The immediate bug was caused by a failure to recognize a correct
position to stop the slave applier run in optimistic parallel mode.
There were the following set of issues that the analysis unveil.
1 incorrect estimate for the event binlog position passed to
is_until_satisfied
2 wait for workers to complete by the driver thread did not account non-group events
that could be left unprocessed and thus to mix up the last executed
binlog group's file and position:
the file remained old and the position related to the new rotated file
3 incorrect 'slave reached file:pos' by the parallel slave report in the error log
4 relay log UNTIL missed out the parallel slave branch in
is_until_satisfied.
The patch addresses all of them to simplify logics of log change
notification in either the master and relay-log until case.
P.1 is addressed with passing the event into is_until_satisfied()
for proper analisis by the function.
P.2 is fixed by changes in handle_queued_pos_update().
P.4 required removing relay-log change notification by workers.
Instead the driver thread updates the notion of the current relay-log
fully itself with aid of introduced
bool Relay_log_info::until_relay_log_names_defer.
An extra print out of the requested until file:pos is arranged
with --log-warning=3.
The immediate bug was caused by a failure to recognize a correct
position to stop the slave applier run in optimistic parallel mode.
There were the following set of issues that the analysis unveil.
1 incorrect estimate for the event binlog position passed to
is_until_satisfied
2 wait for workers to complete by the driver thread did not account non-group events
that could be left unprocessed and thus to mix up the last executed
binlog group's file and position:
the file remained old and the position related to the new rotated file
3 incorrect 'slave reached file:pos' by the parallel slave report in the error log
4 relay log UNTIL missed out the parallel slave branch in
is_until_satisfied.
The patch addresses all of them to simplify logics of log change
notification in either the master and relay-log until case.
P.1 is addressed with passing the event into is_until_satisfied()
for proper analisis by the function.
P.2 is fixed by changes in handle_queued_pos_update().
P.4 required removing relay-log change notification by workers.
Instead the driver thread updates the notion of the current relay-log
fully itself with aid of introduced
bool Relay_log_info::until_relay_log_names_defer.
An extra print out of the requested until file:pos is arranged
with --log-warning=3.
e.g.
- dont -> don't
- occurence -> occurrence
- succesfully -> successfully
- easyly -> easily
Also remove trailing space in selected files.
These changes span:
- server core
- Connect and Innobase storage engine code
- OQgraph, Sphinx and TokuDB storage engines
Related to MDEV-21769.
Lifted long standing limitation to the XA of rolling it back at the
transaction's
connection close even if the XA is prepared.
Prepared XA-transaction is made to sustain connection close or server
restart.
The patch consists of
- binary logging extension to write prepared XA part of
transaction signified with
its XID in a new XA_prepare_log_event. The concusion part -
with Commit or Rollback decision - is logged separately as
Query_log_event.
That is in the binlog the XA consists of two separate group of
events.
That makes the whole XA possibly interweaving in binlog with
other XA:s or regular transaction but with no harm to
replication and data consistency.
Gtid_log_event receives two more flags to identify which of the
two XA phases of the transaction it represents. With either flag
set also XID info is added to the event.
When binlog is ON on the server XID::formatID is
constrained to 4 bytes.
- engines are made aware of the server policy to keep up user
prepared XA:s so they (Innodb, rocksdb) don't roll them back
anymore at their disconnect methods.
- slave applier is refined to cope with two phase logged XA:s
including parallel modes of execution.
This patch does not address crash-safe logging of the new events which
is being addressed by MDEV-21469.
CORNER CASES: read-only, pure myisam, binlog-*, @@skip_log_bin, etc
Are addressed along the following policies.
1. The read-only at reconnect marks XID to fail for future
completion with ER_XA_RBROLLBACK.
2. binlog-* filtered XA when it changes engine data is regarded as
loggable even when nothing got cached for binlog. An empty
XA-prepare group is recorded. Consequent Commit-or-Rollback
succeeds in the Engine(s) as well as recorded into binlog.
3. The same applies to the non-transactional engine XA.
4. @@skip_log_bin=OFF does not record anything at XA-prepare
(obviously), but the completion event is recorded into binlog to
admit inconsistency with slave.
The following actions are taken by the patch.
At XA-prepare:
when empty binlog cache - don't do anything to binlog if RO,
otherwise write empty XA_prepare (assert(binlog-filter case)).
At Disconnect:
when Prepared && RO (=> no binlogging was done)
set Xid_cache_element::error := ER_XA_RBROLLBACK
*keep* XID in the cache, and rollback the transaction.
At XA-"complete":
Discover the error, if any don't binlog the "complete",
return the error to the user.
Kudos
-----
Alexey Botchkov took to drive this work initially.
Sergei Golubchik, Sergei Petrunja, Marko Mäkelä provided a number of
good recommendations.
Sergei Voitovich made a magnificent review and improvements to the code.
They all deserve a bunch of thanks for making this work done!
Instrumenting parallel slave worker thread with wsrep replication hooks.
Added mtr test for testing parallel slave support.
The test is based on the test attached in MDEV-6860 jira tracker.
Analysis:
========
In general if there are three groups.
1 - Inserts 32 which fails due to local entry '32' on slave.
2 - Inserts 33
3 - Inserts 34
Each group considers itself as a waiter and it waits for prior group 'waitee'.
This is done in 'register_wait_for_prior_event_group_commit'. If there is no
other parallel group being scheduled then no waitee will be there.
Let us assume 3 groups are being scheduled in parallel.
3-> waits for 2-> waits for->1
'1' upon completion it checks is there any registered subsequent waiter. If
so it wakes up the subsequent waiter with its execution status. This execution
status is stored in wakeup_error.
If '1' failed then it sends corresponding wakeup_error to 2. Then '2' aborts
and it propagates error to '3'. So all further commits are aborted. This
mechanism works only when all transactions reach a stage where they are
waiting for their prior commit to complete.
In case of optimistic following scenario occurs.
1,2,3 are scheduled in parallel.
3 - Reaches group_commit_code waits for 2 to complete.
1 - errors out sets stop_on_error_sub_id=1.
When a group execution results in error its corresponding sub_id is set to
'stop_on_error_sub_id'. Any new groups queued for execution will check if
their sub_id is > stop_on_error_sub_id. If it is true their execution will be
skipped as prior group execution failed. 'skip_event_group=1' will be set.
Since the execution of SQL thread is about to stop we just skip execution of
all the following event groups. We still do all the normal waiting and wakeup
processing between the event groups as a simple way to ensure that everything
is stopped and cleaned up correctly.
Upon error '1' transaction checks for registered waiters. Since no one is
there it simply goes away.
2 - Starts the execution. It checks do I have a waitee.
Since wait_commit_sub_id == entry->last_committed_sub_id no waitee is set.
Secondly: 'entry->stop_on_error_sub_id' is set by '1'st execution. Now
'handle_parallel_thread' code checks if the current group 'sub_id' is greater
than the 'sub_id' set within 'stop_on_error_sub_id'.
Since the above is true 'skip_event_group=true' is set. Simply call
'wait_for_prior_commit' to wakeup all waiters. Group '2' didn't had any
waitee and its execution is skipped. Hence its wakeup_error=0.It sends a
positive wakeup signal to '3'. Which commits. This results in a missed
transaction. i.e 33 is missed and 34 is committed.
Fix:
===
When a worker learns that an earlier transaction execution has failed, and it
should not proceed for further execution, it should mark its own execution
status as failed so that it alerts its followers to abort as well.
Implemented and integrated THD_list as a replacement for the global
thread list. It uses own mutex instead of LOCK_thread_count for THD
list protection.
Removed unused first_global_thread() and next_global_thread().
delayed_insert_threads is now protected by LOCK_delayed_insert. Although
this patch doesn't fix very wrong synchronization of this variable.
After this patch there are only 2 legitimate uses of LOCK_thread_count
left, both in mysqld.cc: thread_count and ready_to_exit.
Aim is to reduce usage of LOCK_thread_count and COND_thread_count.
Part of MDEV-15135.
The bug appears as a slave SQL thread hanging in
rpl_parallel_thread_pool::get_thread() while there are no slave worker
threads to awake it.
The reason of the hang is that at the parallel slave worker pool
activation the being stared SQL thread could read the worker pool size
concurrently with pool deactivation. At reading the SQL thread did not
employ necessary protection from a race.
Fixed with making the SQL thread at the pool activation first
to grab the same lock as potential deactivator also does prior
to access the pool size.
Fixed by deleting the sequence if we where not able to initialize it
I also noticed that we didn't always set the error message when
check_killed(), which could lead to aborted queries without error
beeing properly set. Fixed by default setting error message if
check_error() noticed that killed had been called.
This allowed me to remove a lot of calls to thd->send_kill_message().
Introduced new alter algorithm type called NOCOPY & INSTANT for
inplace alter operation.
NOCOPY - Algorithm refuses any alter operation that would
rebuild the clustered index. It is a subset of INPLACE algorithm.
INSTANT - Algorithm allow any alter operation that would
modify only meta data. It is a subset of NOCOPY algorithm.
Introduce new variable called alter_algorithm. The values are
DEFAULT(0), COPY(1), INPLACE(2), NOCOPY(3), INSTANT(4)
Message to deprecate old_alter_table variable and make it alias
for alter_algorithm variable.
alter_algorithm variable for slave is always set to default.
out of order at retry
The test failures were of two sorts. One is that the number of retries
what the slave thought as a temporary error exceeded
the default value of the slave retry option.
The 2nd issue was an out of order commit by transactions that
were supposed to error out instead.
Both issues are caused by the same reason that the post-temporary-error
retry did not check possibly already existing error status.
This is mended with refining conditions to retry. Specifically, a retrying
worker checks `rpl_parallel_entry::stop_on_error_sub_id` that
a potential failing predecessor could set to its own sub id.
Now should the member be set the retrying follower errors out with
ER_PRIOR_COMMIT_FAILED.
This was done in, among other things:
- thd->db and thd->db_length
- TABLE_LIST tablename, db, alias and schema_name
- Audit plugin database name
- lex->db
- All db and table names in Alter_table_ctx
- st_select_lex db
Other things:
- Changed a lot of functions to take const LEX_CSTRING* as argument
for db, table_name and alias. See init_one_table() as an example.
- Changed some function arguments from LEX_CSTRING to const LEX_CSTRING
- Changed some lists from LEX_STRING to LEX_CSTRING
- threads_mysql.result changed because process list_db wasn't always
correctly updated
- New append_identifier() function that takes LEX_CSTRING* as arguments
- Added new element tmp_buff to Alter_table_ctx to separate temp name
handling from temporary space
- Ensure we store the length after my_casedn_str() of table/db names
- Removed not used version of rename_table_in_stat_tables()
- Changed Natural_join_column::table_name and db_name() to never return
NULL (used for print)
- thd->get_db() now returns db as a printable string (thd->db.str or "")
Assertion `count > 0' failed in rpl_parallel_thread_pool::
get_thread, rpl.rpl_parallel failed in buildbot
The reason for this is that one thread can call
rpl_parallel_resize_pool_if_no_slaves() while
another thread calls at the same time
rpl_parallel_activate_pool(). If rpl_parallel_active_pool() is
called before rpl_parallel_resize_pool_if_no_slaves() has
finished, pool->count will be set to 0 even if there exists
active slave threads.
Added a mutex lock in rpl_parallel_activate_pool() to protect against this scenario, which seams to fix this issue.
Part of MDEV-13073 AliSQL Optimize performance of semisync
The idea it to use a dedicated lock detecting if there is new data in
the master's binary log instead of the overused LOCK_log.
Changes:
- Use dedicated COND variables for the relay and binary log signaling.
This was needed as we where the old 'update_cond' variable was used
with different mutex's, which could cause deadlocks.
- Relay log uses now COND_relay_log_updated and LOCK_log
- Binary log uses now COND_bin_log_updated and LOCK_binlog_end_pos
- Renamed signal_cnt to relay_signal_cnt (as we now have two signals)
- Added some missing error handling in MYSQL_BIN_LOG::new_file_impl()
- Reformatted some comments with old style
- Renamed m_key_LOCK_binlog_end_pos to key_LOCK_binlog_end_pos
- Changed 'signal_update()' to update_binlog_end_pos() which works for
both relay and binary log
This fixes MDEV-7742 and MDEV-8305 (Allow user to specify if stored
procedures should be logged in the slow and general log)
New functionality:
- Added new variables log_slow_disable_statements and log_disable_statements
that can be used to disable logging of certain queries to slow and
general log. Currently supported options are 'admin', 'call', 'slave'
and 'sp'.
Defaults are as before. Only 'sp' (stored procedure statements) is
disabled for slow and general_log.
- Slow log to files now includes the following new information:
- When logging stored procedure statements the name of stored
procedure is logged.
- Number of created tmp_tables, tmp_disk_tables and the space used
by temporary tables.
- When logging 'call', the logged status now contains the sum of all
included statements. Before only 'time' was correct.
- Added filsort_priority_queue as an option for log_slow_filter (this
variable existed before, but was not exposed)
- Added support for BIT types in my_getopt()
Mapped some old variables to bitmaps (old variables can still be used)
- Variable 'log_queries_not_using_indexes' is mapped to
log_slow_filter='not_using_index'
- Variable 'log_slow_slave_statements' is mapped to
log_slow_disabled_statements='slave'
- Variable 'log_slow_admin_statements' is mapped to
log_slow_disabled_statements='admin'
- All the above variables are changed to session variables from global
variables
Other things:
- Simplified LOGGER::log_command. We don't need to check for super if
OPTION_LOG_OFF is set as this flag can only be set if one is a super
user.
- Removed some setting of enable_slow_log as it's guaranteed to be set by
mysql_parse()
- mysql_admin_table() now sets thd->enable_slow_log
- Added prepare_logs_for_admin_command() to reset thd->enable_slow_log if
needed.
- Added new functions to store, restore and add slow query status
- Added new functions to store and restore query start time
- Reorganized Sub_statement_state according to types
- Added code in dispatch_command() to ensure that
thd->reset_for_next_command() is always called for a query.
- Added thd->last_sql_command to simplify checking of what was the type
of the last command. Needed when logging to slow log as lex->sql_command
may have changed before slow logging is called.
- Moved QPLAN_TMP_... to where status for tmp tables are updated
- Added new THD variable, affected_rows, to be able to correctly log
number of affected rows to slow log.
- Added sql/mariadb.h file that should be included first by files in sql
directory, if sql_plugin.h is not used (sql_plugin.h adds SHOW variables
that must be done before my_global.h is included)
- Removed a lot of include my_global.h from include files
- Removed include's of some files that my_global.h automatically includes
- Removed duplicated include's of my_sys.h
- Replaced include my_config.h with my_global.h
The problem was that the introduction of max-thread-mem-used can cause
an allocation error very early, even before mysql_parse() is called.
As mysql_parse() calls thd->reset_for_next_command(), which called
clear_error(), the error number was lost.
Fixed by adding an option to have unique messages for each KILL
signal and change max-thread-mem-used to use this new feature.
This removes a lot of problems with the original approach, where
one could get errors signaled silenty almost any time.
ixed by moving clear_error() from reset_for_next_command() to
do_command(), before any memory allocation for the thread.
Related changes:
- reset_for_next_command() now have an optional parameter if we should
call clear_error() or not. By default it's called, but not anymore from
dispatch_command() which was the original problem.
- Added optional paramater to clear_error() to force calling of
reset_diagnostics_area(). Before clear_error() only called
reset_diagnostics_area() if there was no error, so we normally
called reset_diagnostics_area() twice.
- This change removed several duplicated calls to clear_error()
when starting a query.
- Reset max_mem_used on COM_QUIT, to protect against kill during
quit.
- Use fatal_error() instead of setting is_fatal_error (cleanup)
- Set fatal_error if max_thead_mem_used is signaled.
(Same logic we use for other places where we are out of resources)
Intermediate commit.
Move the discovery of mysql.gtid_slave_pos* tables into the SQL thread.
This avoids doing things like opening tables and scanning the mysql
schema for tables inside of the START SLAVE statement, which might
interact badly with existing transaction or table locks.
(Even though START SLAVE is documented to implicitly commit any active
transactions, this appears not to be the case in current code).
Table discovery fits naturally in the SQL thread init code, next to
the loading of mysql.gtid_slave_pos state.
When master_use_gtid=no, the IO thread loads the slave GTID state from
the master during connect. This races with the SQL thread when
gtid_ignore_duplicates=1. If an event is in the relay log from before
the new connect and has not been applied yet, moving the slave
position causes the SQL thread to think that event should be skipped
due to gtid_ignore_duplicates=1.
This patch simply disables gtid_ignore_duplicates when not using GTID,
which seems to be what one would expect.
The following is an updated commit message for the following commit
that was pushed before I had a chance to update the commit message:
c5e25c8b40
Fixed dead locks when doing stop slave while slave was starting.
- Added a separate lock for protecting start/stop/reset of a specific slave.
This solves some possible dead locks when one calls stop slave while
the slave is starting as the old run_locks was over used for other things.
- Set hash->records to 0 before calling free of all hash elements.
This was set to stop concurrent threads to loop over hash elements and
access members that was already freed.
This was a problem especially in start_all_slaves/stop_all_slaves
as the mutex protecting the hash was temporarily released while a slave
was started/stopped.
- Because of change to hash->records during hash_reset(),
any_slave_sql_running() will return 1 during shutdown as one can't
loop over master_info_index->master_info_hash while hash_reset() of it
is in progress.
This also fixes a potential old bug in any_slave_sql_running() where
during shutdown and ~Master_info_index(), my_hash_free() we could
potentially try to access elements that was already freed.
The problem was that waiting for pause_for_ftwrl was done before
event_group was completed. This caused rpl_pause_for_ftwrl() to wait
forever during FLUSH TABLES WITH READ LOCK.
Now we only wait for FLUSH TABLES WITH READ LOCK when we are changing
to a new event group.
The reason for this is that stop slave takes LOCK_active_mi over the
whole operation while some slave operations will also need LOCK_active_mi
which causes deadlocks.
Fixed by introducing object counting for Master_info and not taking
LOCK_active_mi over stop slave or even stop_all_slaves()
Another benefit of this approach is that it allows:
- Multiple threads can run SHOW SLAVE STATUS at the same time
- START/STOP/RESET/SLAVE STATUS on a slave will not block other slaves
- Simpler interface for handling get_master_info()
- Added some missing unlock of 'log_lock' in error condtions
- Moved rpl_parallel_inactivate_pool(&global_rpl_thread_pool) to end
of stop_slave() to not have to use LOCK_active_mi inside
terminate_slave_threads()
- Changed argument for remove_master_info() to Master_info, as we always
have this available
- Fixed core dump when doing FLUSH TABLES WITH READ LOCK and parallel
replication. Problem was that waiting for pause_for_ftwrl was not done
when deleting rpt->current_owner after a force_abort.
The SQL thread keeps track of the position in the current relay log from
which to read the next event. This position is not normally used, but a
certain interaction with the IO thread can cause the SQL thread to re-open
the relay log and seek to the stored position.
In parallel replication, there were a couple of places where the position
was not updated. This created a race where a re-open of the relay log could
seek to the wrong position and start re-reading and processing events
already handled once, causing various kinds of problems.
Fix this by moving the position update into a single place in
apply_event_and_update_pos(), which should ensure that the position is
always updated in the parallel replication case.
This problem was found from the testcase of MDEV-10863, but it is logically
a separate problem.
Minor review comments/changes:
- A bunch of style-fixes.
- Change macros to static inline functions.
- Update check_event_type() with compressed event types.
- Small .result file update.
Merge feature into 10.2 from feature branch.
Delayed replication adds an option
CHANGE MASTER TO master_delay=<seconds>
Replication will then delay applying events with that many
seconds. This creates a replication slave that reflects the state of
the master some time in the past.
Feature is ported from MySQL source tree.
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
The function apply_event_and_update_pos() is called with the
rli->data_lock mutex held. However, there seems to be nothing in the
function actually needing the mutex to be held. Certainly not in the
parallel replication case, where sql_slave_skip_counter is always 0
since the non-zero case is handled by the SQL driver thread.
So this patch makes parallel replication use a variant of
apply_event_and_update_pos() without the need to take the
rli->data_lock mutex. This avoids one contended global mutex for each
event executed, which might improve performance on CPU-bound workloads
somewhat.
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
When a deadlock kill is detected inside the storage engine, the kill
is not done immediately, to avoid calling back into the storage engine
kill_query method with various lock subsystem mutexes held. Instead the
kill is queued and done later by a slave background thread.
This patch in preparation for fixing TokuDB optimistic parallel
replication, as well as for removing locking hacks in InnoDB/XtraDB in
10.2.
Signed-off-by: Kristian Nielsen <knielsen at knielsen-hq.org>
- When waiting for events, start time is now counted from start of wait
- Instead of having "Connect" as "Command" for all replication threads we
now have:
- Slave_IO for Slave thread reading relay log
- Slave_SQL for slave executing SQL commands or distribution queries to
Slave workers
- Slave_worker for slave threads executin SQL commands in parallel replication
- To ensure that mallocs are marked for the correct THD, even if it's
allocated in another thread, I added the thread_id to the THD constructor
- Added st_my_thread_var to thr_lock_info_init() to avoid a call to my_thread_var
- Moved things from THD::THD() to THD::init()
- Moved some things to THD::cleanup()
- Added THD::free_connection() and THD::reset_for_reuse()
- Added THD to CONNECT::create_thd()
- Added THD::thread_dbug_id and st_my_thread_var->dbug_id. These are needed
to ensure that we have a constant thread_id used for debugging with a THD,
even if it changes thread_id (=connection_id)
- Set variables.pseudo_thread_id in constructor. Removed not needed sets.
- Removed some QQ markers
- Removed some rows not compatible with valgrind 3.9.0
- Made mysql_install_db.sh more silent by default. --verbose now gives more information
- Added assert that auto-increment doesn't generate 0 (safety)
- Removed thd->set_time() in some places as it's set in init_for_queries()
- Fixed some --big tests in tokudb
- Fixed a bug in mysql_client_test.cc where sql_mode was not properly reset
Creating a CONNECT object on client connect and pass this to the working thread which creates the THD.
Split LOCK_thread_count to different mutexes
Added LOCK_thread_start to syncronize threads
Moved most usage of LOCK_thread_count to dedicated functions
Use next_thread_id() instead of thread_id++
Other things:
- Thread id now starts from 1 instead of 2
- Added cast for thread_id as thread id is now of type my_thread_id
- Made THD->host const (To ensure it's not changed)
- Removed some DBUG_PRINT() about entering/exiting mutex as these was already logged by mutex code
- Fixed that aborted_connects and connection_errors_internal are counted in all cases
- Don't take locks for current_linfo when we set it (not needed as it was 0 before)
This includes fixing all utilities to not have any memory leaks,
as safemalloc warnings stopped tests from passing on MacOSX.
- Ensure that all clients takes character-set-dir, as the
libmysqlclient library will use it.
- mysql-test-run now passes character-set-dir to all external clients.
- Changed dynstr_free() so that it can be called twice (made freeing code easier)
- Changed rpl_global_gtid_slave_state to be allocated dynamicly as it
includes a mutex that needs to be initizlied/destroyed before my_end() is called.
- Removed rpl_slave_state::init() and rpl_slave_stage::deinit() as
their job are better handling by constructor and delete.
- Print alias instead of table_name in check_duplicate_key as
table_name may have been converted to lower case.
Other things:
- Fixed a case in time_to_datetime_with_warn() where we where
using && instead of & in tests
Problem was that we used same condition variable with 2 different mutex.
Fixed by changing to use COND_rpl_thread_stop instead of COND_parallel_entry
for stopping threads.
Patch by Kristian Nielsen
Problem is that FLUSH TABLES WITH READ LOCK first blocks threads from
starting new commits, then waits for running commits to complete. But
in-order parallel replication needs commits to happen in a particular
order, so this can easily deadlock.
To fix this problem, this patch introduces a way to temporarily pause
the parallel replication worker threads. Before starting FTWRL, we let
all worker threads complete in-progress transactions, and then
wait. Then we proceed to take the global read lock. Once the lock is
obtained, we unpause the worker threads. Now commits are blocked from
starting by the global read lock, so the deadlock will no longer occur.
Before, the Seconds_behind_master was updated already when an event
was queued for a worker thread to execute later. This might lead users
to interpret a low value as the slave being almost up to date with the
master, while in reality there might still be lots and lots of events
still queued up waiting to be applied by the slave.
See https://lists.launchpad.net/maria-developers/msg08958.html for
more detailed discussions.
The code was using the wrong variable when comparing the binlog name
for the UNTIL position. This could cause the comparison to fail after
binlog rotation, in turn causing the UNTIL clause to not trigger slave
stop.
The assertion is there to catch cases where we rollback while
mark_start_commit() is active. This can allow following event groups
to be replicated too early, causing conflicts.
But in this case, we have an _explicit_ ROLLBACK event in the binlog,
which should not assert.
We fix this by delaying the mark_start_commit() in the explicit
ROLLBACK case. It seems safest to delay this in ROLLBACK case anyway,
and there should be no reason to try to optimise this corner case.
This bug is essentially another variant of MDEV-7458.
If a transaction conflict caused a deadlock kill of T2 in record_gtid()
during commit, the code would do a rollback _before_ running
rgi->unmark_start_commit(). This creates a race where following transactions
could start too early (before T2 has completed its transaction retry). This
in turn could lead to replication failure, if there was a conflict that
caused eg. duplicate key error or similar.
The fix is to remove these rollbacks (in Query_log_event::do_apply_event()
and Xid_log_event::do_apply_event(). They seem out-of-place; code in
log_event.cc generally does not roll back on error, this is handled higher
up.
In addition, because of the extreme difficulty of reproducing bugs like
MDEV-7458 and MDEV-8302, this patch adds some extra precations to try to
detect (in debug builds) or prevent (in release builds) similar bugs.
ha_rollback_trans() will now call unmark_start_commit() if needed (and
assert in debug build when a caller does rollback without unmark first).
We also add an extra check for thd->killed() so that we avoid doing
mark_start_commit() if we already have a pending deadlock kill.
And we add a missing unmark_start_commit() call in the error case, found by
the above assertion.
When the slave processes the master restart format_description event,
parallel replication needs to complete any prior events before processing
the restart event (which closes temporary tables and such stuff).
This happens in wait_for_workers_idle(), however it was not waiting long
enough. The wait was using wait_for_prior_commit(), but at that points table
can still be open. This lead to assertion in this case.
So change wait_for_workers_idle() to wait until all worker threads have
reached finish_event_group(), at which point all tables should have been
closed.
In optimistic parallel replication, it is not safe to try to run a following
transaction in parallel with a DDL statement, and there is code to prevent
this.
However, the code was missing the case where the DDL is the very first event
after slave start. In this case, following transactions could run in
parallel with the DDL, which can cause the slave to hang or even corrupt
slave in unlucky cases.
There was a rare race, where a deadlock error might not be correctly
handled, causing the slave to stop with something like this in the error
log:
150423 14:04:10 [ERROR] Slave SQL: Connection was killed, Gtid 0-1-2, Internal MariaDB error code: 1927
150423 14:04:10 [Warning] Slave: Connection was killed Error_code: 1927
150423 14:04:10 [Warning] Slave: Deadlock found when trying to get lock; try restarting transaction Error_code: 1213
150423 14:04:10 [Warning] Slave: Connection was killed Error_code: 1927
150423 14:04:10 [Warning] Slave: Connection was killed Error_code: 1927
150423 14:04:10 [ERROR] Error running query, slave SQL thread aborted. Fix the problem, and restart the slave SQL thread with "SLAVE START". We stopped at log 'master-bin.000001 position 1234
The problem was incorrect error handling. When a deadlock is detected, it
causes a KILL CONNECTION on the offending thread. This error is then later
converted to a deadlock error, and the transaction is retried.
However, the deadlock error was not cleared at the start of the retry, nor
was the lingering kill signal. So it was possible to get another deadlock
kill early during retry. If this happened with particular thread
scheduling/timing, it was possible that the new KILL CONNECTION error was
masked by the earlier deadlock error, so that the second kill was not
properly converted into a deadlock error and retry.
This patch adds code that clears the old error and killed flag before
starting the retry. It also adds code to handle a deadlock kill caught in a
couple of places where it was not handled before.
The hangs occur when the group_commit_orderer object is freed before the last
mark_start_commit() call on it - this loses the wakeup to other waiting worker
threads, causing them to hang until killed manually.
The object was freed because wakeup_subsequent_commits() was called two early
in two places. For MDEV-7888, during ANALYZE TABLE, and for MDEV-7929 during
record_gtid() after processing a DDL event. The group_commit_orderer object
can be freed when its last transaction has called wait_for_prior_commit().
Fix by implementing a suspend/resume mechanism for wakeup_subsequent_commits()
that can be used in places where a transaction is committed without this being
the commit of the actual replication event group.
Also add a protection mechanism (that asserts in debug builds) which can
prevent the too-early free and hang if other similar bugs should remain in
other parts of the code.
This patch fixes a bug in the error handling in parallel replication, when one
worker thread gets a failure and other worker threads processing later
transactions have to rollback and abort.
The problem was with the lifetime of group_commit_orderer objects (GCOs).
A GCO is freed when we register that its last event group has committed. This
relies on register_wait_for_prior_commit() and wait_for_prior_commit() to
ensure that the fact that T2 has committed implies that any earlier T1 has
also committed, and can thus no longer execute mark_start_commit().
However, in the error case, the code was skipping the
register_wait_for_prior_commit() and wait_for_prior_commit() calls. Thus
commit ordering was not guaranteed, and a GCO could be freed too early. Then a
later mark_start_commit() would reference deallocated GCO, which could lead to
lost wakeup (causing slave threads to hang) or other corruption.
This patch makes also the error case respect commit order. This way, also the
error case gets the GCO lifetime correct, and the hang no longer occurs.
When a transaction in parallel replication needs to retry (eg. because of
deadlock kill), first wait for all prior transactions to commit before doing
the retry. This way, we avoid the retry once again conflicting with a prior
transaction, requiring yet another retry.
Without this patch, we saw "in the wild" that transactions had to be retried
more than 10 times to succeed, which exceeds the default
--slave_transaction_retries value and is in any case undesirable.
(We already do this in 10.1 in "optimistic" parallel replication mode; this
patch just makes the code use the same logic for "conservative" mode (only
mode in 10.0)).
The patch for optimistic parallel replication as a memory optimisation moved
the gco->installed field into a bit in gco->flags. However, that is just plain
wrong. The gco->flags field is owned by the SQL driver thread, but
gco->installed is used by the worker threads, so this will cause a race
condition.
The user-visible problem might be conflicts between transactions and/or slave
threads hanging.
So revert this part of the optimistic parallel replication patch, going back
to using a separate field gco->installed like in 10.0.
Delay spawning parallel replication worker threads until a slave SQL
thread is running, and de-spawn them when the last SQL thread stops.
This is especially useful to avoid needless threads on a master in a
setup where same my.cnf is used on masters and slaves.
