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.
All changes (except one) is of type
thd->transaction. -> thd->transaction->
thd->transaction points by default to 'thd->default_transaction'
This allows us to 'easily' have multiple active transactions for a
THD object, like when reading data from the mysql.proc table
in fact, in MariaDB it cannot, but it can show spurious slaves
in SHOW SLAVE HOSTS.
slave was registered in COM_REGISTER_SLAVE and un-registered after
COM_BINLOG_DUMP. If there was no COM_BINLOG_DUMP, it would never
unregister.
This is a backport of the applicable part of
commit 93475aff8d and
commit 2c39f69d34
from 10.4.
Before 10.4 and Galera 4, WSREP_ON is a macro that points to
a global Boolean variable, so it is not that expensive to
evaluate, but we will add an unlikely() hint around it.
WSREP_ON_NEW: Remove. This macro was introduced in
commit c863159c32
when reverting WSREP_ON to its previous definition.
We replace some use of WSREP_ON with WSREP(thd), like it was done
in 93475aff8d. Note: the macro
WSREP() in 10.1 is equivalent to WSREP_NNULL() in 10.4.
Item_func_rand::seed_random(): Avoid invoking current_thd
when WSREP is not enabled.
MDEV-21605 Clean up and speed up interfaces for binary row logging
MDEV-21617 Bug fix for previous version of this code
The intention is to have as few 'if' as possible in ha_write() and
related functions. This is done by pre-calculating once per statement the
row_logging state for all tables.
Benefits are simpler and faster code both when binary logging is disabled
and when it's enabled.
Changes:
- Added handler->row_logging to make it easy to check it table should be
row logged. This also made it easier to disabling row logging for system,
internal and temporary tables.
- The tables row_logging capabilities are checked once per "statements
that updates tables" in THD::binlog_prepare_for_row_logging() which
is called when needed from THD::decide_logging_format().
- Removed most usage of tmp_disable_binlog(), reenable_binlog() and
temporary saving and setting of thd->variables.option_bits.
- Moved checks that can't change during a statement from
check_table_binlog_row_based() to check_table_binlog_row_based_internal()
- Removed flag row_already_logged (used by sequence engine)
- Moved binlog_log_row() to a handler::
- Moved write_locked_table_maps() to THD::binlog_write_table_maps() as
most other related binlog functions are in THD.
- Removed binlog_write_table_map() and binlog_log_row_internal() as
they are now obsolete as 'has_transactions()' is pre-calculated in
prepare_for_row_logging().
- Remove 'is_transactional' argument from binlog_write_table_map() as this
can now be read from handler.
- Changed order of 'if's in handler::external_lock() and wsrep_mysqld.h
to first evaluate fast and likely cases before more complex ones.
- Added error checking in ha_write_row() and related functions if
binlog_log_row() failed.
- Don't clear check_table_binlog_row_based_result in
clear_cached_table_binlog_row_based_flag() as it's not needed.
- THD::clear_binlog_table_maps() has been replaced with
THD::reset_binlog_for_next_statement()
- Added 'MYSQL_OPEN_IGNORE_LOGGING_FORMAT' flag to open_and_lock_tables()
to avoid calculating of binary log format for internal opens. This flag
is also used to avoid reading statistics tables for internal tables.
- Added OPTION_BINLOG_LOG_OFF as a simple way to turn of binlog temporary
for create (instead of using THD::sql_log_bin_off.
- Removed flag THD::sql_log_bin_off (not needed anymore)
- Speed up THD::decide_logging_format() by remembering if blackhole engine
is used and avoid a loop over all tables if it's not used
(the common case).
- THD::decide_logging_format() is not called anymore if no tables are used
for the statement. This will speed up pure stored procedure code with
about 5%+ according to some simple tests.
- We now get annotated events on slave if a CREATE ... SELECT statement
is transformed on the slave from statement to row logging.
- In the original code, the master could come into a state where row
logging is enforced for all future events if statement could be used.
This is now partly fixed.
Other changes:
- Ensure that all tables used by a statement has query_id set.
- Had to restore the row_logging flag for not used tables in
THD::binlog_write_table_maps (not normal scenario)
- Removed injector::transaction::use_table(server_id_type sid, table tbl)
as it's not used.
- Cleaned up set_slave_thread_options()
- Some more DBUG_ENTER/DBUG_RETURN, code comments and minor indentation
changes.
- Ensure we only call THD::decide_logging_format_low() once in
mysql_insert() (inefficiency).
- Don't annotate INSERT DELAYED
- Removed zeroing pos_in_table_list in THD::open_temporary_table() as it's
already 0
If async slave thread (slave SQL handler), becomes a BF victim, it may occasionally happen that rollbacker thread is used to carry out the rollback instead of the async slave thread.
This can happen, if async slave thread has flagged "idle" state when BF thread tries to figure out how to kill the victim.
The issue was possible to test by using a galera cluster as slave for external master, and issuing high load of conflicting writes through async replication and directly against galera cluster nodes.
However, a deterministic mtr test for the "conflict window" has not yet been worked on.
The fix, in this patch makes sure that async slave thread state is never set to IDLE. This prevents the rollbacker thread to intervene.
The wsrep_query_state change was refactored to happen by dedicated function to make controlling the idle state change in one place.
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!
If async replication slave thread conflicts with cluster replication,
then the async slave transaction should be BF aborted, and depending on the
state of async slave transaction execution, potentially also replayed.
There were problems in such BF abort implementation and the replaying was not
started.
This pull request contains fixes which make sure that if async slave thread is
marked to abort and replay, it will complete carry out the rollback and
release all locks and resources before starting the replaying. After replaying,
async slave transactions is treated as successful, so the slave thread will
continue as usual, handling next replication event.
There is also new mtr test: galera.galera_slave_replay, which stresses both a
certification failure for async slave thread and a successful BF abort
followed by replaying.
Parallel slave server shutdown found to be hanging in
close_connections() triggered by shutdown due to a slave worker thread
would not be notified to exit in case the worker was sitting idle.
Fixed with destroying the worker pool earlier that is in
slave_prepare_for_shutdown() when all their driver threads have already left.
A test file is added to simulate the bug condition as well as check
multi-sourced and not-idle worker cases.
The string doesn't appear to be null-terminated when binlog checksums are
enabled. This causes a corrupt binlog name in the error message when a
slave is ahead of the master.
In MariaDB 10.2 master could have been configured so that there
is extra annotate events. When we peak next event type for CTAS we
need to skip annotate events.
The problem happens when MariaDB master replicates writes for only non InnoDB
tables (e.g. writes to MyISAM table(s)). Async slave node, in Galera cluster,
can apply these writes successfully, but it will, in the end, write gtid position in
mysql.gtid_slave_pos table. mysql.gtid_slave_pos table is InnoDB engine, and
this write makes innodb handlerton part of the replicated "transaction".
Note that wsrep patch identifies that write to gtid_slave_pos should not be replicated
and skips appending wsrep keys for these writes. However, as InnoDB was present
in the transaction, and there are replication events (for MyISAM table) in transaction
cache, but there are no appended keys, wsrep raises an error, and this makes the söave
thread to stop.
The fix is simply to not treat it as an error if async slave tries to replicate a write
set with binlog events, but no keys. We just skip wsrep replication and return successfully.
This commit contains also a mtr test which forces mysql.gtid_slave_pos table isto be
of InnoDB engine, and executes MyISAM only write through asyn replication.
There is additional fix for declaring IO and background slave threads as non wsrep.
These threads should not write anything for wsrep replication, and this is just a safeguard
to make sure nothing leaks into cluster from these slave threads.
This PR contains a mtr test for reproducing a failure with replicating create table as select statement (CTAS) through asynchronous mariadb replication to mariadb galera cluster.
The problem happens when CTAS replication contains both create table statement followed by row events for populating the table. In such situation, the galera node operating as mariadb replication slave, will first replicate only the create table part into the cluster, and then perform another replication containing both the create table and row events. This will lead all other nodes to fail for duplicate table create attempt, and crash due to this failure.
PR contains also a fix, which identifies the situation when CTAS has been replicated, and makes further scan in async replication stream to see if there are following row events. The slave node will replicate either single TOI in case the CTAS table is empty, or if CTAS table contains rows, then single bundled write set with create table and row events is replicated to galera cluster.
This fix should keep master server's GTID's for CTAS replication in sync with GTID's in galera cluster.
