Files
repmgr/doc/repmgrd-operation.xml
2020-05-15 10:33:47 +09:00

402 lines
17 KiB
XML

<chapter id="repmgrd-operation" xreflabel="repmgrd operation">
<title>repmgrd operation</title>
<indexterm>
<primary>repmgrd</primary>
<secondary>operation</secondary>
</indexterm>
<sect1 id="repmgrd-pausing" xreflabel="pausing the repmgrd service">
<title>Pausing the repmgrd service</title>
<indexterm>
<primary>repmgrd</primary>
<secondary>pausing</secondary>
</indexterm>
<indexterm>
<primary>pausing repmgrd</primary>
</indexterm>
<para>
In normal operation, &repmgrd; monitors the state of the
PostgreSQL node it is running on, and will take appropriate action if problems
are detected, e.g. (if so configured) promote the node to primary, if the existing
primary has been determined as failed.
</para>
<para>
However, &repmgrd; is unable to distinguish between
planned outages (such as performing a <link linkend="performing-switchover">switchover</link>
or installing PostgreSQL maintenance released), and an actual server outage. In versions prior to
&repmgr; 4.2 it was necessary to stop &repmgrd; on all nodes (or at least
on all nodes where &repmgrd; is
<link linkend="repmgrd-automatic-failover">configured for automatic failover</link>)
to prevent &repmgrd; from making unintentional changes to the
replication cluster.
</para>
<para>
From <link linkend="release-4.2">&repmgr; 4.2</link>, &repmgrd;
can now be &quot;paused&quot;, i.e. instructed not to take any action such as performing a failover.
This can be done from any node in the cluster, removing the need to stop/restart
each &repmgrd; individually.
</para>
<note>
<para>
For major PostgreSQL upgrades, e.g. from PostgreSQL 11 to PostgreSQL 12,
&repmgrd; should be shut down completely and only started up
once the &repmgr; packages for the new PostgreSQL major version have been installed.
</para>
</note>
<sect2 id="repmgrd-pausing-prerequisites">
<title>Prerequisites for pausing &repmgrd;</title>
<para>
In order to be able to pause/unpause &repmgrd;, following
prerequisites must be met:
<itemizedlist spacing="compact" mark="bullet">
<listitem>
<simpara><link linkend="release-4.2">&repmgr; 4.2</link> or later must be installed on all nodes.</simpara>
</listitem>
<listitem>
<simpara>The same major &repmgr; version (e.g. 4.2) must be installed on all nodes (and preferably the same minor version).</simpara>
</listitem>
<listitem>
<simpara>
PostgreSQL on all nodes must be accessible from the node where the
<literal>pause</literal>/<literal>unpause</literal> operation is executed, using the
<varname>conninfo</varname> string shown by <link linkend="repmgr-cluster-show"><command>repmgr cluster show</command></link>.
</simpara>
</listitem>
</itemizedlist>
</para>
<note>
<para>
These conditions are required for normal &repmgr; operation in any case.
</para>
</note>
</sect2>
<sect2 id="repmgrd-pausing-execution">
<title>Pausing/unpausing &repmgrd;</title>
<para>
To pause &repmgrd;, execute <link linkend="repmgr-service-pause"><command>repmgr service pause</command></link>
(&repmgr; 4.2 - 4.4: <link linkend="repmgr-service-pause"><command>repmgr daemon pause</command></link>),
e.g.:
<programlisting>
$ repmgr -f /etc/repmgr.conf service pause
NOTICE: node 1 (node1) paused
NOTICE: node 2 (node2) paused
NOTICE: node 3 (node3) paused</programlisting>
</para>
<para>
The state of &repmgrd; on each node can be checked with
<link linkend="repmgr-service-status"><command>repmgr service status</command></link>
(&repmgr; 4.2 - 4.4: <link linkend="repmgr-service-status"><command>repmgr daemon status</command></link>),
e.g.:
<programlisting>$ repmgr -f /etc/repmgr.conf service status
ID | Name | Role | Status | repmgrd | PID | Paused?
----+-------+---------+---------+---------+------+---------
1 | node1 | primary | running | running | 7851 | yes
2 | node2 | standby | running | running | 7889 | yes
3 | node3 | standby | running | running | 7918 | yes</programlisting>
</para>
<note>
<para>
If executing a switchover with <link linkend="repmgr-standby-switchover"><command>repmgr standby switchover</command></link>,
&repmgr; will automatically pause/unpause the &repmgrd; service as part of the switchover process.
</para>
</note>
<para>
If the primary (in this example, <literal>node1</literal>) is stopped, &repmgrd;
running on one of the standbys (here: <literal>node2</literal>) will react like this:
<programlisting>
[2019-08-28 12:22:21] [WARNING] unable to connect to upstream node "node1" (node ID: 1)
[2019-08-28 12:22:21] [INFO] checking state of node 1, 1 of 5 attempts
[2019-08-28 12:22:21] [INFO] sleeping 1 seconds until next reconnection attempt
...
[2019-08-28 12:22:24] [INFO] sleeping 1 seconds until next reconnection attempt
[2019-08-28 12:22:25] [INFO] checking state of node 1, 5 of 5 attempts
[2019-08-28 12:22:25] [WARNING] unable to reconnect to node 1 after 5 attempts
[2019-08-28 12:22:25] [NOTICE] node is paused
[2019-08-28 12:22:33] [INFO] node "node2" (ID: 2) monitoring upstream node "node1" (node ID: 1) in degraded state
[2019-08-28 12:22:33] [DETAIL] repmgrd paused by administrator
[2019-08-28 12:22:33] [HINT] execute "repmgr service unpause" to resume normal failover mode</programlisting>
</para>
<para>
If the primary becomes available again (e.g. following a software upgrade), &repmgrd;
will automatically reconnect, e.g.:
<programlisting>
[2019-08-28 12:25:41] [NOTICE] reconnected to upstream node "node1" (ID: 1) after 8 seconds, resuming monitoring</programlisting>
</para>
<para>
To unpause the &repmgrd; service, execute
<link linkend="repmgr-service-unpause"><command>repmgr service unpause</command></link>
((&repmgr; 4.2 - 4.4: <link linkend="repmgr-service-unpause"><command>repmgr daemon unpause</command></link>),
e.g.:
<programlisting>
$ repmgr -f /etc/repmgr.conf service unpause
NOTICE: node 1 (node1) unpaused
NOTICE: node 2 (node2) unpaused
NOTICE: node 3 (node3) unpaused</programlisting>
</para>
<note>
<para>
If the previous primary is no longer accessible when &repmgrd;
is unpaused, no failover action will be taken. Instead, a new primary must be manually promoted using
<link linkend="repmgr-standby-promote"><command>repmgr standby promote</command></link>,
and any standbys attached to the new primary with
<link linkend="repmgr-standby-follow"><command>repmgr standby follow</command></link>.
</para>
<para>
This is to prevent execution of <link linkend="repmgr-service-unpause"><command>repmgr service unpause</command></link>
resulting in the automatic promotion of a new primary, which may be a problem particularly
in larger clusters, where &repmgrd; could select a different promotion
candidate to the one intended by the administrator.
</para>
</note>
</sect2>
<sect2 id="repmgrd-pausing-details">
<title>Details on the &repmgrd; pausing mechanism</title>
<para>
The pause state of each node will be stored over a PostgreSQL restart.
</para>
<para>
<link linkend="repmgr-service-pause"><command>repmgr service pause</command></link> and
<link linkend="repmgr-service-unpause"><command>repmgr service unpause</command></link> can be
executed even if &repmgrd; is not running; in this case,
&repmgrd; will start up in whichever pause state has been set.
</para>
<note>
<para>
<link linkend="repmgr-service-pause"><command>repmgr service pause</command></link> and
<link linkend="repmgr-service-unpause"><command>repmgr service unpause</command></link>
<emphasis>do not</emphasis> start/stop &repmgrd;.
</para>
<para>
The commands <link linkend="repmgr-daemon-start"><command>repmgr daemon start</command></link>
and <link linkend="repmgr-daemon-stop"><command>repmgr daemon stop</command></link>
(<link linkend="repmgrd-service-configuration">if correctly configured</link>) can be used to start/stop
&repmgrd; on individual nodes.
</para>
</note>
</sect2>
</sect1>
<sect1 id="repmgrd-wal-replay-pause">
<title>repmgrd and paused WAL replay</title>
<indexterm>
<primary>repmgrd</primary>
<secondary>paused WAL replay</secondary>
</indexterm>
<para>
If WAL replay has been paused (using <command>pg_wal_replay_pause()</command>,
on PostgreSQL 9.6 and earlier <command>pg_xlog_replay_pause()</command>),
in a failover situation &repmgrd; will
automatically resume WAL replay.
</para>
<para>
This is because if WAL replay is paused, but WAL is pending replay,
PostgreSQL cannot be promoted until WAL replay is resumed.
</para>
<note>
<para>
<command><link linkend="repmgr-standby-promote">repmgr standby promote</link></command>
will refuse to promote a node in this state, as the PostgreSQL
<command>promote</command> command will not be acted on until
WAL replay is resumed, leaving the cluster in a potentially
unstable state. In this case it is up to the user to
decide whether to resume WAL replay.
</para>
</note>
</sect1>
<sect1 id="repmgrd-degraded-monitoring" xreflabel="repmgrd degraded monitoring">
<title>"degraded monitoring" mode</title>
<indexterm>
<primary>repmgrd</primary>
<secondary>degraded monitoring</secondary>
</indexterm>
<indexterm>
<primary>degraded monitoring</primary>
</indexterm>
<para>
In certain circumstances, &repmgrd; is not able to fulfill its primary mission
of monitoring the node's upstream server. In these cases it enters &quot;degraded monitoring&quot;
mode, where &repmgrd; remains active but is waiting for the situation
to be resolved.
</para>
<para>
Situations where this happens are:
<itemizedlist spacing="compact" mark="bullet">
<listitem>
<simpara>a failover situation has occurred, no nodes in the primary node's location are visible</simpara>
</listitem>
<listitem>
<simpara>a failover situation has occurred, but no promotion candidate is available</simpara>
</listitem>
<listitem>
<simpara>a failover situation has occurred, but the promotion candidate could not be promoted</simpara>
</listitem>
<listitem>
<simpara>a failover situation has occurred, but the node was unable to follow the new primary</simpara>
</listitem>
<listitem>
<simpara>a failover situation has occurred, but no primary has become available</simpara>
</listitem>
<listitem>
<simpara>a failover situation has occurred, but automatic failover is not enabled for the node</simpara>
</listitem>
<listitem>
<simpara>repmgrd is monitoring the primary node, but it is not available (and no other node has been promoted as primary)</simpara>
</listitem>
</itemizedlist>
</para>
<para>
Example output in a situation where there is only one standby with <literal>failover=manual</literal>,
and the primary node is unavailable (but is later restarted):
<programlisting>
[2017-08-29 10:59:19] [INFO] node "node2" (ID: 2) monitoring upstream node "node1" (ID: 1) in normal state (automatic failover disabled)
[2017-08-29 10:59:33] [WARNING] unable to connect to upstream node "node1" (ID: 1)
[2017-08-29 10:59:33] [INFO] checking state of node 1, 1 of 5 attempts
[2017-08-29 10:59:33] [INFO] sleeping 1 seconds until next reconnection attempt
(...)
[2017-08-29 10:59:37] [INFO] checking state of node 1, 5 of 5 attempts
[2017-08-29 10:59:37] [WARNING] unable to reconnect to node 1 after 5 attempts
[2017-08-29 10:59:37] [NOTICE] this node is not configured for automatic failover so will not be considered as promotion candidate
[2017-08-29 10:59:37] [NOTICE] no other nodes are available as promotion candidate
[2017-08-29 10:59:37] [HINT] use "repmgr standby promote" to manually promote this node
[2017-08-29 10:59:37] [INFO] node "node2" (ID: 2) monitoring upstream node "node1" (ID: 1) in degraded state (automatic failover disabled)
[2017-08-29 10:59:53] [INFO] node "node2" (ID: 2) monitoring upstream node "node1" (ID: 1) in degraded state (automatic failover disabled)
[2017-08-29 11:00:45] [NOTICE] reconnected to upstream node "node1" (ID: 1) after 68 seconds, resuming monitoring
[2017-08-29 11:00:57] [INFO] node "node2" (ID: 2) monitoring upstream node "node1" (ID: 1) in normal state (automatic failover disabled)</programlisting>
</para>
<para>
By default, <literal>repmgrd</literal> will continue in degraded monitoring mode indefinitely.
However a timeout (in seconds) can be set with <varname>degraded_monitoring_timeout</varname>,
after which &repmgrd; will terminate.
</para>
<note>
<para>
If &repmgrd; is monitoring a primary mode which has been stopped
and manually restarted as a standby attached to a new primary, it will automatically detect
the status change and update the node record to reflect the node's new status
as an active standby. It will then resume monitoring the node as a standby.
</para>
</note>
</sect1>
<sect1 id="repmgrd-monitoring" xreflabel="Storing monitoring data">
<title>Storing monitoring data</title>
<indexterm>
<primary>repmgrd</primary>
<secondary>monitoring</secondary>
</indexterm>
<indexterm>
<primary>monitoring</primary>
<secondary>with repmgrd</secondary>
</indexterm>
<para>
When &repmgrd; is running with the option <literal>monitoring_history=true</literal>,
it will constantly write standby node status information to the
<varname>monitoring_history</varname> table, providing a near-real time
overview of replication status on all nodes
in the cluster.
</para>
<para>
The view <literal>replication_status</literal> shows the most recent state
for each node, e.g.:
<programlisting>
repmgr=# select * from repmgr.replication_status;
-[ RECORD 1 ]-------------+------------------------------
primary_node_id | 1
standby_node_id | 2
standby_name | node2
node_type | standby
active | t
last_monitor_time | 2017-08-24 16:28:41.260478+09
last_wal_primary_location | 0/6D57A00
last_wal_standby_location | 0/5000000
replication_lag | 29 MB
replication_time_lag | 00:00:11.736163
apply_lag | 15 MB
communication_time_lag | 00:00:01.365643</programlisting>
</para>
<para>
The interval in which monitoring history is written is controlled by the
configuration parameter <varname>monitor_interval_secs</varname>;
default is 2.
</para>
<para>
As this can generate a large amount of monitoring data in the table
<literal>repmgr.monitoring_history</literal>. it's advisable to regularly
purge historical data using the <xref linkend="repmgr-cluster-cleanup"/>
command; use the <literal>-k/--keep-history</literal> option to
specify how many day's worth of data should be retained.
</para>
<para>
It's possible to use &repmgrd; to run in monitoring
mode only (without automatic failover capability) for some or all
nodes by setting <literal>failover=manual</literal> in the node's
<filename>repmgr.conf</filename> file. In the event of the node's upstream failing,
no failover action will be taken and the node will require manual intervention to
be reattached to replication. If this occurs, an
<link linkend="event-notifications">event notification</link>
<varname>standby_disconnect_manual</varname> will be created.
</para>
<para>
Note that when a standby node is not streaming directly from its upstream
node, e.g. recovering WAL from an archive, <varname>apply_lag</varname> will always appear as
<literal>0 bytes</literal>.
</para>
<tip>
<para>
If monitoring history is enabled, the contents of the <literal>repmgr.monitoring_history</literal>
table will be replicated to attached standbys. This means there will be a small but
constant stream of replication activity which may not be desirable. To prevent
this, convert the table to an <literal>UNLOGGED</literal> one with:
<programlisting>
ALTER TABLE repmgr.monitoring_history SET UNLOGGED;</programlisting>
</para>
<para>
This will however mean that monitoring history will not be available on
another node following a failover, and the view <literal>repmgr.replication_status</literal>
will not work on standbys.
</para>
</tip>
</sect1>
</chapter>