percona/percona-toolkit
1
0
Fork 0
mirror of https://github.com/percona/percona-toolkit.git synced 2025-09-28 08:51:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge Kenny's pt-slave-restart-gtid-support branch.

Browse Source
This commit is contained in:
Daniel Nichter
2014-05-28 15:26:26 -07:00
parent d917571f86 b50c35c828
commit 0ffcec054e
7 changed files with 419 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

138
bin/pt-slave-restart
View File

@@ -4766,6 +4766,12 @@ use sigtrap qw(handler finish untrapped normal-signals);
use Percona::Toolkit;
use constant PTDEBUG => $ENV{PTDEBUG} || 0;
use Data::Dumper;
local $Data::Dumper::Indent = 1;
local $Data::Dumper::Sortkeys = 1;
local $Data::Dumper::Quotekeys = 0;
$OUTPUT_AUTOFLUSH = 1;
my $o;
@@ -4976,10 +4982,36 @@ sub watch_server {
$start_sql .= " UNTIL RELAY_LOG_FILE = '$file', RELAY_LOG_POS = $pos";
}
my $set_skip = $dbh->prepare("SET GLOBAL SQL_SLAVE_SKIP_COUNTER = "
. $o->get('skip-count'));
my $start = $dbh->prepare($start_sql);
my $stop = $dbh->prepare('STOP SLAVE');
my $start = $dbh->prepare($start_sql);
my $stop = $dbh->prepare('STOP SLAVE');
# ########################################################################
# Detect if GTID is enabled. Skipping an event is done differently.
# ########################################################################
# When MySQL 5.6.5 or higher is used and gtid is enabled, skipping a
# transaction is not possible with SQL_SLAVE_SKIP_COUNTER
my $skip_event;
my $have_gtid = 0;
if ( VersionParser->new($dbh) >= '5.6.5' ) {
my $row = $dbh->selectrow_arrayref('SELECT @@GLOBAL.gtid_mode');
PTDEBUG && _d('@@GLOBAL.gtid_mode:', $row->[0]);
if ( $row && $row->[0] eq 'ON' ) {
$have_gtid = 1;
}
}
PTDEBUG && _d('Have GTID:', $have_gtid);
# If GTID is enabled, slave_parallel_workers should be == 0.
# It's currently not possible to know what GTID event the failed trx is.
if ( $have_gtid ) {
my $threads = $dbh->selectrow_hashref(
'SELECT @@GLOBAL.slave_parallel_workers AS threads');
if ( $threads->{threads} > 0 ) {
die "Cannot skip transactions properly because GTID is enabled "
. "and slave_parallel_workers > 0. See 'GLOBAL TRANSACTION IDS' "
. "in the tool's documentation.\n";
}
}
# ########################################################################
# Lookup tables of things to do when a problem is detected.
@@ -4989,7 +5021,7 @@ sub watch_server {
[ qr/Could not parse relay log event entry/ => 'refetch_relay_log' ],
[ qr/Incorrect key file for table/ => 'repair_table' ],
# This must be the last one. It's a catch-all rule: skip and restart.
[ qr/./ => 'skip' ],
[ qr/./ => ($have_gtid ? 'skip_gtid' : 'skip') ],
);
# ########################################################################
@@ -5012,9 +5044,61 @@ sub watch_server {
},
skip => sub {
my ( $stat, $dbh ) = @_;
PTDEBUG && _d('Found non-relay-log error');
my $set_skip = $dbh->prepare("SET GLOBAL SQL_SLAVE_SKIP_COUNTER = "
. $o->get('skip-count'));
$set_skip->execute();
},
skip_gtid => sub {
my ( $stat, $dbh ) = @_;
# Get master_uuid from SHOW SLAVE STATUS if a UUID is not specified
# with --master-uuid.
my $gtid_uuid = $o->get('master-uuid');
if ( !$gtid_uuid ) {
$gtid_uuid = $stat->{master_uuid};
die "No master_uuid" unless $gtid_uuid; # shouldn't happen
}
# We need the highest transaction in the executed_gtid_set.
# and then we need to increase it by 1 (the one we want to skip)
# Notes:
# - does not work with parallel replication
# - it skips the next transaction from the master_uuid
# (when a slaveB is replicating from slaveA,
# the master_uuid is it's own master, slaveA)
my ($gtid_exec_ids) = ($stat->{executed_gtid_set} || '') =~ m/$gtid_uuid([0-9-:]*)/;
$gtid_exec_ids =~ s/:[0-9]-/:/g;
die "No executed GTIDs" unless $gtid_exec_ids;
my @gtid_exec_ranges = split(/:/, $gtid_exec_ids);
delete $gtid_exec_ranges[0]; # undef the first value, it's always empty
# Get the highest id by sorting the array, removing the undef value.
my @gtid_exec_sorted = sort { $a <=> $b }
grep { defined($_) } @gtid_exec_ranges;
my $gtid_exec_last = $gtid_exec_sorted[-1];
PTDEBUG && _d("\n",
"GTID: master_uuid:", $gtid_uuid, "\n",
"GTID: executed_gtid_set:", $gtid_exec_ids, "\n",
"GTID: max for master_uuid:", $gtid_exec_sorted[-1], "\n",
"GTID: last executed gtid:", $gtid_uuid, ":", $gtid_exec_last);
# Set the sessions next gtid, write an empty transaction
my $skipped = 0;
while ( $skipped++ < $o->get('skip-count') ) {
my $gtid_next = $gtid_exec_last + $skipped;
my $sql = "SET GTID_NEXT='$gtid_uuid:$gtid_next'";
PTDEBUG && _d($sql);
my $sth = $dbh->prepare($sql);
$sth->execute();
$dbh->begin_work();
$dbh->commit();
}
# Set the session back to the automatically generated GTID_NEXT.
$dbh->do("SET GTID_NEXT='AUTOMATIC'");
},
repair_table => sub {
my ( $stat, $dbh ) = @_;
PTDEBUG && _d('Found corrupt table');
@@ -5301,6 +5385,28 @@ sleep time, whichever is less.
=back
=head1 GLOBAL TRANSACTION IDS
As of Percona Toolkit 2.2.8, pt-slave-restart supports Global Transaction IDs
introduced in MySQL 5.6.5. It's important to keep in mind that:
=over
=item *
pt-slave-restart will not skip transactions when multiple replication threads
are being used (slave_parallel_workers > 0). pt-slave-restart does not know
what the GTID event is of the failed transaction of a specific slave thread.
=item *
The default behavior is to skip the next transaction from the slave's master.
Writes can originate on different servers, each with their own UUID.
See L<"--master-uuid">.
=back
=head1 EXIT STATUS
An exit status of 0 (sometimes also called a return value or return code)
@@ -5555,6 +5661,26 @@ type: int; default: 1
Number of statements to skip when restarting the slave.
=item --master-uuid
type: string
When using GTID, an empty transaction should be created in order to skip it.
If writes are coming from different nodes in the replication tree above, it is
not possible to know which event from which UUID to skip.
By default, transactions from the slave's master (C<'Master_UUID'> from
C<SHOW SLAVE STATUS>) are skipped.
For example, with
master1 -> slave1 -> slave2
When skipping events on slave2 that were written to master1, you must specify
the UUID of master1, else the tool will use the UUID of slave1 by default.
See L<"GLOBAL TRANSACTION IDS">.
=item --sleep
type: int; default: 1