Operating pg_auto_failover ========================== This section is not yet complete. Please contact us with any questions. Deployment ---------- pg_auto_failover is a general purpose tool for setting up PostgreSQL replication in order to implement High Availability of the PostgreSQL service. Provisioning ------------ It is also possible to register pre-existing PostgreSQL instances with a pg_auto_failover monitor. The ``pg_autoctl create`` command honors the ``PGDATA`` environment variable, and checks whether PostgreSQL is already running. If Postgres is detected, the new node is registered in SINGLE mode, bypassing the monitor's role assignment policy. Postgres configuration management --------------------------------- The :ref:`pg_autoctl_create_postgres` command edits the default Postgres configuration file (``postgresql.conf``) to include pg_auto_failover settings. The include directive is placed on the top of the ``postgresql.conf`` file in a way that you may override any setting by editing it later in the file. Unless using the ``--skip-pg-hba`` option then pg_autoctl edits a minimal set of HBA rules for you, in order for the pg_auto_failover nodes to be able to connect to each other. The HBA rules that are needed for your application to connect to your Postgres nodes still need to be added. As pg_autoctl knows nothing about your applications, then you are responsible for editing the HBA file. Upgrading pg_auto_failover, from versions 1.4 onward ----------------------------------------------------- When upgrading a pg_auto_failover setup, the procedure is different on the monitor and on the Postgres nodes: - on the monitor, the internal pg_auto_failover database schema might have changed and needs to be upgraded to its new definition, porting the existing data over. The pg_auto_failover database contains the registration of every node in the system and their current state. It is not possible to trigger a failover during the monitor update. Postgres operations on the Postgres nodes continue normally. During the restart of the monitor, the other nodes might have trouble connecting to the monitor. The ``pg_autoctl`` command is designed to retry connecting to the monitor and handle errors gracefully. - on the Postgres nodes, the ``pg_autoctl`` command connects to the monitor every once in a while (every second by default), and then calls the ``node_active`` protocol, a stored procedure in the monitor databases. The ``pg_autoctl`` also verifies at each connection to the monitor that it's running the expected version of the extension. When that's not the case, the "node-active" sub-process quits, to be restarted with the possibly new version of the ``pg_autoctl`` binary found on-disk. As a result, here is the standard upgrade plan for pg_auto_failover: 1. Upgrade the pg_auto_failover package on the all the nodes, monitor included. When using a debian based OS, this looks like the following command when from 1.4 to 1.5:: sudo apt-get remove pg-auto-failover-cli-1.4 postgresql-11-auto-failover-1.4 sudo apt-get install -q -y pg-auto-failover-cli-1.5 postgresql-11-auto-failover-1.5 2. Restart the ``pgautofailover`` service on the monitor. When using the systemd integration, all we need to do is:: sudo systemctl restart pgautofailover Then we may use the following commands to make sure that the service is running as expected:: sudo systemctl status pgautofailover sudo journalctl -u pgautofailover At this point it is expected that the ``pg_autoctl`` logs show that an upgrade has been performed by using the ``ALTER EXTENSION pgautofailover UPDATE TO ...`` command. The monitor is ready with the new version of pg_auto_failover. When the Postgres nodes ``pg_autoctl`` process connects to the new monitor version, the check for version compatibility fails, and the "node-active" sub-process exits. The main ``pg_autoctl`` process supervisor then restart the "node-active" sub-process from its on-disk binary executable file, which has been upgraded to the new version. That's why we first install the new packages for pg_auto_failover on every node, and only then restart the monitor. .. important:: Before upgrading the monitor, which is a simple restart of the ``pg_autoctl`` process, it is important that the OS packages for pgautofailover be updated on all the Postgres nodes. When that's not the case, ``pg_autoctl`` on the Postgres nodes will still detect a version mismatch with the monitor extension, and the "node-active" sub-process will exit. And when restarted automatically, the same version of the local ``pg_autoctl`` binary executable is found on-disk, leading to the same version mismatch with the monitor extension. After restarting the "node-active" process 5 times, ``pg_autoctl`` quits retrying and stops. This includes stopping the Postgres service too, and a service downtime might then occur. And when the upgrade is done we can use ``pg_autoctl show state`` on the monitor to see that eveything is as expected. Upgrading from previous pg_auto_failover versions ------------------------------------------------- The new upgrade procedure described in the previous section is part of pg_auto_failover since version 1.4. When upgrading from a previous version of pg_auto_failover, up to and including version 1.3, then all the ``pg_autoctl`` processes have to be restarted fully. To prevent triggering a failover during the upgrade, it's best to put your secondary nodes in maintenance. The procedure then looks like the following: 1. Enable maintenance on your secondary node(s):: pg_autoctl enable maintenance 2. Upgrade the OS packages for pg_auto_failover on every node, as per previous section. 3. Restart the monitor to upgrade it to the new pg_auto_failover version: When using the systemd integration, all we need to do is:: sudo systemctl restart pgautofailover Then we may use the following commands to make sure that the service is running as expected:: sudo systemctl status pgautofailover sudo journalctl -u pgautofailover At this point it is expected that the ``pg_autoctl`` logs show that an upgrade has been performed by using the ``ALTER EXTENSION pgautofailover UPDATE TO ...`` command. The monitor is ready with the new version of pg_auto_failover. 4. Restart ``pg_autoctl`` on all Postgres nodes on the cluster. When using the systemd integration, all we need to do is:: sudo systemctl restart pgautofailover As in the previous point in this list, make sure the service is now running as expected. 5. Disable maintenance on your secondary nodes(s):: pg_autoctl disable maintenance Extension dependencies when upgrading the monitor ------------------------------------------------- Since version 1.4.0 the ``pgautofailover`` extension requires the Postgres contrib extension ``btree_gist``. The ``pg_autoctl`` command arranges for the creation of this dependency, and has been buggy in some releases. As a result, you might have trouble upgrade the pg_auto_failover monitor to a recent version. It is possible to fix the error by connecting to the monitor Postgres database and running the ``create extension`` command manually:: # create extension btree_gist; Cluster Management and Operations --------------------------------- It is possible to operate pg_auto_failover formations and groups directly from the monitor. All that is needed is an access to the monitor Postgres database as a client, such as ``psql``. It's also possible to add those management SQL function calls in your own ops application if you have one. For security reasons, the ``autoctl_node`` is not allowed to perform maintenance operations. This user is limited to what ``pg_autoctl`` needs. You can either create a specific user and authentication rule to expose for management, or edit the default HBA rules for the ``autoctl`` user. In the following examples we're directly connecting as the ``autoctl`` role. The main operations with pg_auto_failover are node maintenance and manual failover, also known as a controlled switchover. Maintenance of a secondary node ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ It is possible to put a secondary node in any group in a MAINTENANCE state, so that the Postgres server is not doing *synchronous replication* anymore and can be taken down for maintenance purposes, such as security kernel upgrades or the like. The command line tool ``pg_autoctl`` exposes an API to schedule maintenance operations on the current node, which must be a secondary node at the moment when maintenance is requested. Here's an example of using the maintenance commands on a secondary node, including the output. Of course, when you try that on your own nodes, dates and PID information might differ:: $ pg_autoctl enable maintenance 17:49:19 14377 INFO Listening monitor notifications about state changes in formation "default" and group 0 17:49:19 14377 INFO Following table displays times when notifications are received Time | ID | Host | Port | Current State | Assigned State ---------+-----+-----------+--------+---------------------+-------------------- 17:49:19 | 1 | localhost | 5001 | primary | wait_primary 17:49:19 | 2 | localhost | 5002 | secondary | wait_maintenance 17:49:19 | 2 | localhost | 5002 | wait_maintenance | wait_maintenance 17:49:20 | 1 | localhost | 5001 | wait_primary | wait_primary 17:49:20 | 2 | localhost | 5002 | wait_maintenance | maintenance 17:49:20 | 2 | localhost | 5002 | maintenance | maintenance The command listens to the state changes in the current node's formation and group on the monitor and displays those changes as it receives them. The operation is done when the node has reached the ``maintenance`` state. It is now possible to disable maintenance to allow ``pg_autoctl`` to manage this standby node again:: $ pg_autoctl disable maintenance 17:49:26 14437 INFO Listening monitor notifications about state changes in formation "default" and group 0 17:49:26 14437 INFO Following table displays times when notifications are received Time | ID | Host | Port | Current State | Assigned State ---------+-----+-----------+--------+---------------------+-------------------- 17:49:27 | 2 | localhost | 5002 | maintenance | catchingup 17:49:27 | 2 | localhost | 5002 | catchingup | catchingup 17:49:28 | 2 | localhost | 5002 | catchingup | secondary 17:49:28 | 1 | localhost | 5001 | wait_primary | primary 17:49:28 | 2 | localhost | 5002 | secondary | secondary 17:49:29 | 1 | localhost | 5001 | primary | primary When a standby node is in maintenance, the monitor sets the primary node replication to WAIT_PRIMARY: in this role, the PostgreSQL streaming replication is now asynchronous and the standby PostgreSQL server may be stopped, rebooted, etc. Maintenance of a primary node ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ A primary node must be available at all times in any formation and group in pg_auto_failover, that is the invariant provided by the whole solution. With that in mind, the only way to allow a primary node to go to a maintenance mode is to first failover and promote the secondary node. The same command ``pg_autoctl enable maintenance`` implements that operation when run on a primary node with the option ``--allow-failover``. Here is an example of such an operation:: $ pg_autoctl enable maintenance 11:53:03 50526 WARN Enabling maintenance on a primary causes a failover 11:53:03 50526 FATAL Please use --allow-failover to allow the command proceed As we can see the option ``allow-failover`` is mandatory. In the next example we use it:: $ pg_autoctl enable maintenance --allow-failover 13:13:42 1614 INFO Listening monitor notifications about state changes in formation "default" and group 0 13:13:42 1614 INFO Following table displays times when notifications are received Time | ID | Host | Port | Current State | Assigned State ---------+-----+-----------+--------+---------------------+-------------------- 13:13:43 | 2 | localhost | 5002 | primary | prepare_maintenance 13:13:43 | 1 | localhost | 5001 | secondary | prepare_promotion 13:13:43 | 1 | localhost | 5001 | prepare_promotion | prepare_promotion 13:13:43 | 2 | localhost | 5002 | prepare_maintenance | prepare_maintenance 13:13:44 | 1 | localhost | 5001 | prepare_promotion | stop_replication 13:13:45 | 1 | localhost | 5001 | stop_replication | stop_replication 13:13:46 | 1 | localhost | 5001 | stop_replication | wait_primary 13:13:46 | 2 | localhost | 5002 | prepare_maintenance | maintenance 13:13:46 | 1 | localhost | 5001 | wait_primary | wait_primary 13:13:47 | 2 | localhost | 5002 | maintenance | maintenance When the operation is done we can have the old primary re-join the group, this time as a secondary:: $ pg_autoctl disable maintenance 13:14:46 1985 INFO Listening monitor notifications about state changes in formation "default" and group 0 13:14:46 1985 INFO Following table displays times when notifications are received Time | ID | Host | Port | Current State | Assigned State ---------+-----+-----------+--------+---------------------+-------------------- 13:14:47 | 2 | localhost | 5002 | maintenance | catchingup 13:14:47 | 2 | localhost | 5002 | catchingup | catchingup 13:14:52 | 2 | localhost | 5002 | catchingup | secondary 13:14:52 | 1 | localhost | 5001 | wait_primary | primary 13:14:52 | 2 | localhost | 5002 | secondary | secondary 13:14:53 | 1 | localhost | 5001 | primary | primary Triggering a failover ^^^^^^^^^^^^^^^^^^^^^ It is possible to trigger a manual failover, or a switchover, using the command ``pg_autoctl perform failover``. Here's an example of what happens when running the command:: $ pg_autoctl perform failover 11:58:00 53224 INFO Listening monitor notifications about state changes in formation "default" and group 0 11:58:00 53224 INFO Following table displays times when notifications are received Time | ID | Host | Port | Current State | Assigned State ---------+-----+-----------+--------+--------------------+------------------- 11:58:01 | 1 | localhost | 5001 | primary | draining 11:58:01 | 2 | localhost | 5002 | secondary | prepare_promotion 11:58:01 | 1 | localhost | 5001 | draining | draining 11:58:01 | 2 | localhost | 5002 | prepare_promotion | prepare_promotion 11:58:02 | 2 | localhost | 5002 | prepare_promotion | stop_replication 11:58:02 | 1 | localhost | 5001 | draining | demote_timeout 11:58:03 | 1 | localhost | 5001 | demote_timeout | demote_timeout 11:58:04 | 2 | localhost | 5002 | stop_replication | stop_replication 11:58:05 | 2 | localhost | 5002 | stop_replication | wait_primary 11:58:05 | 1 | localhost | 5001 | demote_timeout | demoted 11:58:05 | 2 | localhost | 5002 | wait_primary | wait_primary 11:58:05 | 1 | localhost | 5001 | demoted | demoted 11:58:06 | 1 | localhost | 5001 | demoted | catchingup 11:58:06 | 1 | localhost | 5001 | catchingup | catchingup 11:58:08 | 1 | localhost | 5001 | catchingup | secondary 11:58:08 | 2 | localhost | 5002 | wait_primary | primary 11:58:08 | 1 | localhost | 5001 | secondary | secondary 11:58:08 | 2 | localhost | 5002 | primary | primary Again, timings and PID numbers are not expected to be the same when you run the command on your own setup. Also note in the output that the command shows the whole set of transitions including when the old primary is now a secondary node. The database is available for read-write traffic as soon as we reach the state ``wait_primary``. Implementing a controlled switchover ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ It is generally useful to distinguish a *controlled switchover* to a *failover*. In a controlled switchover situation it is possible to organise the sequence of events in a way to avoid data loss and lower downtime to a minimum. In the case of pg_auto_failover, because we use **synchronous replication**, we don't face data loss risks when triggering a manual failover. Moreover, our monitor knows the current primary health at the time when the failover is triggered, and drives the failover accordingly. So to trigger a controlled switchover with pg_auto_failover you can use the same API as for a manual failover:: $ pg_autoctl perform switchover Because the subtelties of orchestrating either a controlled switchover or an unplanned failover are all handled by the monitor, rather than the client side command line, at the client level the two command ``pg_autoctl perform failover`` and ``pg_autoctl perform switchover`` are synonyms, or aliases. Current state, last events -------------------------- The following commands display information from the pg_auto_failover monitor tables ``pgautofailover.node`` and ``pgautofailover.event``: :: $ pg_autoctl show state $ pg_autoctl show events When run on the monitor, the commands outputs all the known states and events for the whole set of formations handled by the monitor. When run on a PostgreSQL node, the command connects to the monitor and outputs the information relevant to the service group of the local node only. For interactive debugging it is helpful to run the following command from the monitor node while e.g. initializing a formation from scratch, or performing a manual failover:: $ watch pg_autoctl show state Monitoring pg_auto_failover in Production ----------------------------------------- The monitor reports every state change decision to a LISTEN/NOTIFY channel named ``state``. PostgreSQL logs on the monitor are also stored in a table, ``pgautofailover.event``, and broadcast by NOTIFY in the channel ``log``. .. _replacing_monitor_online: Replacing the monitor online ---------------------------- When the monitor node is not available anymore, it is possible to create a new monitor node and then switch existing nodes to a new monitor by using the following commands. 1. Apply the STONITH approach on the old monitor to make sure this node is not going to show up again during the procedure. This step is sometimes referred to as “fencing”. 2. On every node, ending with the (current) Postgres primary node for each group, disable the monitor while ``pg_autoctl`` is still running:: $ pg_autoctl disable monitor --force 3. Create a new monitor node:: $ pg_autoctl create monitor ... 4. On the current primary node first, so that it's registered first and as a primary still, for each group in your formation(s), enable the monitor online again:: $ pg_autoctl enable monitor postgresql://autoctl_node@.../pg_auto_failover 5. On every other (secondary) node, enable the monitor online again:: $ pg_autoctl enable monitor postgresql://autoctl_node@.../pg_auto_failover See :ref:`pg_autoctl_disable_monitor` and :ref:`pg_autoctl_enable_monitor` for details about those commands. This operation relies on the fact that a ``pg_autoctl`` can be operated without a monitor, and when reconnecting to a new monitor, this process reset the parts of the node state that comes from the monitor, such as the node identifier. Trouble-Shooting Guide ---------------------- pg_auto_failover commands can be run repeatedly. If initialization fails the first time -- for instance because a firewall rule hasn't yet activated -- it's possible to try ``pg_autoctl create`` again. pg_auto_failover will review its previous progress and repeat idempotent operations (``create database``, ``create extension`` etc), gracefully handling errors.