Upgrading PostgreSQL Using Slony

This guide describes the steps one can take to upgrade their PostgreSQL database to the latest version without the need for hours of downtime. This guide assumes you have two database servers: one database server running an older version of PostgreSQL (e.g. 9.2.18) and one server running a newer version (e.g. 9.6.0).

For this process we'll use a PostgreSQL replication tool called "Slony". Slony allows replication between different PostgreSQL versions and as such can be used to upgrade a cluster with a minimal amount of downtime.

In various places we'll refer to the user gitlab-psql. This user should be the user used to run the various PostgreSQL OS processes. If you're using a different user (e.g. postgres) you should replace gitlab-psql with the name of said user. This guide also assumes your database is called gitlabhq_production. If you happen to use a different database name you should change this accordingly.

Database Dumps

Slony only replicates data and not any schema changes. As a result we must ensure that all databases have the same database structure.

To do so we'll generate a dump of our current database. This dump will only contain the structure, not any data. To generate this dump run the following command on your active database server:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/pg_dump -h /var/opt/gitlab/postgresql -p 5432 -U gitlab-psql -s -f /tmp/structure.sql gitlabhq_production

If you're not using GitLab's Omnibus package you may have to adjust the paths to pg_dump and the PostgreSQL installation directory to match the paths of your configuration.

Once the structure dump is generated we also need to generate a dump for the schema_migrations table. This table doesn't have any primary keys and as such can't be replicated easily by Slony. To generate this dump run the following command on your active database server:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/pg_dump -h /var/opt/gitlab/postgresql/ -p 5432 -U gitlab-psql -a -t schema_migrations -f /tmp/migrations.sql gitlabhq_production

Next we'll need to move these files somewhere accessible by the new database server. The easiest way is to simply download these files to your local system:

scp your-user@production-database-host:/tmp/*.sql /tmp

This will copy all the SQL files located in /tmp to your local system's /tmp directory. Once copied you can safely remove the files from the database server.

Installing Slony

Slony will be used to upgrade the database without requiring long downtimes. Slony can be downloaded from https://www.slony.info/. If you have installed PostgreSQL using your operating system's package manager you may also be able to install Slony using said package manager.

When compiling Slony from source you must use the following commands to do so:

./configure --prefix=/path/to/installation/directory --with-perltools --with-pgconfigdir=/path/to/directory/containing/pg_config/bin
make
make install

Omnibus users can use the following commands:

./configure --prefix=/opt/gitlab/embedded --with-perltools --with-pgconfigdir=/opt/gitlab/embedded/bin
make
make install

This assumes you have installed GitLab into /opt/gitlab.

To test if Slony is installed properly, run the following commands:

test -f /opt/gitlab/embedded/bin/slonik && echo 'Slony installed' || echo 'Slony not installed'
test -f /opt/gitlab/embedded/bin/slonik_init_cluster && echo 'Slony Perl tools are available' || echo 'Slony Perl tools are not available'
/opt/gitlab/embedded/bin/slonik -v

This assumes Slony was installed to /opt/gitlab/embedded. If Slony was installed properly the output of these commands will be (the mentioned "slonik" version may be different):

Slony installed
Slony Perl tools are available
slonik version 2.2.5

Slony User

Next we must set up a PostgreSQL user that Slony can use to replicate your database. To do so, log in to your production database using psql using a super user account. Once done run the following SQL queries:

CREATE ROLE slony WITH SUPERUSER LOGIN REPLICATION ENCRYPTED PASSWORD 'password string here';
ALTER ROLE slony SET statement_timeout TO 0;

Make sure you replace "password string here" with the actual password for the user. A password is required. This user must be created on both the old and new database server using the same password.

Once the user has been created make sure you note down the password as we will need it later on.

Configuring Slony

Now we can finally start configuring Slony. Slony uses a configuration file for most of the work so we'll need to set this one up. This configuration file specifies where to put log files, how Slony should connect to the databases, etc.

First we'll need to create some required directories and set the correct permissions. To do so, run the following commands on both the old and new database server:

sudo mkdir -p /var/log/gitlab/slony /var/run/slony1 /var/opt/gitlab/postgresql/slony
sudo chown gitlab-psql:root /var/log/gitlab/slony /var/run/slony1 /var/opt/gitlab/postgresql/slony

Here gitlab-psql is the user used to run the PostgreSQL database processes. If you're using a different user you should replace this with the name of said user.

Now that the directories are in place we can create the configuration file. For this we can use the following template:

if ($ENV{"SLONYNODES"}) {
    require $ENV{"SLONYNODES"};
} else {
    $CLUSTER_NAME = 'slony_replication';
    $LOGDIR = '/var/log/gitlab/slony';
    $MASTERNODE = 1;
    $DEBUGLEVEL = 2;

    add_node(host => 'OLD_HOST', dbname => 'gitlabhq_production', port =>5432,
        user=>'slony', password=>'SLONY_PASSWORD', node=>1);

    add_node(host => 'NEW_HOST', dbname => 'gitlabhq_production', port =>5432,
        user=>'slony', password=>'SLONY_PASSWORD', node=>2, parent=>1 );
}

$SLONY_SETS = {
    "set1" => {
        "set_id"       => 1,
        "table_id"     => 1,
        "sequence_id"  => 1,
        "pkeyedtables" => [
            TABLES
        ],
    },
};

if ($ENV{"SLONYSET"}) {
    require $ENV{"SLONYSET"};
}

# Please do not add or change anything below this point.
1;

In this configuration file you should replace a few placeholders before you can use it. The following placeholders should be replaced:

  • OLD_HOST: the address of the old database server.
  • NEW_HOST: the address of the new database server.
  • SLONY_PASSWORD: the password of the Slony user created earlier.
  • TABLES: the tables to replicate.

The list of tables to replicate can be generated by running the following command on your old PostgreSQL database:

sudo gitlab-psql gitlabhq_production -c "select concat('\"', schemaname, '.', tablename, '\",') from pg_catalog.pg_tables where schemaname = 'public' and tableowner = 'gitlab' and tablename != 'schema_migrations' order by tablename asc;" -t

If you're not using Omnibus you should replace gitlab-psql with the appropriate path to the psql executable.

The above command outputs a list of tables in a format that can be copy-pasted directly into the above configuration file. Make sure to replace TABLES with this output, don't just append it below it. Once done you'll end up with something like this:

"pkeyedtables" => [
    "public.abuse_reports",
    "public.appearances",
    "public.application_settings",
    ... more rows here ...
]

Once you have the configuration file generated you must install it on both the old and new database. To do so, place it in /var/opt/gitlab/postgresql/slony/slon_tools.conf (for which we created the directory earlier on).

Now that the configuration file is in place we can finally start replicating our database. First we must set up the schema in our new database. To do so make sure that the SQL files we generated earlier can be found in the /tmp directory of the new server. Once these files are in place start a psql session on this server:

sudo gitlab-psql gitlabhq_production

Now run the following commands:

\i /tmp/structure.sql
\i /tmp/migrations.sql

To verify if the structure is in place close the session, start it again, then run \d. If all went well you should see output along the lines of the following:

                               List of relations
 Schema |                    Name                     |   Type   |    Owner
--------+---------------------------------------------+----------+-------------
 public | abuse_reports                               | table    | gitlab
 public | abuse_reports_id_seq                        | sequence | gitlab
 public | appearances                                 | table    | gitlab
 public | appearances_id_seq                          | sequence | gitlab
 public | application_settings                        | table    | gitlab
 public | application_settings_id_seq                 | sequence | gitlab
 public | approvals                                   | table    | gitlab
 ... more rows here ...

Now we can initialize the required tables and what not that Slony will use for its replication process. To do so, run the following on the old database:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slonik_init_cluster --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf | /opt/gitlab/embedded/bin/slonik

If all went well this will produce something along the lines of:

<stdin>:10: Set up replication nodes
<stdin>:13: Next: configure paths for each node/origin
<stdin>:16: Replication nodes prepared
<stdin>:17: Please start a slon replication daemon for each node

Next we need to start a replication node on every server. To do so, run the following on the old database:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slon_start 1 --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf

If all went well this will produce output such as:

Invoke slon for node 1 - /opt/gitlab/embedded/bin/slon -p /var/run/slony1/slony_replication_node1.pid -s 1000 -d2  slony_replication 'host=192.168.0.7 dbname=gitlabhq_production user=slony port=5432 password=hieng8ezohHuCeiqu0leeghai4aeyahp' > /var/log/gitlab/slony/node1/gitlabhq_production-2016-10-06.log 2>&1 &
Slon successfully started for cluster slony_replication, node node1
PID [26740]
Start the watchdog process as well...

Next we need to run the following command on the new database server:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slon_start 2 --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf

This will produce similar output if all went well.

Next we need to tell the new database server what it should replicate. This can be done by running the following command on the new database server:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slonik_create_set 1 --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf | /opt/gitlab/embedded/bin/slonik

This should produce output along the lines of the following:

<stdin>:11: Subscription set 1 (set1) created
<stdin>:12: Adding tables to the subscription set
<stdin>:16: Add primary keyed table public.abuse_reports
<stdin>:20: Add primary keyed table public.appearances
<stdin>:24: Add primary keyed table public.application_settings
... more rows here ...
<stdin>:327: Adding sequences to the subscription set
<stdin>:328: All tables added

Finally we can start the replication process by running the following on the new database server:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slonik_subscribe_set 1 2 --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf | /opt/gitlab/embedded/bin/slonik

This should produce the following output:

<stdin>:6: Subscribed nodes to set 1

At this point the new database server will start replicating the data of the old database server. This process can take anywhere from a few minutes to hours, if not days. Unfortunately Slony itself doesn't really provide a way of knowing when the two databases are in sync. To get an estimate of the progress you can use the following shell script:

#!/usr/bin/env bash

set -e

user='slony'
pass='SLONY_PASSWORD'

function main {
    while :
    do
        local source
        local target

        source=$(PGUSER="${user}" PGPASSWORD="${pass}" /opt/gitlab/embedded/bin/psql -h OLD_HOST gitlabhq_production -c "select pg_size_pretty(pg_database_size('gitlabhq_production'));" -t -A)
        target=$(PGUSER="${user}" PGPASSWORD="${pass}" /opt/gitlab/embedded/bin/psql -h NEW_HOST gitlabhq_production -c "select pg_size_pretty(pg_database_size('gitlabhq_production'));" -t -A)

        echo "$(date): ${target} of ${source}" >> progress.log
        echo "$(date): ${target} of ${source}"

        sleep 60
    done
}

main

This script will compare the sizes of the old and new database every minute and print the result to STDOUT as well as logging it to a file. Make sure to replace SLONY_PASSWORD, OLD_HOST, and NEW_HOST with the correct values.

Stopping Replication

At some point the two databases are in sync. Once this is the case you'll need to plan for a few minutes of downtime. This small downtime window is used to stop the replication process, remove any Slony data from both databases, restart GitLab so it can use the new database, etc.

First, let's stop all of GitLab. Omnibus users can do so by running the following on their GitLab server(s):

sudo gitlab-ctl stop unicorn
sudo gitlab-ctl stop sidekiq
sudo gitlab-ctl stop mailroom

If you have any other processes that use PostgreSQL you should also stop those.

Once everything has been stopped you should update any configuration settings, DNS records, etc so they all point to the new database.

Once the settings have been taken care of we need to stop the replication process. It's crucial that no new data is written to the databases at this point as this data will be lost.

To stop replication, run the following on both database servers:

sudo -u gitlab-psql /opt/gitlab/embedded/bin/slon_kill --conf /var/opt/gitlab/postgresql/slony/slon_tools.conf

This will stop all the Slony processes on the host the command was executed on.

Resetting Sequences

The above setup does not replicate database sequences, as such these must be reset manually in the target database. You can use the following script for this:

#!/usr/bin/env bash
set -e

function main {
    local fix_sequences
    local fix_owners

    fix_sequences='/tmp/fix_sequences.sql'
    fix_owners='/tmp/fix_owners.sql'

    # The SQL queries were taken from
    # https://wiki.postgresql.org/wiki/Fixing_Sequences
    sudo gitlab-psql gitlabhq_production -t -c "
    SELECT 'ALTER SEQUENCE '|| quote_ident(MIN(schema_name)) ||'.'|| quote_ident(MIN(seq_name))
           ||' OWNED BY '|| quote_ident(MIN(TABLE_NAME)) ||'.'|| quote_ident(MIN(column_name)) ||';'
    FROM (
        SELECT
            n.nspname AS schema_name,
            c.relname AS TABLE_NAME,
            a.attname AS column_name,
            SUBSTRING(d.adsrc FROM E'^nextval\\(''([^'']*)''(?:::text|::regclass)?\\)') AS seq_name
        FROM pg_class c
        JOIN pg_attribute a ON (c.oid=a.attrelid)
        JOIN pg_attrdef d ON (a.attrelid=d.adrelid AND a.attnum=d.adnum)
        JOIN pg_namespace n ON (c.relnamespace=n.oid)
        WHERE has_schema_privilege(n.oid,'USAGE')
          AND n.nspname NOT LIKE 'pg!_%' escape '!'
          AND has_table_privilege(c.oid,'SELECT')
          AND (NOT a.attisdropped)
          AND d.adsrc ~ '^nextval'
    ) seq
    GROUP BY seq_name HAVING COUNT(*)=1;
    " > "${fix_owners}"

    sudo gitlab-psql gitlabhq_production -t -c "
    SELECT 'SELECT SETVAL(' ||
           quote_literal(quote_ident(PGT.schemaname) || '.' || quote_ident(S.relname)) ||
           ', COALESCE(MAX(' ||quote_ident(C.attname)|| '), 1) ) FROM ' ||
           quote_ident(PGT.schemaname)|| '.'||quote_ident(T.relname)|| ';'
    FROM pg_class AS S,
         pg_depend AS D,
         pg_class AS T,
         pg_attribute AS C,
         pg_tables AS PGT
    WHERE S.relkind = 'S'
        AND S.oid = D.objid
        AND D.refobjid = T.oid
        AND D.refobjid = C.attrelid
        AND D.refobjsubid = C.attnum
        AND T.relname = PGT.tablename
    ORDER BY S.relname;
    " > "${fix_sequences}"

    sudo gitlab-psql gitlabhq_production -f "${fix_owners}"
    sudo gitlab-psql gitlabhq_production -f "${fix_sequences}"

    rm "${fix_owners}" "${fix_sequences}"
}

main

Upload this script to the target server and execute it as follows:

bash path/to/the/script/above.sh

This will correct the ownership of sequences and reset the next value for the id column to the next available value.

Removing Slony

Next we need to remove all Slony related data. To do so, run the following command on the target server:

sudo gitlab-psql gitlabhq_production -c "DROP SCHEMA _slony_replication CASCADE;"

Once done you can safely remove any Slony related files (e.g. the log directory), and uninstall Slony if desired. At this point you can start your GitLab instance again and if all went well it should be using your new database server.