Clear any pending deadlock kill after completing XA PREPARE, and before
updating the mysql.gtid_slave_pos table in a separate transaction.
Reviewed-by: Andrei Elkin <andrei.elkin@mariadb.com>
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
Keep track of each recently active XID, recording which worker it was queued
on. If an XID might still be active, choose the same worker to queue event
groups that refer to the same XID to avoid conflicts.
Otherwise, schedule the XID freely in the next round-robin slot.
This way, XA PREPARE can normally be scheduled without restrictions (unless
duplicate XID transactions come close together). This improves scheduling
and parallelism over the old method, where the worker thread to schedule XA
PREPARE on was fixed based on a hash value of the XID.
XA COMMIT will normally be scheduled on the same worker as XA PREPARE, but
can be a different one if the XA PREPARE is far back in the event history.
Testcase and code for trimming dynamic array due to Andrei.
Reviewed-by: Andrei Elkin <andrei.elkin@mariadb.com>
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
This is a preparatory patch to facilitate the next commit to improve
the scheduling of XA transactions in parallel replication.
When choosing the scheduling bucket for the next event group in
rpl_parallel_entry::choose_thread(), use an explicit FIFO for the
round-robin selection instead of a simple cyclic counter i := (i+1) % N.
This allows to schedule XA COMMIT/ROLLBACK dependencies explicitly without
changing the round-robin scheduling of other event groups.
Reviewed-by: Andrei Elkin <andrei.elkin@mariadb.com>
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
AKA rpl.rpl_parallel, binlog_encryption.rpl_parallel fails in
buildbot with timeout in include
A replication parallel worker thread can deadlock with another
connection running SHOW SLAVE STATUS. That is, if the replication
worker thread is in do_gco_wait() and is killed, it will already
hold the LOCK_parallel_entry, and during error reporting, try to
grab the err_lock. SHOW SLAVE STATUS, however, grabs these locks in
reverse order. It will initially grab the err_lock, and then try to
grab LOCK_parallel_entry. This leads to a deadlock when both threads
have grabbed their first lock without the second.
This patch implements the MDEV-31894 proposed fix to optimize the
workers_idle() check to compare the last in-use relay log’s
queued_count==dequeued_count for idleness. This removes the need for
workers_idle() to grab LOCK_parallel_entry, as these values are
atomically updated.
Huge thanks to Kristian Nielsen for diagnosing the problem!
Reviewed By:
============
Kristian Nielsen <knielsen@knielsen-hq.org>
Andrei Elkin <andrei.elkin@mariadb.com>
The problem was an incorrect unmark_start_commit() in
signal_error_to_sql_driver_thread(). If an event group gets an error, this
unmark could run after the following GCO started, and the subsequent
re-marking could access de-allocated GCO.
The offending unmark_start_commit() looks obviously incorrect, and the fix
is to just remove it. It was introduced in the MDEV-8302 patch, the commit
message of which suggests it was added there solely to satisfy an assertion
in ha_rollback_trans(). So update this assertion instead to not trigger for
event groups that experienced an error (rgi->worker_error). When an error
occurs in an event group, all following event groups are skipped anyway, so
the unmark should never be needed in this case.
Reviewed-by: Andrei Elkin <andrei.elkin@mariadb.com>
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
At STOP SLAVE, worker threads will continue applying event groups until the
end of the current GCO before stopping. This is a left-over from when only
conservative mode was available. In optimistic and aggressive mode, often
_all_ queued event will be in the same GCO, and slave stop will be
needlessly delayed.
This patch instead records at STOP SLAVE time the latest (highest sub_id)
event group that has started. Then worker threads will continue to apply
event groups up to that event group, but skip any following. The result is
that each worker thread will complete its currently running event group, and
then the slave will stop.
If the slave is caught up, and STOP SLAVE is run in the middle of an event
group that is already executing in a worker thread, then that event group
will be rolled back and the slave stop immediately, as normal.
Reviewed-by: Andrei Elkin <andrei.elkin@mariadb.com>
Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>
…with: Test assertion failed
Problem:
=======
Assertion text: 'Value returned by SSS and PS table for Last_Error_Number
should be same.'
Assertion condition: '"1146" = "0"'
Assertion condition, interpolated: '"1146" = "0"'
Assertion result: '0'
Analysis:
========
In parallel replication when slave is started the worker pool gets
activated and it gets cleared when slave stops. Each time the worker pool
gets activated a backup worker pool also gets created to store worker
specific perforance schema information in case of errors. On error, all
relevant information is copied from rpl_parallel_thread to rli and it gets
cleared from thread. Then server waits for all workers to complete their
work, during this stage performance schema table specific worker info is
stored into the backup pool and finally the actual pool gets cleared. If
users query the performance schema table to know the status of workers the
information from backup pool will be used. The test simulates
ER_NO_SUCH_TABLE error and verifies the worker information in pfs table.
Test works fine if execution occurs in following order.
Step 1. Error occurred 'worker information is copied to backup pool'.
Step 2. handle_slave_sql invokes 'rpl_parallel_resize_pool_if_no_slaves' to
deactivate worker pool, it marks the pool->count=0
Step 3. PFS table is queried, since actual pool is deactivated backup pool
information is read.
If the Step 3 happens prior to Step2 the pool is yet to be deactivated and
the actual pool is read, which doesn't have any error details as they were
cleared. Hence test ocasionally fails.
Fix:
===
Upon error mark the back pool as being active so that if PFS table is
quried since the backup pool is flagged as valid its information will be
read, in case it is not flagged regular pool will be read.
This work is one of the last pieces created by the late Sujatha Sivakumar.
The hang could be seen as show slave status displaying an error like
Last_Error: Could not execute Write_rows_v1
along with
Slave_SQL_Running: Yes
accompanied with one of the replication threads in show-processlist
characteristically having status like
2394 | system user | | NULL | Slave_worker | 50852| closing tables
It turns out that closing tables worker got entrapped in endless looping
in mark_start_commit_inner() across already garbage-collected gco items.
The reclaimed gco links are explained with actually possible
out-of-order groups of events termination due to the Last_Error.
This patch reinforces the correct ordering to perform
finish_event_group's cleanup actions, incl unlinking gco:s
from the active list.
This commit implements two phase binloggable ALTER.
When a new
@@session.binlog_alter_two_phase = YES
ALTER query gets logged in two parts, the START ALTER and the COMMIT
or ROLLBACK ALTER. START Alter is written in binlog as soon as
necessary locks have been acquired for the table. The timing is
such that any concurrent DML:s that update the same table are either
committed, thus logged into binary log having done work on the old
version of the table, or will be queued for execution on its new
version.
The "COMPLETE" COMMIT or ROLLBACK ALTER are written at the very point
of a normal "single-piece" ALTER that is after the most of
the query work is done. When its result is positive COMMIT ALTER is
written, otherwise ROLLBACK ALTER is written with specific error
happened after START ALTER phase.
Replication of two-phase binloggable ALTER is
cross-version safe. Specifically the OLD slave merely does not
recognized the start alter part, still being able to process and
memorize its gtid.
Two phase logged ALTER is read from binlog by mysqlbinlog to produce
BINLOG 'string', where 'string' contains base64 encoded
Query_log_event containing either the start part of ALTER, or a
completion part. The Query details can be displayed with `-v` flag,
similarly to ROW format events. Notice, mysqlbinlog output containing
parts of two-phase binloggable ALTER is processable correctly only by
binlog_alter_two_phase server.
@@log_warnings > 2 can reveal details of binlogging and slave side
processing of the ALTER parts.
The current commit also carries fixes to the following list of
reported bugs:
MDEV-27511, MDEV-27471, MDEV-27349, MDEV-27628, MDEV-27528.
Thanks to all people involved into early discussion of the feature
including Kristian Nielsen, those who helped to design, implement and
test: Sergei Golubchik, Andrei Elkin who took the burden of the
implemenation completion, Sujatha Sivakumar, Brandon
Nesterenko, Alice Sherepa, Ramesh Sivaraman, Jan Lindstrom.
Problem:
=======
Test fails with 3 different symptoms
connection slave;
Assertion text: 'Last_Seen_Transaction should show .'
Assertion condition: '"0-1-1" = ""'
Assertion condition, interpolated: '"0-1-1" = ""'
Assertion result: '0'
connection slave;
Assertion text: 'Value returned by SSS and PS table for Last_Error_Number
should be same.'
Assertion condition: '"1146" = "0"'
Assertion condition, interpolated: '"1146" = "0"'
Assertion result: '0'
connection slave;
Assertion text: 'Value returned by PS table for worker_idle_time should be
>= 1'
Assertion condition: '"0" >= "1"'
Assertion condition, interpolated: '"0" >= "1"'
Assertion result: '0'
Fix1:
====
Performance schema table's Last_Seen_Transaction is compared with 'SELECT
gtid_slave_pos'. Since DDLs are not transactional changes to user table and
gtid_slave_pos table are not guaranteed to be synchronous. To fix the
issue Gtid_IO_Pos value from SHOW SLAVE STATUS command will be used to
verify the correctness of Performance schema specific
Last_Seen_Transaction.
Fix2:
====
On error worker thread information is stored as part of backup pool. Access
to this backup pool should be protected by 'LOCK_rpl_thread_pool' mutex so
that simultaneous START SLAVE cannot destroy the backup pool, while it is
being queried by performance schema.
Fix3:
====
When a worker is waiting for events if performance schema table is queried,
at present it just returns the difference between current_time and
start_time. This is incorrect. It should be worker_idle_time +
(current_time - start_time).
For example a worker thread was idle for 10 seconds and then it got events
to process. Upon completion it goes to idle state, now if the pfs table is
queried it should return current_idle time + worker_idle_time.
Step 3:
======
Preserve worker pool information on either STOP SLAVE/Error. In case STOP
SLAVE is executed worker threads will be gone, hence worker threads will be
unavailable. Querying the table at this stage will give empty rows. To
address this case when worker threads are about to stop, due to an error or
forced stop, create a backup pool and preserve the data which is relevant to
populate performance schema table. Clear the backup pool upon slave start.
Step2:
=====
Add two extra columns mentioned below.
---------------------------------------------------------------------------
|Column Name: | Description: |
|-------------------------------------------------------------------------|
| | |
|WORKER_IDLE_TIME | Total idle time in seconds that the worker |
| | thread has spent waiting for work from |
| | co-ordinator thread |
| | |
|LAST_TRANS_RETRY_COUNT | Total number of retries attempted by last |
| | transaction |
---------------------------------------------------------------------------
Step1:
=====
Backport 'replication_applier_status_by_worker' from upstream.
Iterate through rpl_parallel_thread_pool and display slave worker thread
specific information as part of 'replication_applier_status_by_worker'
table.
---------------------------------------------------------------------------
|Column Name: | Description: |
|-------------------------------------------------------------------------|
| | |
|CHANNEL_NAME | Name of replication channel through which the |
| | transaction is received. |
| | |
|THREAD_ID | Thread_Id as displayed in 'performance_schema. |
| | threads' table for thread with name |
| | 'thread/sql/rpl_parallel_thread' |
| | |
| | THREAD_ID will be NULL when worker threads are |
| | stopped due to an error/force stop |
| | |
|SERVICE_STATE | Thread is running or not |
| | |
|LAST_SEEN_TRANSACTION | Last GTID executed by worker |
| | |
|LAST_ERROR_NUMBER | Last Error that occured on a particular worker |
| | |
|LAST_ERROR_MESSAGE | Last error specific message |
| | |
|LAST_ERROR_TIMESTAMP | Time stamp of last error |
| | |
---------------------------------------------------------------------------
CHANNEL_NAME will be empty when the worker has not processed any
transaction. Channel_name points to valid source channel_name when it is
processing a transaction/event group.
Problem:- When we issue FTWRL with shutdown in parallel, there is race between
FTWRL and shutdown. Shutdown might destroy the mutex (pool->LOCK_rpl_thread_pool)
before FTWRL can lock it. So we can get crash on FTWRL thread
Solution:- mysql_mutex_destroy(pool->LOCK_rpl_thread_pool) should wait for
FTWRL thread to complete its work , and then destroy.
So slave_prepare_for_shutdown will just deactivate the pool, and mutex is destroyed
later in end_slave()
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!
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.
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.
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.
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.
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.
The problem occurs in parallel replication in GTID mode, when we are using
multiple replication domains. In this case, if the SQL thread stops, the
slave GTID position may refer to a different point in the relay log for each
domain.
The bug was that when the SQL thread was stopped and restarted (but the IO
thread was kept running), the SQL thread would resume applying the relay log
from the point of the most advanced replication domain, silently skipping all
earlier events within other domains. This caused replication corruption.
This patch solves the problem by storing, when the SQL thread stops with
multiple parallel replication domains active, the current GTID
position. Additionally, the current position in the relay logs is moved back
to a point known to be earlier than the current position of any replication
domain. Then when the SQL thread restarts from the earlier position, GTIDs
encountered are compared against the stored GTID position. Any GTID that was
already applied before the stop is skipped to avoid duplicate apply.
This patch should have no effect if multi-domain GTID parallel replication is
not used. Similarly, if both SQL and IO thread are stopped and restarted, the
patch has no effect, as in this case the existing relay logs are removed and
re-fetched from the master at the current global @@gtid_slave_pos.
The bug occurs when a transaction does a retry after all transactions have
done mark_start_commit() in a batch of group commit from the master. In this
case, the retrying transaction can unmark_start_commit() after the following
batch has already started running and de-allocated the GCO. Then after retry,
the transaction will re-do mark_start_commit() on a de-allocated GCO, and also
wakeup of later GCOs can be lost.
This was seen "in the wild" by a user, even though it is not known exactly
what circumstances can lead to retry of one transaction after all transactions
in a group have reached the commit phase.
The lifetime around GCO was somewhat clunky anyway. With this patch, a GCO
lives until rpl_parallel_entry::last_committed_sub_id has reached the last
transaction in the GCO. This guarantees that the GCO will still be alive when
a transaction does mark_start_commit(). Also, we now loop over the list of
active GCOs for wakeup, to ensure we do not lose a wakeup even in the
problematic case.
Implement a new mode for parallel replication. In this mode, all transactions
are optimistically attempted applied in parallel. In case of conflicts, the
offending transaction is rolled back and retried later non-parallel.
This is an early-release patch to facilitate testing, more changes to user
interface / options will be expected. The new mode is not enabled by default.
