable. (Bug #36515172)

  * Replication: When running GTID-based replication with relay_log_space_limit
    enabled, a restart of the auto positioning protocol sometimes resulted in
    an infinite loop, leading to a deadlock in replication. This was because
    relay_log_space_limit was not honoured, not only for transactions whose
    size exceed this limit, but when the replica could not purge previous logs
    as well.

    To fix this issue, we make the following changes:

      o The receiver respects relay_log_space_limit as set by the user, unless
        a transaction received by the receiver cannot fit into the purged relay
        log. Before queueing the received transaction, receiver now checks
        whether scheduling a full transaction is possible. If not, the receiver
        performs the following actions:

          # Sets the flag indicating that receiver is waiting

          # Rotates the relay log

          # Waits until it is notified that relay log purge was executed and
            that the applier has purged all available relay logs; after this,
            the receiver may queue a transaction without checking the limit
            again

      o Before moving to the next file, the coordinator checks whether the
        receiver is waiting for available relay log space. If so, the
        coordinator forcibly purges the applied logs, including the current
        relay log file. To purge the current relay log file safely, the
        coordinator must do the following:

          # Synchronize all of its workers before moving to the next file

          # Forcibly update group positions, which is necessary to allow
            current purging of the relay log

          # Update the variable read by the receiver which contains relay log
            filename to which applier was moved

        These operations are allowed because we know that receiver waits at a
        transaction boundary and rotates the relay log before waiting.

    (Bug #36507020)

  * Replication: Worker jobs now contain information about the relay log file
    which initiated the transaction, instead of using the default defined by
    relay_log. (Bug #36395631)

  * Replication: Handling an incident while transactions were being committed
    to the binary log caused MySQL to wait indefinitely. (Bug #35671897)

  * Group Replication: Removed a memory leak from /xcom/gcs_xcom_networking.cc.
    (Bug #36532199)

  * Group Replication: Under certain circumstances, if a primary's host
    experienced network inactivity of 20 seconds or more, the secondaries could
    stop unexpectedly. (Bug #36306144)

  * Group Replication: Under certain circumstances, if garbage collection
    occurred just before a relay log rotation, it could cause the applier to
    stop applying new transactions on the secondary members.

    This was caused by garbage collection incrementing the relay log's
    last_committed and sequence_number, creating a gap in the recorded
    sequence_number after the log rotation. The applier was unaffected if the
    gap occurred anywhere else in the relay log.

    As of this release, only last_committed is updated during garbage
    collection. (Bug #36280130, Bug #36446250)

  * JSON: Added missing checks for error handling to NULLIF(), COALESCE(), and
    the shift (>>) operator. (Bug #113668, Bug #35513196, Bug #36198403)

    References: See also: Bug #31358416.

  * MySQL NDB ClusterJ: Running the ClusterJ test suite resulted in an error
    message saying a number of threads did not exist. That was due to some
    wrong handling of threads and connections, which was corrected by this
    patch. (Bug #36086735)

  * Averages of certain numbers were not always computed correctly. (Bug #
    36563773)

  * The following files in strings contained incorrect license information:

      o mb_wc.h

      o ctype-uca.cc

      o ctype-ucs2.cc

      o ctype-utf8.cc

      o dtoa.cc

      o strxmov.cc

      o strxnmov.cc

    (Bug #36506181)

  * In certain unusual cases, the UpdateXML() function did not process all of
    its arguments correctly. (Bug #36479091)

  * Explaining a query which used FORCE INDEX on a spatial index containing a
    column with SRID attributes led to an unplanned exit. (Bug #36418426)

  * When incrementing the reference count for an expression, underlying
    expressions within this expression are not looked at. While removing an
    expression, after decrementing the reference count, even the underlying
    expressions were examined, which led to unintentional deletion of the
    underlying expressions. This issue manifested in Item_ref::real_item() as
    well as in an assert in sql/item.h. We fix this by not looking at the
    underlying expression unless the current expression contains the only
    remaining reference. (Bug #36204344, Bug #36356279)

  * Under certain conditions, EXPLAIN FORMAT=JSON FOR CONNECTION sometimes led
    to an unplanned exit. (Bug #36189820)

  * Under certain conditions, a race condition could result in the amount of
    RAM used by TABLE_HANDLES increasing to a maximum of 9GB. (Bug #36170903)

  * Some CREATE USER statements were not handled correctly. (Bug #36022885)

  * For a SELECT with ORDER BY and LIMIT, the optimizer first chose a full
    table scan with a very expensive cost, then performed another check and
    used the perform_order_index type of path, but this was not reflected by
    the cost in the optimizer plan. (Bug #35930969)

  * Client connections were not alway terminated correctly during shutdown.
    (Bug #35854919)

  * All internal ACL bitmask variables are now explicitly 32 bits (uint32_t).
    (Bug #35507223)

  * It was not possible to add a functional index on FIND_IN_SET(). (Bug #
    35352161)

  * Running two concurrent OPTIMIZE TABLE statements on the same table with
    fulltext indexes and innodb_optimize_fulltext_only enabled sometimes caused
    the server to exit. (Bug #34929814)

  * Removed a memory leak observed while running authentication_kerberos under
    Valgrind. (Bug #34482788, Bug #36570929)

  * The gen_range() function as implemented by the (deprecated) data masking
    plugin did not always return the correct result.

    This issue affected the data masking plugin only, and did not affect the
    data masking component which supersedes it. (Bug #34163992)

  * Fixed an erroneous comment in include/my_command.h.

    Our thanks to Sho Nakazono for the contribution. (Bug #114507, Bug #
    36455468)

  * It was possible for a deterministic stored function to return an incorrect
    result when the function used JOIN ON inside the return statement. If the
    query needed to be reprepared due to a table metadata caused by, for
    example, FLUSH TABLES between two executions, the ON clause was sometimes
    lost. (Bug #114235, Bug #36379879)

  * The PROCESSLIST_INFO column of THREADS was not updated when executing a
    prepared statement.

    Thanks to Daniel Lenski and Amazon for the contribution. (Bug #104121, Bug
    #33057164)

Changes in MySQL 8.0.37 (2024-04-30, General Availability)

  * Audit Log Notes

  * Character Set Support

  * Compilation Notes

  * Component Notes

  * Optimizer Notes

  * Performance Schema Notes

  * Functionality Added or Changed

  * Bugs Fixed

Audit Log Notes

  * MySQL now calls plugin->deinit() with a valid plugin struct as an argument
    regardless of the plugin's type.

    Our thanks to Martin Alderete for the contribution.

Character Set Support

  * When the character_set_server system variable was set using SET PERSIST or
    SET GLOBAL, it did not take effect for new client sessions or for a client
    establishing a connection to the server after the server was restarted. The
    only workaround was to set the corresponding command-line option when
    starting the server.

    To fix this, we now make sure that, at the time of server restart, the
    configuration data is read in the correct order so that the variable
    setting takes effect as expected. (Bug #35529604)

Compilation Notes

  * Warnings for unused variables are turned into compilation errors when
    compiling with -DMYSQL_MAINTAINER_MODE=1. To avoid this use
    -DMYSQL_MAINTAINER_MODE=0 to disable such errors. (Bug #113662, Bug #
    36198423)

  * Improved the cycle timer for the s390x architecture.

    Our thanks to Jonathan Albrecht for the contribution. (Bug #112845, Bug #
    35949958)

Component Notes

  * The values for component options set using the --loose prefix were not read
    when the component was installed. (Bug #28341329)

Optimizer Notes

  * The multi-range read (MRR) optimization did not perform as well as in
    previous releases. (Bug #113711, Bug #36220640)

Performance Schema Notes

  * User variables assigned decimal values were rounded up in the
    user_variables_by_thread table. (Bug #35781732)

  * Executing SELECT * from performance_schema.data_locks on a server under
    heavy load could cause MySQL to consume too much memory and close
    unexpectedly.

    As of this release, memory used executing such a query is now instrumented
    with memory/performance_schema/data_container, enabling you to observe
    memory consumption. (Bug #35240825)

  * Executing SELECT * from performance_schema.data_locks on a server under
    heavy load could cause a deadlock in InnoDB. (Bug #35068461)

    References: See also: Bug #35240825.

Functionality Added or Changed

  * Important Change; Group Replication: When issued with
    group_replication_consistency set to BEFORE_ON_PRIMARY_FAILOVER, the MySQL
    KILL statement now ignores any consistency guarantees, with any interrupted
    transactions now being rolled back.

  * Important Change: For platforms on which OpenSSL libraries are bundled, the
    linked OpenSSL library for MySQL Server has been updated to version 3.0.13.
    Issues fixed in OpenSSL version 3.0.13 are described at https://
    www.openssl.org/news/cl30.txt. (Bug #36261675)

  * Packaging: Added support for Fedora 40 and Ubuntu 24.04.

  * Microsoft Windows: MySQL Windows binary files (.exe and .dll files) now
    display additional information when their properties are viewed. (Bug #
    36379291)

  * clone plugin version requirements were relaxed to allow cloning between
    different point releases in the same series. In other words, only the major
    and minor version numbers must match when previously the release number
    also had to match.

    For example, cloning of MySQL 8.0.37 to (a future) MySQL 8.0.42, or the
    reverse, is supported. Previous restrictions still apply to versions older
    than 8.0.37, for example, so cloning of MySQL 8.0.36 to any other MySQL
    release (or from a different MySQL release to 8.0.36) remains unsupported.
    (Bug #36293529, WL #15989)

Bugs Fixed

  * Important Change: The Robin Hood hashing library has been replaced with
    unordered_dense. (Bug #36158022)

  * InnoDB; Microsoft Windows: Improved redo log performance on Windows by
    opening redo log files in overlapped mode. (Bug #36154818)

    References: This issue is a regression of: Bug #12527.

  * InnoDB: The log writer calls functions that temporarily release
    log.writer_mutex; when innodb_log_writer_threads=OFF, this potentially led
    to other threads writing to the redo log in between these times. (Bug #
    36425219)

  * InnoDB: Some FTS operations on tables with FTS indexes led to inconsistent
    results. For example, if the server terminated while synchronizing the FTS
    cache or when synchronization occurred concurrently with another FTS
    operation.

    Our thanks to Yin Peng and the Tencent team for the contribution. (Bug #
    36347647)

  * InnoDB: When creating an index on a table containing data, valgrind
    occasionally reported reads of uninitialized memory from
    ddl::Builder::bulk_add_row(). (Bug #36342792)

  * InnoDB: On Windows, keeping a file open without a shared write lock and
    attempting to acquire the fil_shard mutex caused a deadlock with another
    thread that had acquired the fil_shard mutex and was attempting to access
    the same file. (Bug #36159317)

    References: See also: Bug #32808809.

  * InnoDB: Fixed a potential redo log rotation issue that could emit a "Found
    existing redo log files, but at least one is missing" error during
    recovery. (Bug #36124625)

  * InnoDB: Improved os_innodb_umask handling, and made it read-only. (Bug #
    35932118)

    References: This issue is a regression of: Bug #29472125.

  * InnoDB: Found and fixed an assertion failure related to full-text indexes.
    (Bug #35836581)

  * InnoDB: Improved buffer handling during the tablespace deletion process, a
    situation that could have potentially caused an assertion failure. (Bug #
    35676106, Bug #36343647)

  * InnoDB: If a MySQL table in a system schema had an INSTANT ADD column that
    was added before 8.0.29 (they are not allowed as of that version), and
    after MySQL was upgraded to a version greater than 8.0.29, DMLs on these
    tables would result in the server unexpectedly closing.

    Our thanks to Richard Dang for the contribution. (Bug #35625510, Bug #
    35981565, Bug #36180360)

  * InnoDB: The redo log would potentially not log a column order change with
    instant DDL, which could cause an incorrect log replay during recovery.
    (Bug #35183686)

  * InnoDB: Results for SHOW ENGINE INNODB STATUS showed an invalid value
    (NULL) as the first result. (Bug #113819, Bug #36118112)

    References: This issue is a regression of: Bug #34992157.

  * InnoDB: With innodb_parallel_read_threads set to a value greater than 1,
    InnoDB unnecessarily disabled read-ahead heuristics which resulted in
    stalls when pages were not already in the buffer pool. (Bug #113482, Bug #
    36142806)

  * InnoDB: Running a query that used a unique hash index with the TempTable
    storage engine could take significantly more time compared to running the
    query with the MEMORY engine.

    Our thanks to xiaoyang chen for the contribution. (Bug #113178, Bug #
    36037224, Bug #36224958)

  * InnoDB: In debug builds, there was an assertion failure in InnoDB's
    background when a transaction it wanted to acquire an MDL lock on was no
    longer active.

    This fix is based on a patch from Genze Wu with Alibaba, thank you for the
    contribution. (Bug #112424, Bug #35835864)

    References: This issue is a regression of: Bug #33700835.

  * InnoDB: With innodb_parallel_read_threads set to a value greater than 1,
    InnoDB would unnecessarily request asynchronous reads which required more
    synchronization during I/O completion and created a bottleneck due to the
    limited number of available threads (innodb_read_io_threads) for handling I
    /O operations. Now this performs synchronous instead of asynchronous reads.
    (Bug #112137, Bug #35740866)

  * InnoDB: A trx would unexpectedly halt after encountering an incorrect trx->
    in_innodb value.

    Our thanks to Shaohua Wang for the contribution. (Bug #110652, Bug #
    35277407)

  * InnoDB: SELECT ... GROUP BY queries were at least twice as slow with the
    TempTable engine than the Memory engine. (Bug #107700, Bug #34338001)

  * InnoDB: MySQL no longer ignores the optimizer hint to use a secondary index
    scan, which instead forced a clustered (parallel) index scan. (Bug #100597,
    Bug #112767, Bug #31791868, Bug #35952353)

  * Group Replication: Problems arose when members M1 and M2 were in a group,
    with M1 using u1 as its recovery user and M2 using u2 as its own recovery
    user, and both users u1 and u2 existing on M1 and M2 with all necessary
    privileges, and when a new member M3 joined the group using u2 as its
    recovery user. M3 knew only of user u2, but did not know of user u1,
    leading START GROUP REPLICATION on M2 to be rejected since M1 was unable to
    connect to M2. This also generated a new view_id listing the group members
    as M1 and M2, but M1 nevertheless continued trying to connect to M3, with
    M1 logging Error in establishing mysql connection and M3 logging Access
    denied errors for the connection attempts from M1.

    By design, XCom stores the last three known configurations, including
    references to physical connections shared among all past and present
    configurations. This is done to facilitate quick reconnections by nodes
    rejoining the group, explicitly or implicitly, and that were already
    present in any of those configurations.

    A side effect of this was that we might keep attempting to connect to a
    node that was currently not in the group. To solve this problem, we inhibit
    error logging if the node is not in the current configuration, in order to
    avoid false negatives which might lead a DBA or an operator to think
    mistakenly that there is a problem in the system. (Bug #36210988)

    References: See also: Bug #32592027.

  * Group Replication: Improved handling of GTID sets. (Bug #36093405)

  * Group Replication: Two cases were found in which a member exited the group
    and moved to the ERROR state, but did not honor the action specified by
    group_replication_exit_state_action; these are listed here:

      o When an error occurred while enabling super_read_only

      o When member join recovery was not possible, due to missing binary logs
        and clone groups on group members

    Example: When the value of group_replication_exit_state_action was
    OFFLINE_MODE and one of these events took place, offline mode was not
    enabled as expected. (Bug #36076308)

  * Group Replication: After successfully setting a new primary,
    group_replication_set_as_primary() in some cases waited indefinitely for
    the operation to complete. (Bug #36059098)

  * JSON: JOIN and GROUP BY handled some JSON column values differently. (Bug #
    101048, Bug #31969607)

  * MySQL NDB ClusterJ: The setLimits() method can now be chained to
    deletePersistentAll() to limit the number of items to delete. See the
    description of deletePersistentAll() for details. (Bug #36049906)

  * Updated BuildRequire rules to align with versions now required for CMake
    and Bison. (Bug #36343254)

  * SET GLOBAL offline_mode=ON did not always perform correctly when issued
    under high loads. (Bug #36275182)

    References: See also: Bug #36405894.

  * Upgraded curl to version 8.6.0. (Bug #36267545)

  * mysqldump did not always interpret the server version correctly. (Bug #
    36248967)

    References: See also: Bug #36405879.

  * Condition pushdown to a view was rejected with a collation mismatch if the
    view was created with a different character set than the character set used
    when querying the view. (Bug #36246859)

  * Use of the deprecated exec_program() command has been replaced by
    execute_process() to provide compatibility with CMake 3.28.1 and later.
    (Bug #36220656)

  * Upgraded the protobuf library to version 25.1. (Bug #36108397)

  * For building Enterprise Linux RPMs, the build scripts now point to a newer
    strip command (under /opt/rh/gcc-toolset-12), and they now check that the
    corresponding dwz tool is available. Previously this was only implemented
    for EL8. (Bug #36090069)

  * We now look for gcc-ar and gcc-ranlib when building on Oracle Linux with
    link-time optimization. (Bug #36089900)

  * The server sometimes terminated unexpectedly in response to a specific
    query. (Bug #35957627)

  * Improved view and trigger definer handling by view and table DDL. (Bug #
    35942937)

  * The server did not always return metadata to the client correctly for
    certain queries. (Bug #35904044)

  * When running queries against a table with a multi-value index, the server
    sometimes exited unexpectedly, often while executing a complex SELECT query
    which used this index. (Bug #35789759)

  * Improved code in sql/item_subselect.cc. (Bug #35733778, Bug #35738531, Bug
    #35779012)

  * CREATE USER IF EXISTS was not always logged correctly. (Bug #35530823)

  * The server did not disallow subqueries in partition expressions properly.
    These are invalid, and should cause a syntax error. (Bug #35476172)

  * Some RANK() ... OVER() queries raised an assertion in sql/sql_executor.cc.
    (Bug #35228083)

  * When successive ALTER TABLE ... ALGORITHM=COPY statements were issued
    within 10 seconds of one another, the n_rows value became 0. (Bug #
    35127747)

  * Updated the URL used for downloading the Boost C++ libraries. (Bug #113576,
    Bug #36164514)

  * The fix for a previous issue, first addressed in MySQL 8.0.30, was
    incomplete.

    Our thanks to Hao Lu for the contribution. (Bug #113174, Bug #36035044)

    References: This issue is a regression of: Bug #110801, Bug #35328028.

  * On s390x, we now compile the FMA test with -O2 to avoid overoptimization.

    Our thanks to Jonathan Albrecht for the contribution. (Bug #113096, Bug #
    36016140)

  * Although s390x is a big-endian platform, the little-endian ICU data
    directory was used for compiling.

    Our thanks to Jonathan Albrecht for the contribution. (Bug #113095, Bug #
    36016141)

  * Incorrect results were sometimes obtained from a query that used a group by
    loose index scan. (Bug #112541, Bug #35854362)

  * An assertion failed in debug builds when inserting data with a zero-length
    column, such as CHAR(0) or BINARY(0), into a table. Now, a less strict
    assertion more accurately fails only if it detects that a non-zero number
    of bytes copied from a source is identical to the target. (Bug #111450, Bug
    #35507763)

  * A VALUES statement in a correlated lateral or (other) dependent subquery
    yielded an incorrect result. (Bug #109252, Bug #110076, Bug #34852090, Bug
    #35087820)

  * IN should return NULL not only if the expression on the left hand side is
    NULL, but when no match is found in the list and one of the expressions in
    the list is NULL as well. In some cases, it was found that adding rows to a
    table caused a subsequent prepared SELECT NULL IN (SELECT ...) query to
    return a different result when the result should have remained the same.
    (Bug #94857, Bug #29602835)