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.
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.
If somehow the COMMIT or XID event in an event group was missing, the code in
parallel replication to handle this was not sufficient, leading to server
deadlock.
Adjust the configuration options, as discussed on the
maria-developers@ mailing list.
The option to hint a transaction to not be replicated in parallel is
now called @@skip_parallel_replication, consistent with
@@skip_replication.
And the --slave-parallel-mode is now simplified to have just one of
the following values:
none
minimal
conservative
optimistic
aggressive
This reflects successively harder efforts to find opportunities to run
things in parallel on the slave. It allows to extend the server with
more automatic heuristics in the future without having to introduce a
new configuration option for each and every one.
pool->threads is freed before being reassigned the new pool.
Although not really a memory barrier I though it prudent to keep the pool
thread count to be the lower of the old/new thread list before the new threads
is allocated.
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.
The replication relay log position was sometimes updated incorrectly at the
end of a transaction in parallel replication. This happened because the relay
log file name was taken from the current Relay_log_info (SQL driver thread),
not the correct value for the transaction in question.
The result was that if a transaction was applied while the SQL driver thread
was at least one relay log file ahead, _and_ the SQL thread was subsequently
stopped before applying any events from the most recent relay log file, then
the relay log position would be incorrect - wrong relay log file name. Thus,
when the slave was started again, usually a relay log read error would result,
or in rare cases, if the position happened to be readable, the slave might
even skip arbitrary amounts of events.
In GTID mode, the relay log position is reset when both slave threads are
restarted, so this bug would only be seen in non-GTID mode, or in GTID mode
when only the SQL thread, not the IO thread, was stopped.
When a master server restarts, it writes a restart format_description event as
the first event in the next binlog file. The parallel slave SQL thread queues
a special restart entry for the current worker thread to signal this, so that
the worker thread can roll back any prior partial transaction that might have
been written to the binlog due to master crashing.
This queueing was missing a mysql_cond_signal() to notify the worker
thread. This could cause the worker thread to not process the restart entry,
and this in turn would cause the SQL thread to hang infinitely waiting for the
worker thread to complete processing.
Fix by adding the missing wakeup signalling for this case.
The retry of an event group in parallel replication set the wrong value for
the end log position of the event that was retried
(qev->future_event_relay_log_pos). It was too large by the size of the event,
so it pointed into the middle of the following event.
If the retry happened in the very last event of the event group, _and_ the SQL
thread was stopped just after successfully retrying that event, then the SQL
threads's relay log position would be left incorrect. Restarting the SQL
thread could then try to read events from a garbage offset in the relay log,
usually leading to an error about not being able to read the event.
The code that handles free lists of various objects passed to worker threads
in parallel replication handles freeing in batches, to avoid taking and
releasing LOCK_rpl_thread too often. However, it was possible for freeing to
be delayed to the point where one thread could stall the SQL driver thread due
to full queue, while other worker threads might be idle. This could
significantly degrade possible parallelism and thus performance.
Clean up the batch freeing code so that it is more robust and now able to
regularly free batches of object, so that normally the queue will not run full
unless the SQL driver thread is really far ahead of the worker threads.
The bug occured in parallel replication when re-trying transactions that
failed due to deadlock. In this case, the relay log file is re-opened and the
events are read out again. This reading requires a format description event of
the appropriate version. But the code was using a description event stored in
rli, which is not thread-safe. This could lead to various rare races if the
format description event was replaced by the SQL driver thread at the exact
moment where a worker thread was trying to use it.
The fix is to instead make the retry code create and maintain its own format
description event. When the relay log file is opened, we first read the format
description event from the start of the file, before seeking to the current
position. This now uses the same code as when the SQL driver threads starts
from a given relay log position. This also makes sure that the correct format
description event version will be used in cases where the version of the
binlog could change during replication.
In parallel replication, threads can do two different waits for a prior
transaction. One is for the prior transaction to start commit, the other is
for it to complete commit.
It turns out that the same PSI_stage_info message was errorneously used in
both cases (probably a merge error), causing SHOW PROCESSLIST to be
misleading.
Fix by using correct, distinct message in each case.
After-review fixes.
Mainly catching if the wait in wait_for_workers_idle() is aborted due to kill.
In this case, we should return an error and not proceed to execute the format
description event, as other threads might still be running for a bit until the
error is caught in all threads.
Follow-up patch, fixing a possible deadlock issue.
If the master crashes in the middle of an event group, there can be an active
transaction in a worker thread when we encounter the following master restart
format description event. In this case, we need to notify that worker thread
to abort and roll back the partial event group. Otherwise a deadlock occurs:
the worker thread waits for the commit that never arrives, and the SQL driver
thread waits for the worker thread to complete its event group, which it never
does.
mysqldump:
- Added --log-queries to allow one to disable logging for the dump
sql/log_event.cc:
- Removed setting of enable_slow_log as it's not required anymore.
sql/sql_parse.cc:
- Set enable_slow_log to value of thd->variables.sql_log_slow as this will speed up tests if slow log is disabled.
- opt_log_slow_admin_statements can now only disable slow log, not enable it.
sql/sql_explain.cc:
- Minor cleanup
Other things:
- Added sql_log_slow to system variables.
- Changed opt_slow_log to global_system_variables.sql_log_slow in all files
- Updated tests to reflect changes
When a master server starts up, it logs a special format_description event at
the start of a new binlog to mark that is has restarted. This is used by a
slave to drop all temporary tables - this is needed in case the master crashed
and did not have a chance to send explicit DROP TEMPORARY TABLE statements to
the slave.
In parallel replication, we need to be careful when dropping the temporary
tables - we need to be sure that no prior events are still executing that
might be using the temporary tables to be dropped, _and_ that no following
events have started executing that might have created new temporary tables
that should not be dropped.
This was not handled correctly, which could cause errors about access to not
existing temporary tables or even crashes. This patch implements that such
format_description events cause serialisation of event execution; all prior
events are executed to completion first, then the format_description event is
executed, dropping temporary tables, then following events are queued for
execution.
Master restarts should be sufficiently infrequent that the resulting loss of
parallelism should be of minimal impact.
The problem occured when using parallel replication, and an error occured that
caused the SQL thread to stop when the IO thread had already reached a
following binlog file from the master (or otherwise performed a relay log
rotation).
In this case, the Rotate Event at the end of the relay log file could still be
executed, even though an earlier event in that relay log file had gotten an
error. This would cause the position to be incorrectly updated, so that upon
restart of the SQL thread, the event that had failed would be silently skipped
and ignored, causing replication corruption.
Fixed by checking before executing Rotate Event, whether an earlier event
has failed. If so, the Rotate Event is not executed, just dequeued, same as
for other normal events following a failing event.
The bug was that in some cases, if a replicated transaction was rolled back
due to deadlock, during the subsequent retry of that transaction, the
gtid_slave_pos would _not_ be updated with the new GTID, leaving the GTID
position of the slave incorrect.
Fix this by ensuring during the retry that we clear the flag that marks that
the GTID has already been recorded in gtid_slave_pos, so that the update of
gtid_slave_pos will be done again during the retry.
In the original bug, the symptom was an assertion due to OPTION_GTID_BEGIN not
being cleared during the retry of the transaction. The reason was some code in
handling of a COMMIT query event, which would not clear the flag when not
recording a GTID in gtid_slave_pos. This commit also fixes that code to always
clear the OPTION_GTID_BEGIN flag for clarity, though it is actually not
possible for OPTION_GTID_BEGIN to become set unless a GTID is pending for
update (after fixing the bug described above).
When a MyISAM query is killed midway, the query is logged to the binlog marked
with the error.
The slave does not attempt to run the query, but aborts with a suitable error
message in the error log for the DBA to act on.
In this case, the parallel replication code would check the sql_errno() code,
even no my_error() had been set. In debug builds, this causes an assertion.
Fixed the code to check that we actually have an error set before querying for
an error code.
After-review changes.
For this patch in 10.0, we do not introduce a new public storage engine API,
we just fix the InnoDB/XtraDB issues. In 10.1, we will make a better public
API that can be used for all storage engines (MDEV-6429).
Eliminate the background thread that did deadlock kills asynchroneously.
Instead, we ensure that the InnoDB/XtraDB code can handle doing the kill from
inside the deadlock detection code (when thd_report_wait_for() needs to kill a
later thread to resolve a deadlock).
(We preserve the part of the original patch that introduces dedicated mutex
and condition for the slave init thread, to remove the abuse of
LOCK_thread_count for start/stop synchronisation of the slave init thread).
replication causing replication to fail.
Remove the temporary fix for MDEV-5914, which used READ COMMITTED for parallel
replication worker threads. Replace it with a better, more selective solution.
The issue is with certain edge cases of InnoDB gap locks, for example between
INSERT and ranged DELETE. It is possible for the gap lock set by the DELETE to
block the INSERT, if the DELETE runs first, while the record lock set by
INSERT does not block the DELETE, if the INSERT runs first. This can cause a
conflict between the two in parallel replication on the slave even though they
ran without conflicts on the master.
With this patch, InnoDB will ask the server layer about the two involved
transactions before blocking on a gap lock. If the server layer tells InnoDB
that the transactions are already fixed wrt. commit order, as they are in
parallel replication, InnoDB will ignore the gap lock and allow the two
transactions to proceed in parallel, avoiding the conflict.
Improve the fix for MDEV-6020. When InnoDB itself detects a deadlock, it now
asks the server layer for any preferences about which transaction to roll
back. In case of parallel replication with two transactions T1 and T2 fixed to
commit T1 before T2, the server layer will ask InnoDB to roll back T2 as the
deadlock victim, not T1. This helps in some cases to avoid excessive deadlock
rollback, as T2 will in any case need to wait for T1 to complete before it can
itself commit.
Also some misc. fixes found during development and testing:
- Remove thd_rpl_is_parallel(), it is not used or needed.
- Use KILL_CONNECTION instead of KILL_QUERY when a parallel replication
worker thread is killed to resolve a deadlock with fixed commit
ordering. There are some cases, eg. in sql/sql_parse.cc, where a KILL_QUERY
can be ignored if the query otherwise completed successfully, and this
could cause the deadlock kill to be lost, so that the deadlock was not
correctly resolved.
- Fix random test failure due to missing wait_for_binlog_checkpoint.inc.
- Make sure that deadlock or other temporary errors during parallel
replication are not printed to the the error log; there were some places
around the replication code with extra error logging. These conditions can
occur occasionally and are handled automatically without breaking
replication, so they should not pollute the error log.
- Fix handling of rgi->gtid_sub_id. We need to be able to access this also at
the end of a transaction, to be able to detect and resolve deadlocks due to
commit ordering. But this value was also used as a flag to mark whether
record_gtid() had been called, by being set to zero, losing the value. Now,
introduce a separate flag rgi->gtid_pending, so rgi->gtid_sub_id remains
valid for the entire duration of the transaction.
- Fix one place where the code to handle ignored errors called reset_killed()
unconditionally, even if no error was caught that should be ignored. This
could cause loss of a deadlock kill signal, breaking deadlock detection and
resolution.
- Fix a couple of missing mysql_reset_thd_for_next_command(). This could
cause a prior error condition to remain for the next event executed,
causing assertions about errors already being set and possibly giving
incorrect error handling for following event executions.
- Fix code that cleared thd->rgi_slave in the parallel replication worker
threads after each event execution; this caused the deadlock detection and
handling code to not be able to correctly process the associated
transactions as belonging to replication worker threads.
- Remove useless error code in slave_background_kill_request().
- Fix bug where wfc->wakeup_error was not cleared at
wait_for_commit::unregister_wait_for_prior_commit(). This could cause the
error condition to wrongly propagate to a later wait_for_prior_commit(),
causing spurious ER_PRIOR_COMMIT_FAILED errors.
- Do not put the binlog background thread into the processlist. It causes
too many result differences in mtr, but also it probably is not useful
for users to pollute the process list with a system thread that does not
really perform any user-visible tasks...
replication causing replication to fail.
In parallel replication, we run transactions from the master in parallel, but
force them to commit in the same order they did on the master. If we force T1
to commit before T2, but T2 holds eg. a row lock that is needed by T1, we get
a deadlock when T2 waits until T1 has committed.
Usually, we do not run T1 and T2 in parallel if there is a chance that they
can have conflicting locks like this, but there are certain edge cases where
it can occasionally happen (eg. MDEV-5914, MDEV-5941, MDEV-6020). The bug was
that this would cause replication to hang, eventually getting a lock timeout
and causing the slave to stop with error.
With this patch, InnoDB will report back to the upper layer whenever a
transactions T1 is about to do a lock wait on T2. If T1 and T2 are parallel
replication transactions, and T2 needs to commit later than T1, we can thus
detect the deadlock; we then kill T2, setting a flag that causes it to catch
the kill and convert it to a deadlock error; this error will then cause T2 to
roll back and release its locks (so that T1 can commit), and later T2 will be
re-tried and eventually also committed.
The kill happens asynchroneously in a slave background thread; this is
necessary, as the reporting from InnoDB about lock waits happen deep inside
the locking code, at a point where it is not possible to directly call
THD::awake() due to mutexes held.
Deadlock is assumed to be (very) rarely occuring, so this patch tries to
minimise the performance impact on the normal case where no deadlocks occur,
rather than optimise the handling of the occasional deadlock.
Also fix transaction retry due to deadlock when it happens after a transaction
already signalled to later transactions that it started to commit. In this
case we need to undo this signalling (and later redo it when we commit again
during retry), so following transactions will not start too early.
Also add a missing thd->send_kill_message() that got triggered during testing
(this corrects an incorrect fix for MySQL Bug#58933).
Handle retry of event groups that span multiple relay log files.
- If retry reaches the end of one relay log file, move on to the next.
- Handle refcounting of relay log files, and avoid purging relay log
files until all event groups have completed that might have needed
them for transaction retry.
Implement that if first retry fails, we can do another attempt.
Add testcases to test multi-retry that succeeds in second attempt, and
multi-retry that eventually fails due to exceeding slave_trans_retries.
Start implementing that an event group can be re-tried in parallel replication
if it fails with a temporary error (like deadlock).
Patch is very incomplete, just some very basic retry works.
Stuff still missing (not complete list):
- Handle moving to the next relay log file, if event group to be retried
spans multiple relay log files.
- Handle refcounting of relay log files, to ensure that we do not purge a
relay log file and then later attempt to re-execute events out of it.
- Handle description_event_for_exec - we need to save this somehow for the
possible retry - and use the correct one in case it differs between relay
logs.
- Do another retry attempt in case the first retry also fails.
- Limit the max number of retries.
- Lots of testing will be needed for the various edge cases.
The direct cause of the assertion was missing error handling in
record_gtid(). If ha_commit_trans() fails for the statement commit, there was
missing code to catch the error and do ha_rollback_trans() in this case; this
caused close_thread_tables() to assert.
Normally, this error case is not hit, but in this case it was triggered due to
another bug: When a transaction T1 fails during parallel replication, the code
would signal following transactions that they could start to run without
properly marking the error condition. This caused subsequent transactions to
incorrectly start replicating, only to get an error later during their own
commit step. This was particularly serious if the subsequent transactions were
DDL or MyISAM updates, which cannot be rolled back and would leave replication
in an inconsistent state.
Fixed by 1) in case of error, only signal following transactions to continue
once the error has been properly marked and those transactions will know not
to start; and 2) implement proper error handling in record_gtid() in the case
that statement commit fails.
If replication breaks in GTID mode, it is not trivial to determine the GTID of
the failing event group. This is a problem, as such GTID is needed eg. to
explicitly set @@gtid_slave_pos to skip to after that event group, or to
compare errors on different servers, etc.
Fix by ensuring that relevant slave errors logged to the error log include the
GTID of the event group containing the problem event.
The previous patch for this bug was unfortunately completely wrong.
The purpose of cached_charset is to remember which character set we
have installed currently in the THD, so that in the common case where
charset does not change between queries, we do not need to update it
in the THD. Thus, it is important that the cached_charset field is
tightly coupled to the THD for which it handles caching.
Thus the right place to put cached_charset seems to be in the THD.
This patch introduces a field THD:system_thread_info where such info
in general can be placed without further inflating the THD with unused
data for other threads (THD is already far too big as it is). It then
moves the cached_charset into this slot for the SQL driver thread and
for the parallel replication worker threads.
The THD::rpl_filter field is also moved inside system_thread_info, to
keep the size of THD unchanged. Moving further fields in to reduce the
size of THD is a separate task, filed as MDEV-6164.
In parallel replication, there was an error case where we could call
my_error() in-between events. This causes the assertion, as the previous event
has reported ok status, but the following event has not yet reset the
diagnostics area. This happened when a worker thread detects that the SQL
driver thread is aborting, and when it gets an error from a prior commit at
the same time in wait_for_prior_commit().
Since this is already an error case, the code should be using
unregister_wait_for_prior_commit() instead of wait_for_prior_commit(). But
unregister is already done a bit later (from finish_event_group()), so just
removing the redundant call to wait_for_prior_commit() fixes the issue.
Due to how gap locks work, two transactions could group commit together on the
master, but get lock conflicts and then deadlock due to different thread
scheduling order on slave.
For now, remove these deadlocks by running the parallel slave in READ
COMMITTED mode. And let InnoDB/XtraDB allow statement-based binlogging for the
parallel slave in READ COMMITTED.
We are also investigating a different solution long-term, which is based on
relaxing the gap locks only between the transactions running in parallel for
one slave, but not against possibly external transactions.
When a transaction fails in parallel replication, it should signal the error
to any following transactions doing wait_for_prior_commit() on it. But the
code for this was incorrect, and would not correctly remember a prior error
when sending the signal. This caused corruption when slave stopped due to an
error.
Fix by remembering the error code when we first get an error, and passing the
saved error code to wakeup_subsequent_commits().
Thanks to nanyi607rao who reported this bug on
maria-developers@lists.launchpad.net and analysed the root cause.
Before, the arrival of same GTID twice in multi-source replication
would cause double-apply or in gtid strict mode an error.
Keep the behaviour, but add an option --gtid-ignore-duplicates which
allows to correctly handle duplicates, ignoring all but the first.
This relies on the user ensuring correct configuration so that
sequence numbers are strictly increasing within each replication
domain; then duplicates can be detected simply by comparing the
sequence numbers against what is already applied.
Only one master connection (but possibly multiple parallel worker
threads within that connection) is allowed to apply events within
one replication domain at a time; any other connection that
receives a GTID in the same domain either discards it (if it is
already applied) or waits for the other connection to not have
any events to apply.
Intermediate patch, as proof-of-concept for testing. The main limitation
is that currently it is only implemented for parallel replication,
@@slave_parallel_threads > 0.
Make sure to signal the condition variable for the thread pool after
the new threads have been added to the pool.
Thanks to user nanyi607rao, who reported this bug on maria-developers@.
When an rpl_group_info object was returned from the free list, the
rgi->deferred_events_collecting and rgi->deferred_events was not correctly
re-inited. Additionally, the rgi->deferred_events was incorrectly freed in
free_rgi(), which causes unnecessary malloc/free (or crash when re-init is not
done).
Thanks to user nanyi607rao, who reported this bug on maria-developers@.
