Home - Waterfall Grid T-Grid Console Builders Recent Builds Buildslaves Changesources - JSON API - About

Console View


Categories: connectors experimental galera main
Legend:   Passed Failed Warnings Failed Again Running Exception Offline No data

connectors experimental galera main
Yuchen Pei
MDEV-15621 Auto add RANGE COLUMNS partitions by interval

Allow auto partitioning by interval in PARTITION BY RANGE COLUMNS

PARTITION BY RANGE COLUMNS (col_name)
INTERVAL interval [AUTO]
(
  PARTITION partition_name VALUES LESS THAN (value)
  [, PARTITION partition_name VALUES LESS THAN (value) ... ]
)

where

- col_name is the name of one column of type DATE or DATETIME
- at least one partition is supplied
- INTERVAL interval is a time interval

When performing DML on such a table, it will first add partitions
by the specified interval until the partition covers the current time.

Partition addition will not cause an implicit commit like DDL normally
does.

The partitions are named pN.

Otherwise the table behaves exactly the same as a normal RANGE COLUMNS
partitioned table.
Hemant Dangi
High priority MDL_BACKUP_COMMIT_HIGH_PRIO and Slave-side BACKUP-COMMIT
tracking to prevent parallel-slave/backup deadlock

A parallel-slave worker holding MDL_BACKUP_COMMIT and waiting on a
prior commit can deadlock with an incoming BACKUP STAGE BLOCK_COMMIT:
the prior commit also needs MDL_BACKUP_COMMIT but fairness denies it
while the backup X-request is in the waiting queue.

Add MDL_BACKUP_COMMIT_HIGH_PRIO — a sibling of MDL_BACKUP_COMMIT
whose only matrix difference is that it is granted past a waiting
backup X. Compatibility with already-granted holders is unchanged.

The slave decides at request time, via per-Relay_log_info state in
Backup_commit_team (every in-flight teammate plus the subset that
picked C but is not yet granted). The atomic join_and_decide() rule:

Pick HIGH_PRIO iff
  (a) some larger-sub_id teammate is registered, AND
  (b) no smaller-sub_id teammate is currently pending-C.
Otherwise pick MDL_BACKUP_COMMIT and mark pending-C.

(a) breaks the deadlock. (b) prevents a larger sub_id from being
granted H while a smaller is queued as C behind X — which would
re-deadlock through wait_for_prior_commit. Atomicity closes the
register/decide race. mark_granted() clears pending-C on grant;
leave() cleans up on release or error.
Yuchen Pei
MDEV-15621 [demo] Allow PARTITION BY RANGE COLUMNS of a timestamp column
Vladislav Vaintroub
improve MTR bootstrap performance on slow disk

use --debug-no-sync to workaround slow disk flush.
On Github actions Windows ARM64, this alone cuts the test time from
ca 55 min to ca 45 min.
Georgi (Joro) Kodinov
MDEV-39718: Produce Markdown plugin API documentation

Implemented a cmake utility macro to generate markdown documentation
for the plugin API headers.
Fixed some doxygen comment mistakes in these.
Yuchen Pei
MDEV-15621 Auto add RANGE COLUMNS partitions by interval

Allow auto partitioning by interval in PARTITION BY RANGE COLUMNS

PARTITION BY RANGE COLUMNS (col_name)
INTERVAL interval [AUTO]
(
  PARTITION partition_name VALUES LESS THAN (value)
  [, PARTITION partition_name VALUES LESS THAN (value) ... ]
)

where

- col_name is the name of one column of type DATE or DATETIME
- at least one partition is supplied
- INTERVAL interval is a time interval

When performing DML on such a table, it will first add partitions
by the specified interval until the partition covers the current time.

Partition addition will not cause an implicit commit like DDL normally
does.

The partitions are named pN.

Otherwise the table behaves exactly the same as a normal RANGE COLUMNS
partitioned table.
Yuchen Pei
MDEV-15621 [refactor] Partitioning cleanup

change p_column_list_val::fixed to a bool
remove redundant end label in partition_info::fix_column_value_functions
sjaakola
MDEV-38260 applier hang due to sequence table access

Avoiding to mark sequence table write as a DDL transaction
for galera applying.

Skipping commit for DDL marked GTID log event in applying as this
would lead to double commit, if the transaction has also a real
transaction. Such scnenario could happen if transaction has sequence
table writes together with other innodb table writes.
Rex Johnston
MDEV-39492 Parallel Query: Study how to create worker threads

Introduces parallel_worker_threads variable to control the number
of worker threads created by a parallel execution query.

2 new files, sql_parallel_workers.h sql_parallel_workers.cc which
contain structures for the creation, management and deletion of
parallel worker threads (pwt_ in the name).  Main management
class created in the stack in JOIN::exec, implemented for the
top level select.

Current parallel_worker_thread_func sleeps for 10 seconds, generates
a warning, signals the main thread, sleeps 10 seconds, signals the
main thread again, sets it's finished flag and cleans it's THD.

The main thread loops through worker threads, looking for finished
thread and cleans them up if they have finished.
It then waits for a signal, then processes it's message queue.

Threads are registed in server_threads, so are visible in
information_schema.processlist and the show processlist command.

We check that a kill query on a parallel worker is passed onto it's
manager and the query is properly aborted, and that a kill connection
is handled properly in parallel_worker.test.

Review input 1: cleanup earlier

Do cleanup before we've finished sending the result to the client.
This way, one can see the errors (and eventually warnings) marshalled
back to the main thread and returned to the user:

MariaDB [test]> set parallel_worker_threads=10;
Query OK, 0 rows affected (0.001 sec)

MariaDB [test]> select seq from seq_1_to_10;
ERROR 4103 (HY000): Argument to the worker_busted_function() function does not belong to the range [0,1]

Assisted by Sergei Petrunia and Claude Code.
bsrikanth-mariadb
MDEV-39538: Different costs when same range is read twice

When same range is used as a filter, once in the outer query block, and
the second inside a sub query such as: -

select * from t1
  where year(a) = 2010 and c < (select count(*) from t1 where year(a) = 2010);

The Optimizer Context had two records for multi_range_read_info_const() call,
but had different cost vector members.
The cause is that the first table considered index-only scan on the range,
while the second considered non-index only scan.

However, when replaying the context, the same range got matched
twice, and the costs corresponding to that got returned twice.
Hence the costs in the explain plan output differed as well.

Solution
========
Include a new field called "call_number", while recording range contexts
into the overall context. This way, we could even match the call_number
and return the appropriate cost during replay.
Marko Mäkelä
Re-enable 3 tests
Marko Mäkelä
fixup! 2a7950aa1cfdd4bb212eec4317b7bea4e24c830b

FIXME: Sometimes we have both a hard-linked ib_*.log and ib_logfile101
with identical contents.
PranavKTiwari
MDEV-28610 : Assertion  marked_for_read() failed upon range select with virtual column in index.

Fix read_set propagation for virtual columns during keyread
The keyread initialization path was populating  using , which internally uses the non-recursive index marking path. This only marks direct index fields and does not resolve dependencies of virtual/generated columns.
As a result, when a covering index contained a virtual column, its dependent base columns could be missing from , leading to  assertions during virtual column evaluation.
Replace the call with , which uses the recursive read-marking path and correctly propagates virtual/generated column dependencies into  during keyread setup.
Rex Johnston
MDEV-39492 Parallel Query: Study how to create worker threads

Introduces parallel_worker_threads variable to control the number
of worker threads created by a parallel execution query.

2 new files, sql_parallel_workers.h sql_parallel_workers.cc which
contain structures for the creation, management and deletion of
parallel worker threads (pwt_ in the name).  Main management
class created in the stack in JOIN::exec, implemented for the
top level select.

Current parallel_worker_thread_func sleeps for 10 seconds, generates
a warning, signals the main thread, sleeps 10 seconds, signals the
main thread again, sets it's finished flag and cleans it's THD.

The main thread loops through worker threads, looking for finished
thread and cleans them up if they have finished.
It then waits for a signal, then processes it's message queue.

Threads are registed in server_threads, so are visible in
information_schema.processlist and the show processlist command.

We check that a kill query on a parallel worker is passed onto it's
manager and the query is properly aborted, and that a kill connection
is handled properly in parallel_worker.test.

Review input 1: cleanup earlier

Do cleanup before we've finished sending the result to the client.
This way, one can see the errors (and eventually warnings) marshalled
back to the main thread and returned to the user:

MariaDB [test]> set parallel_worker_threads=10;
Query OK, 0 rows affected (0.001 sec)

MariaDB [test]> select seq from seq_1_to_10;
ERROR 4103 (HY000): Argument to the worker_busted_function() function does not belong to the range [0,1]

Assisted by Sergei Petrunia and Claude Code.
Yuchen Pei
MDEV-15621 Auto add RANGE COLUMNS partitions by interval

Allow auto partitioning by interval in PARTITION BY RANGE COLUMNS

PARTITION BY RANGE COLUMNS (col_name)
INTERVAL interval [AUTO]
(
  PARTITION partition_name VALUES LESS THAN (value)
  [, PARTITION partition_name VALUES LESS THAN (value) ... ]
)

where

- col_name is the name of one column of type DATE or DATETIME or
  TIMESTAMP
- at least one partition is supplied
- INTERVAL interval is a time interval

When performing DML on such a table, it will first add partitions
by the specified interval until the partition covers the current time.

Partition addition will not cause an implicit commit like DDL normally
does.

The partitions are named pN.

Otherwise the table behaves exactly the same as a normal RANGE COLUMNS
partitioned table.

Note that TIMESTAMP is not allowed as a type for PARTITION BY RANGE
COLUMNS otherwise.
Vladislav Vaintroub
MDEV-39719 Fix memory allocation errors on Windows ARM64 CI

Do not overload the machine. Huge reservation seem to stress
Windows on ARM64 memory management, so that random OOM can appear
when multiple processes are constantly allocating/deallocating
memory. Reduce number of parallel threads to 2. The  bottleneck
on this machine is disk IO, so it does not reduce total test time.
Vladislav Vaintroub
improve MTR bootstrap performance on slow disk

use --debug-no-sync to workaround slow disk flush.
On Github actions Windows ARM64, this alone cuts the test time from
ca 55 min to ca 45 min.
Vladislav Vaintroub
x
PranavKTiwari
Added code change.
Alexey Botchkov
Progress on 26.05.26.
Andrei Elkin
MDEV-36025 Parallel slave concurrent BACKUP MDL locking with ongoing backup..

The problem the patch tackles is that the binlogging parallel slave
can expose prepared state of transactions to the backup process.
That is the latter can enroll such transactions, typically waiting for
prior commits - in other words *undecided* (of whether they are going
to commit at all - sic!), into backup image.

The technical possibility of that owes to earlier deadlocks fixes in this area
that made the parallel slave worker to re-acquire its BACKUP MDL within
the wait-for-prior commit phase (which is post- engine prepare) one.

Note the --skip-log-bin or --log-slave-update=OFF parallel slave is
not vulnerable to the exposure of its prepared transactions.

The fixes this patch provides make sure the backup image contains only
transactions that are (binlog-order) committed, while
parallel slave does not deadlock (MDEV-23586).

The principal part of the fixes implements leapfrogging of a waiting
"high-priority" BACKUP MDL by a slave parallel worker which is
exemplified in the following. First mind about notations:

  here (as elsewhere in the patch) `Fn` group commit followers
  indexed by the gtid seq-no `n`; `F|xyz` indicates the execution
  status (timing) of the MDL request.  `B` - stands for backup
  thread; `Si,Xk` are BACKUP share and exclusive MDL sequenced by
  `i,k` that are logical times of the MDL acquisition; The
  semi-column separates the granted head of the lock queue from the
  waiters.

The queue below is the BACKUP MDL queue.

| which initially is
|
| [Initial Queue State]
| GRANTED            ; WAITING QUEUE
| ----------------------------------
| S1(F2)            ; X2(B)
|
| Next it receives
|                    <- Incoming Request: S3(L1)
| the scheduler having the following context
| [Scheduler Evaluation]
| Requestor: L1 (gtid_sub_id = 1)
| Holder  : F2 (gtid_sub_id = 2)
| Condition: holder->gtid_sub_id > requestor->gtid_sub_id  (2 > 1) -> TRUE!
|  Action  : ignore_mdl_priority = true. L1 bypasses the waiting X2 lock.
|
| decides where to place it
|
| [Final Queue State]
| GRANTED            ; WAITING QUEUE
| ----------------------------------
| S1(F2), S3(L1)    ; X2(B)

| so it chooses to team up L1,F2 correctly. Any next S4(F_n), where n > 2
| won't be granted which means F_n will wait for the backup completion.
| Like this
|
| [Final Queue State]
| GRANTED            ; WAITING QUEUE
| ----------------------------------
| S1(F2), S3(L1)    ; X2(B),S4(F3)

The decision to allow S3 leapfrog X2 requires the current lock holder
F2 to have a greater gtid_sub_id (2 > 1).

To implement this scheduling policy requires the following modifications.

P1 sql/mdl.h
    The new method determines whether a BACKUP lock requester is a teammate
    of a group that can share the MDL lock and that there exists at least
    on granted member of the group.

    +MDL_request::bool (*is_teammate_callback)(const THD*, const THD*);

P2. sql/log.cc
    dismantling of mdl_context.{release,acquire}_lock() in the parallel worker
    wait-for-prior commit

P3. sql/handler.cc

    setting the P1 MDL_request::is_teammate_callback to the slave parallel worker
    as hint to try acquiring the MDL lock by team membership.

P4. sql/rpl_rli.cc
    defines the teammate callback for the parallel slave.

    +rpl_group_info::ignore_mdl_priority

P5. sql/handler.cc
    Logics of "careful" release of the MDL lock could not be streamlined.
    The failed parallel slave worker still has to abide with the former policy of
    ``` as there is extra replication
    book-keeping to be done before rolling back and allowing a conflicting
    transaction to continue (MDEV-7458).``` [ha_commit_trans].

    For that reason of MDEV-7458 the BACKUP MDL request's memory is now allocated
    for the worker in THD::st_transaction::mem_root. Also guards are deployed to not let
    the mem-root be be cleaned too early, before the lock gets released. It can be
    released fast to follow with transaction->cleanup() for
    not failing trx:s.
    Failing to commit parallel workers defer that to Relay_log_info::cleanup_context()
    from where now ha_rollback_trans() will make it.

Aslo necessary changes are caused by to the gtid implicit statement's design.

1. sql/sql_class.cc
  Removed earlier backup-on-parallel-slave bugfixes of MDEV-23586.
  The parallel worker does not release anymore BACKUP MDL at its wait-for-prior
  commit stage.

2. sql/sql_lex.h, sql/sql_base.cc, sql/rpl_gtid.cc
  Has to be introduced
    +#define TL_OPTION_GTID_TABLE_SLAVE      64

  as a part of a method to find out (see open_table() hunk) at
  executing record_gtid() by implicit GTID statement that its
  BACKUP MDL lock (however it is necessary - it is not challenged
  here) needs the ignore-priority hint. Note this hint applies for granting
  a "special" S' share locked not yet the commit time  ha_commit_trans()'s
  MDL_BACKUP_COMMIT (denoted as S) lock, which is going to be requested later.

  Let's exemplify it on the following diagrams, calling S' a sort of
  shared lock, compatible with with S of the commit time.

  *Without* the teammate hint the request for S'3(F2) at record_gtid() time

        S1(F3|wait_for_prior_commit); X2(B) <- S'3(F2|record_gtid)

  would end up to wait behind X2

        S1; /* waiting */ X2, S'3

  and we'd be regress back to hang/deadlock of earlier bugs: F3 would not be
  awaken by F2 who is blocked by B.

Tested with extended set of parallel slave with backup scenarios.

This commit must require more and extensive testing and may cause
followup amendments, through review comments as well.

Co-authored-with: [email protected].
Thirunarayanan Balathandayuthapani
MDEV-34358  Encryption threads consume CPU when no work available

Problem:
========
1. Encryption threads busy-wait when no work is available. When reaching
fil_system.space_list.end(), fil_crypt_return_iops() is called with
wake=true, causing pthread_cond_broadcast() to wake all threads
unnecessarily, leading to CPU waste.

2. Tablespaces with CLOSING/STOPPING flags (set during DDL operations)
are skipped during iteration. Since DDL completion doesn't wake
encryption threads, these spaces may never be encrypted if threads
sleep indefinitely.

3. For default_encrypt_list iteration, when spaces exist but none are
acquirable, threads need to wake others for cooperative retry, but
this case was not distinguished from fil_system.space_list.end().

4. IOPS are allocated before searching for tablespaces, wasting resources
during iteration when no I/O occurs.

5. Encryption threads use fil_crypt_threads_cond for two different purposes:
waiting for encryption work and waiting for IOPS allocation. When
fil_crypt_return_iops() or fil_crypt_realloc_iops() broadcasts after
releasing IOPS, it wakes ALL threads including those correctly waiting
for work, causing spurious wakeups and CPU waste.

6. When innodb_encrypt_tables or innodb_encryption_rotate_key_age is
changed during encryption thread iteration, threads continue with
stale configuration values, potentially missing tablespaces that
should be encrypted or rotated under the new settings.


Solution:
=========
1. Implement timed wait with exponential backoff when space ==
fil_system.space_list.end() (applies to both default_encrypt_list and
space_list iteration when no acquirable spaces are found):
- First timeout: 5 seconds
- Subsequent timeouts: (timed_wait_count + 1) * 5 seconds
  (10s, 20s, 40s, 60s)
- After 5 consecutive timeouts (~135 seconds total), continue with
60-second timed waits to ensure threads periodically recheck for
state changes without requiring explicit signals
- Timeout counter resets to 0 when woken by signal (config change,
new tablespace) or when work is found

2. Move IOPS allocation from before tablespace search to after finding a
space that needs rotation. If allocation fails, set recheck=true to
skip waiting and immediately try next space.

3. Introduce separate condition variable fil_crypt_iops_cond specifically
for IOPS allocation synchronization to prevent spurious wakeups:

- fil_crypt_threads_cond: Used in wait_for_work() for waiting when no
tablespaces need encryption. Signaled when settings change, new
tablespaces are created, or thread count changes.

- fil_crypt_iops_cond (NEW): Used in fil_crypt_alloc_iops() for waiting
when IOPS limit is reached. Signaled when IOPS are returned via
fil_crypt_return_iops(), released via fil_crypt_realloc_iops(), or
when srv_n_fil_crypt_iops is increased.

4. Added atomic version counter fil_crypt_settings_version that is
incremented whenever innodb_encrypt_tables or
innodb_encryption_rotate_key_age changes. Encryption threads capture
the version at iteration start and check for changes during iteration.
If config changed, threads immediately restart iteration from the
beginning to ensure complete coverage with new settings.

fil_crypt_settings_version: Atomic counter to track the
innodb_encrypt_tables or innodb_encryption_rotate_key_age changes

rotate_thread_t::timed_wait_count (uint8_t):
Counts consecutive timeouts for exponential backoff

rotate_thread_t::wait_for_work(): Implements timed/indefinite
wait strategy

rotate_thread_t::settings_version: To compare the
existing fil_crypt_settings_version to restart the
encryption from the beginning
Daniel Black
MDEV-37664: main.lotofstack fails with MSAN+Debug

Under MSAN+Debug this test case exceeds the stack
early without the stack detection mechanism to
identify that it passed its limit.
Raghunandan Bhat
MDEV-39227: Schema Versioning - Server-side

- SQL layer infrastructure for multiple versions of a table definition.
- No storage engine changes; any DDLs that depends on the engine or .frm
  re-write for persistance is not yet supported.
- binlog/replication, DDL log for crash recovery are not supported yet.
PranavKTiwari
MDEV-28610 : Assertion  marked_for_read() failed upon range select with virtual column in index.
Fix read_set propagation for virtual columns during keyread

The keyread initialization path was populating  using , which internally uses the non-recursive index marking path. This only marks direct index fields and does not resolve dependencies of virtual/generated columns.

As a result, when a covering index contained a virtual column, its dependent base columns could be missing from , leading to  assertions during virtual column evaluation.

Replace the call with , which uses the recursive read-marking path and correctly propagates virtual/generated column dependencies into  during keyread setup.
Vladislav Vaintroub
MDEV-39719 Fix memory allocation errors on Windows ARM64 CI

Do not overload the machine. Huge reservation seem to stress
Windows on ARM64 memory management, so that random OOM can appear
when multiple processes are constantly allocating/deallocating
memory. Reduce number of parallel threads to 2. The  bottleneck
on this machine is disk IO, so it does not reduce total test time.
Marko Mäkelä
Implement compatible copy_file() on macOS
Thirunarayanan Balathandayuthapani
MDEV-34358  Encryption threads consume CPU when no work available

Problem:
========
The InnoDB encryption thread could deadlock with
purge or DML operations when rotating encryption keys.
The encryption thread called fseg_page_is_allocated()
to check if a page was allocated before encrypting it.
This function acquired an exclusive latch (fil_space_t::latch)
on the tablespace to read the allocation bitmap.

However, purge or DML threads could already be holding page
latches while waiting to acquire the same tablespace latch,
creating a lock order violation. Specifically, the encryption
thread would hold the tablespace latch and then try to acquire a
page latch, while purge held a page latch and tried to
acquire the tablespace latch, resulting in a deadlock.

Solution:
=========
fseg_page_is_allocated(): Encryption thread passes an mtr,
allocation bitmap page latch is now correctly held in that mtr
until the caller commits it.

fil_crypt_get_page_throttle(): Use a trylock mechanism to acquire
the tablespace exclusive latch non-blockingly. If the trylock fails,
back off and retry the same page to avoid deadlock.  The bitmap
page S-latch is added to the mtr and held throughout the page read
and encryption operations. Release the tablespace exclusive latch
immediately after the allocation check to minimize contention,
while the bitmap page S-latch remains held in the mtr.
Georgi (Joro) Kodinov
MDEV-39718: Produce Markdown plugin API documentation

Implemented a cmake utility macro to generate markdown documentation
for the plugin API headers.
Fixed some doxygen comment mistakes in these.
Andrei Elkin
High priority MDL_BACKUP_BLOCK_COMMIT_RPL

When the standard request of MDL_BACKUP_COMMIT is not granted to a worker
it first searches for a possible grantee that is a parallel slave
worker and if there exist one with greater commit order the request
type is escalate to the highest which is re-submitted.

To the changes in sql_base.cc,
note there also exist a request of the highest (new) type at time of
Xid_apply_log_event::do_record_gtid which is a committing phase of
slave transaction when it is executing an implicit statement.
It must be able to bypass any waiting lock by backup process even
without yet knowing whether the transaction is going to be enrolled
into a pending to commit slave group. When later it turns out
not the case the strong MDL_STATEMENT requested lock would've been
already relinquished.

Tested with previous commit's  refined and extended BASE's
  rpl.parallel_backup, rpl.rpl_parallel_backup_worker_retry
and new added
  rpl.rpl_parallel_backup_waits_worker_retry

Todo:
According to Svoj's review, this patch still needs
removing the slave specific code from the MDL.
That is changes in MDL_context::acquire_lock().

Probably
ha_commit_trans() would call first
    try_acquire_lock_impl()
when it returns empty ticket a part of the added to acquire_lock
Teammate-Aware Escalation Check would be invoked just to compute
`has_teammate` and when that is true to call the whole BASE version
of MDL_context::acquire_lock() having the highest type of request.
Alexey Botchkov
fix to the XMLTYPE.
Vladislav Vaintroub
MDEV-39719 Fix memory allocation errors on Windows ARM64 CI

Do not overload the machine. Huge reservation seem to stress
Windows on ARM64 memory management, so that random OOM can appear
when multiple processes are constantly allocating/deallocating
huge chunks of memory. Thus, reduce number of parallel threads to 2,
and do not run replication tests.
PranavKTiwari
removed inlin.
Marko Mäkelä
InnoDB_backup::fini(): Trim to the same log file size
Alexey Botchkov
Progress report.
        17.04.2026
Alexander Barkov
MDEV-39518 Allow prepared statements in stored functions in assignment right hand

In progress
bsrikanth-mariadb
MDEV-39538: Different costs when same range is read twice

When same range is used as a filter, once in the outer query block, and
the second inside a sub query such as: -

select * from t1
  where year(a) = 2010 and c < (select count(*) from t1 where year(a) = 2010);

The Optimizer Context had two records for multi_range_read_info_const() call,
but had different cost vector members.
The cause is that the first table considered index-only scan on the range,
while the second considered non-index only scan.

However, when replaying the context, the same range got matched
twice, and the costs corresponding to that got returned twice.
Hence the costs in the explain plan output differed as well.

Solution
========
Include a new field called "call_number", while recording range contexts
into the overall context. This way, we could even match the call_number
and return the appropriate cost during replay.
Vladislav Vaintroub
MDEV-39719 Fix memory allocation errors on Windows ARM64 CI

Do not overload the machine. Huge reservation seem to stress
Windows on ARM64 memory management, so that random OOM can appear
when multiple processes are constantly allocating/deallocating
memory. Reduce number of parallel threads to 2. The  bottleneck
on this machine is disk IO, so it does not reduce total test time.
Rex Johnston
MDEV-39492 Parallel Query: Study how to create worker threads

Introduces parallel_worker_threads variable to control the number
of worker threads created by a parallel execution query.

2 new files, sql_parallel_workers.h sql_parallel_workers.cc which
contain structures for the creation, management and deletion of
parallel worker threads (pwt_ in the name).  Main management
class created in the stack in JOIN::exec, implemented for the
top level select.

Current parallel_worker_thread_func sleeps for 10 seconds, generates
a warning, signals the main thread, sleeps 10 seconds, signals the
main thread again, sets it's finished flag and cleans it's THD.

The main thread loops through worker threads, looking for finished
thread and cleans them up if they have finished.
It then waits for a signal, then processes it's message queue.

Threads are registed in server_threads, so are visible in
information_schema.processlist and the show processlist command.

We check that a kill query on a parallel worker is passed onto it's
manager and the query is properly aborted, and that a kill connection
is handled properly in parallel_worker.test.

Review input 1: cleanup earlier

Do cleanup before we've finished sending the result to the client.
This way, one can see the errors (and eventually warnings) marshalled
back to the main thread and returned to the user:

MariaDB [test]> set parallel_worker_threads=10;
Query OK, 0 rows affected (0.001 sec)

MariaDB [test]> select seq from seq_1_to_10;
ERROR 4103 (HY000): Argument to the worker_busted_function() function does not belong to the range [0,1]

Assisted by Sergei Petrunia and Claude Code.
Marko Mäkelä
Fix a regression in mtr_t::is_named_space