openstack
/
puppet-tripleo
Code Issues Proposed changes

Add option for innodb_flush_log_at_trx_commit = 2 for Galera only 

The innodb_flush_log_at_trx_commit flag changes the timing
of when the log buffer is written to disk for writes.
At its default of 1, transactions are written to disk
and the buffer flushed on a per-transaction basis; but when
set to 2, the flush of the buffer proceeds only once per
second.  This removes the durability guarantee for the
single node.  However the central concept of Galera is
that durability is achieved via the cluster as a whole,
in that transactions are replicated to other nodes before
the commit succeeds (though not necessarily written to disk
unless wsrep_causal_reads is set).  In this model,
data would only be lost of all nodes of the Galera cluster
were killed within one second of each other.  Percona's
blog post at https://www.percona.com/blog/2014/11/17/typical-misconceptions-on-galera-for-mysql/
recommends that the value of 2 should be considered "safe"
for a Galera cluster unless you are in fact worried that
all three nodes will be powered off simultaneously.

The value here is added as an option only, defaulting
to the usual default of "1", flush per transaction.

Change-Id: Id5a30f1daf978e094a74db2d284febbc9ae64bb3
This commit is contained in:
Mike Bayer 2017-07-03 10:35:30 -04:00
parent ad6c4e9c91
commit 8b2d404777
2 changed files with 61 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

79
manifests/profile/pacemaker/database/mysql.pp
View File

@ -52,6 +52,13 @@
# connections from other nodes in the cluster.
# Defaults to hiera('mysql_bind_host')
#
# [*innodb_flush_log_at_trx_commit*]
# (Optional) Disk flush behavior for MySQL under Galera. A value of
# '1' indicates flush to disk per transaction. A value of '2' indicates
# flush to disk every second, flushing all unflushed transactions in
# one step.
# Defaults to hiera('innodb_flush_log_at_trx_commit', '1')
#
# [*step*]
# (Optional) The current step in deployment. See tripleo-heat-templates
# for more details.
@ -62,14 +69,15 @@
# Defaults to hiera('pcs_tries', 20)
#
class tripleo::profile::pacemaker::database::mysql (
$bootstrap_node = hiera('mysql_short_bootstrap_node_name'),
$bind_address = $::hostname,
$ca_file = undef,
$certificate_specs = hiera('tripleo::profile::base::database::mysql::certificate_specs', {}),
$enable_internal_tls = hiera('enable_internal_tls', false),
$gmcast_listen_addr = hiera('mysql_bind_host'),
$step = Integer(hiera('step')),
$pcs_tries = hiera('pcs_tries', 20),
$bootstrap_node = hiera('mysql_short_bootstrap_node_name'),
$bind_address = $::hostname,
$ca_file = undef,
$certificate_specs = hiera('tripleo::profile::base::database::mysql::certificate_specs', {}),
$enable_internal_tls = hiera('enable_internal_tls', false),
$gmcast_listen_addr = hiera('mysql_bind_host'),
$innodb_flush_log_at_trx_commit = hiera('innodb_flush_log_at_trx_commit', '1'),
$step = Integer(hiera('step')),
$pcs_tries = hiera('pcs_tries', 20),
) {
if $::hostname == downcase($bootstrap_node) {
$pacemaker_master = true
@ -109,33 +117,34 @@ class tripleo::profile::pacemaker::database::mysql (
$mysqld_options = {
'mysqld' => {
'skip-name-resolve' => '1',
'binlog_format' => 'ROW',
'default-storage-engine' => 'innodb',
'innodb_autoinc_lock_mode' => '2',
'innodb_locks_unsafe_for_binlog'=> '1',
'innodb_file_per_table' => 'ON',
'query_cache_size' => '0',
'query_cache_type' => '0',
'bind-address' => $bind_address,
'max_connections' => hiera('mysql_max_connections'),
'open_files_limit' => '-1',
'wsrep_on' => 'ON',
'wsrep_provider' => '/usr/lib64/galera/libgalera_smm.so',
'wsrep_cluster_name' => 'galera_cluster',
'wsrep_cluster_address' => "gcomm://${galera_nodes}",
'wsrep_slave_threads' => '1',
'wsrep_certify_nonPK' => '1',
'wsrep_max_ws_rows' => '131072',
'wsrep_max_ws_size' => '1073741824',
'wsrep_debug' => '0',
'wsrep_convert_LOCK_to_trx' => '0',
'wsrep_retry_autocommit' => '1',
'wsrep_auto_increment_control' => '1',
'wsrep_drupal_282555_workaround'=> '0',
'wsrep_causal_reads' => '0',
'wsrep_sst_method' => 'rsync',
'wsrep_provider_options' => "gmcast.listen_addr=tcp://${gmcast_listen_addr}:4567;${tls_options}",
'skip-name-resolve' => '1',
'binlog_format' => 'ROW',
'default-storage-engine' => 'innodb',
'innodb_autoinc_lock_mode' => '2',
'innodb_locks_unsafe_for_binlog' => '1',
'innodb_file_per_table' => 'ON',
'innodb_flush_log_at_trx_commit' => $innodb_flush_log_at_trx_commit,
'query_cache_size' => '0',
'query_cache_type' => '0',
'bind-address' => $bind_address,
'max_connections' => hiera('mysql_max_connections'),
'open_files_limit' => '-1',
'wsrep_on' => 'ON',
'wsrep_provider' => '/usr/lib64/galera/libgalera_smm.so',
'wsrep_cluster_name' => 'galera_cluster',
'wsrep_cluster_address' => "gcomm://${galera_nodes}",
'wsrep_slave_threads' => '1',
'wsrep_certify_nonPK' => '1',
'wsrep_max_ws_rows' => '131072',
'wsrep_max_ws_size' => '1073741824',
'wsrep_debug' => '0',
'wsrep_convert_LOCK_to_trx' => '0',
'wsrep_retry_autocommit' => '1',
'wsrep_auto_increment_control' => '1',
'wsrep_drupal_282555_workaround' => '0',
'wsrep_causal_reads' => '0',
'wsrep_sst_method' => 'rsync',
'wsrep_provider_options' => "gmcast.listen_addr=tcp://${gmcast_listen_addr}:4567;${tls_options}",
}
}

17
releasenotes/notes/innodb_flush_log_at_trx_commit-eb7d99749ca3c911.yaml Normal file
View File

@ -0,0 +1,17 @@
---
features:
- |
Enable innodb_flush_log_at_trx_commit configuration for Galera only.
upgrade:
- |
Setting the innodb_flush_log_at_trx_commit flag to the value of "2" instead
of its default value of "1" means that the underlying MySQL/MariaDB engine
will no longer flush transactions to disk on a per-transaction basis;
instead, flushes occur once per second. This leads to far fewer
disk writes and can dramatically improve write performance, at the cost
of durability (e.g. will lose the last second's worth of transactions)
if the database engine is ungracefully shut down. The
clustered nature of Galera mitigates this risk in that transactions
are replicated to other nodes before completion, and the setting of
"2" is considered to be generally safe for a Galera cluster, with the
exception case of simultaneous power loss for all nodes.