As mentioned in the previous section, success of the switchover operation depends on repmgr being able to shut down the current primary server quickly and cleanly.
Ensure that the promotion candidate has sufficient free walsenders available
(PostgreSQL configuration item max_wal_senders), and if replication
slots are in use, at least one free slot is available for the demotion candidate (
PostgreSQL configuration item max_replication_slots).
Ensure that a passwordless SSH connection is possible from the promotion candidate
(standby) to the demotion candidate (current primary). If --siblings-follow
will be used, ensure that passwordless SSH connections are possible from the
promotion candidate to all nodes attached to the demotion candidate
(including the witness server, if in use).
repmgr expects to find the repmgr binary in the same path on the remote server as on the local server.
Double-check which commands will be used to stop/start/restart the current
primary; this can be done by e.g. executing repmgr node service
on the current primary:
repmgr -f /etc/repmgr.conf node service --list-actions --action=stop
repmgr -f /etc/repmgr.conf node service --list-actions --action=start
repmgr -f /etc/repmgr.conf node service --list-actions --action=restart
These commands can be defined in repmgr.conf with
service_start_command, service_stop_command
and service_restart_command.
If repmgr is installed from a package. you should set these commands to use the appropriate service commands defined by the package/operating system as these will ensure PostgreSQL is stopped/started properly taking into account configuration and log file locations etc.
If the service_*_command options aren't defined, repmgr will
fall back to using pg_ctl to stop/start/restart
PostgreSQL, which may not work properly, particularly when executed on a remote
server.
For more details, see service command settings.
On systemd systems we strongly recommend using the appropriate
systemctl commands (typically run via sudo) to ensure
systemd is informed about the status of the PostgreSQL service.
If using sudo for the systemctl calls, make sure the
sudo specification doesn't require a real tty for the user. If not set
this way, repmgr will fail to stop the primary.
See the service command settings documentation section for further details.
Check that access from applications is minimalized or preferably blocked completely, so applications are not unexpectedly interrupted.
If an exclusive backup is running on the current primary, or if WAL replay is paused on the standby, repmgr will not perform the switchover.
Check there is no significant replication lag on standbys attached to the current primary.
If WAL file archiving is set up, check that there is no backlog of files waiting
to be archived, as PostgreSQL will not finally shut down until all of these have been
archived. If there is a backlog exceeding archive_ready_warning WAL files,
repmgr will emit a warning before attempting to perform a switchover; you can also check
manually with repmgr node check --archive-ready.
From repmgr 4.2, repmgr will instruct any running repmgrd instances to pause operations while the switchover is being carried out, to prevent repmgrd from unintentionally promoting a node. For more details, see pausing the repmgrd service.
Users of repmgr versions prior to 4.2 should ensure that repmgrd is not running on any nodes while a switchover is being executed.
Finally, consider executing repmgr standby switchover with the
--dry-run option; this will perform any necessary checks and inform you about
success/failure, and stop before the first actual command is run (which would be the shutdown of the
current primary). Example output:
$ repmgr standby switchover -f /etc/repmgr.conf --siblings-follow --dry-run
NOTICE: checking switchover on node "node2" (ID: 2) in --dry-run mode
INFO: SSH connection to host "node1" succeeded
INFO: archive mode is "off"
INFO: replication lag on this standby is 0 seconds
INFO: all sibling nodes are reachable via SSH
NOTICE: local node "node2" (ID: 2) will be promoted to primary; current primary "node1" (ID: 1) will be demoted to standby
INFO: following shutdown command would be run on node "node1":
"pg_ctl -l /var/log/postgresql/startup.log -D '/var/lib/postgresql/data' -m fast -W stop"
INFO: parameter "shutdown_check_timeout" is set to 60 seconds
Be aware that --dry-run checks the prerequisites
for performing the switchover and some basic sanity checks on the
state of the database which might effect the switchover operation
(e.g. replication lag); it cannot however guarantee the switchover
operation will succeed. In particular, if the current primary
does not shut down cleanly, repmgr will not be able to reliably
execute the switchover (as there would be a danger of divergence
between the former and new primary nodes).
See repmgr standby switchover for a full list of available
command line options and repmgr.conf settings relevant
to performing a switchover.
If the demotion candidate does not shut down smoothly or cleanly, there's a risk it will have a slightly divergent timeline and will not be able to attach to the new primary. To fix this situation without needing to reclone the old primary, it's possible to use the pg_rewind utility, which will usually be able to resync the two servers.
To have repmgr execute pg_rewind if it detects this
situation after promoting the new primary, add the --force-rewind
option.
If repmgr detects a situation where it needs to execute pg_rewind,
it will execute a CHECKPOINT on the new primary before executing
pg_rewind.
For more details on pg_rewind, see: https://www.postgresql.org/docs/current/app-pgrewind.html.
pg_rewind has been part of the core PostgreSQL distribution since
version 9.5. Users of PostgreSQL 9.4 will need to manually install it; the source code is available here:
https://github.com/vmware/pg_rewind.
If the pg_rewind
binary is not installed in the PostgreSQL bin directory, provide
its full path on the demotion candidate with --force-rewind.
Note that building the 9.4 version of pg_rewind requires the PostgreSQL source code.