Prepare PostgreSQL for RDI

Prepare PostgreSQL databases to work with RDI

PostgreSQL supports several logical decoding plug-ins to enable CDC. If you don't want to use the native pgoutput logical replication stream support then you must install your preferred plug-in into the PostgreSQL server. Once you have done this, you must enable a replication slot, and configure a user with privileges to perform the replication.

If you are using a service like Heroku Postgres to host your database then this might restrict the plug-ins you can use. If you can't use your preferred plug-in then could try the pgoutput decoder if you are using PostgreSQL 10 or above. If this doesn't work for you then you won't be able to use RDI with your database.

The following checklist summarizes the steps to prepare a PostgreSQL database for RDI, with links to the sections that explain the steps in full detail. You may find it helpful to track your progress with the checklist as you complete each step.

- [ ] [Install the logical decoding output plug-in](#install-the-logical-decoding-output-plug-in)
- [ ] [Configure the PostgreSQL server](#configure-the-postgresql-server)
- [ ] [Set up permissions](#set-up-permissions)
- [ ] [Check replica identity and choose a publication mode](#check-replica-identity-and-choose-a-publication-mode)
- [ ] [Set privileges for Debezium to create PostgreSQL publications with pgoutput](#set-privileges-for-debezium-to-create-postgresql-publications-with-pgoutput)
- [ ] [Configure PostgreSQL for replication with the Debezium connector host](#configure-postgresql-for-replication-with-the-debezium-connector-host)

Amazon RDS for PostgreSQL

Follow the steps below to enable CDC with Amazon RDS for PostgreSQL:

1. Set the instance parameter rds.logical_replication to 1.

2. Check that the wal_level parameter is set to logical by running the query SHOW wal_level as the database RDS master user. The parameter might not have this value in multi-zone replication setups. You can't change the value manually but it should change automatically when you set the rds.logical_replication parameter to 1. If it doesn't change then you probably just need to restart your database instance. You can restart manually or wait until a restart occurs during your maintenance window.

3. Set the Debezium plugin.name parameter to pgoutput.

4. Initiate logical replication from an AWS account that has the rds_replication role. The role grants permissions to manage logical slots and to stream data using logical slots. By default, only the master user account on AWS has the rds_replication role on Amazon RDS, but if you have administrator privileges, you can grant the role to other accounts using a query like the following:

   GRANT rds_replication TO <my_user>
   

   To enable accounts other than the master account to create an initial snapshot, you must grant SELECT permission to the accounts on the tables to be captured. See the documentation about security for PostgreSQL logical replication for more information.

Azure Database for PostgreSQL

If you are using Azure Database for PostgreSQL you need to manually set the wal_level parameter in the Azure portal for your PostgreSQL server. Go to the server parameters section, search for wal_level and set it to logical. Then save and restart the server.

Install the logical decoding output plug-in

As of PostgreSQL 9.4, the only way to read changes to the write-ahead-log is to install a logical decoding output plug-in. These plug-ins are written in C using PostgreSQL-specific APIs, as described in the PostgreSQL documentation. The PostgreSQL connector uses one of Debezium’s supported logical decoding plug-ins to receive change events from the database in either the default pgoutput format (supplied with PostgreSQL) or the Protobuf format. See the decoderbufs Protobuf plug-in documentation for more details about how to compile it and also its requirements and limitations.

For simplicity, Debezium also provides a container image that compiles and installs the plug-ins on top of the upstream PostgreSQL server image. Use this image as an example of the steps involved in the installation.

Plug-in differences

Plug-ins don't all behave in exactly the same way. All of them refresh information about the database schema when they detect that it has changed, but the pgoutput plug-in is more "eager" than some other plug-ins to do this. For example, pgoutput will refresh when it detects a change to the default value of a column but other plug-ins won't notice this until another, more significant change happens (such as adding a new table column).

The Debezium project maintains a Java class that tracks the known differences between plug-ins.

Configure the PostgreSQL server

If you want to use a logical decoding plug-in other than the default pgoutput then you must first configure it in the postgresql.conf file. Set the shared_preload_libraries parameter to load your plug-in at startup. For example, to load the decoderbufs plug-in, you would add the following line:

# MODULES
shared_preload_libraries = 'decoderbufs'

Add the line below to configure the replication slot (for any plug-in). This instructs the server to use logical decoding with the write-ahead log.

# REPLICATION
wal_level = logical            

You can also set other PostgreSQL streaming replication parameters if you need them. For example, you can use max_wal_senders and max_replication_slots to increase the number of connectors that can access the sending server concurrently, and wal_keep_size to limit the maximum WAL size that a replication slot retains. The configuration parameters documentation describes all the parameters you can use.

PostgreSQL’s logical decoding uses replication slots. These are guaranteed to retain all the WAL segments that Debezium needs even when Debezium suffers an outage. You should monitor replication slots carefully to avoid excessive disk consumption and other conditions such as catalog bloat that can arise if a replication slot is used infrequently. See the PostgreSQL documentation about replication slots for more information. If you are using a synchronous_commit setting other than on, then you should set wal_writer_delay to a value of about 10 milliseconds to ensure a low latency for change events. If you don't set this then the default value of about 200 milliseconds will apply.

Set up permissions

The Debezium connector needs a database user that has the REPLICATION and LOGIN roles so that it can perform replications. By default, a superuser has these roles but for security reasons, you should give the minimum necessary permissions to the Debezium user rather than full superuser permissions.

If you have administrator privileges then you can create a role for your Debezium user using a query like the following. Note that these are the minimum permissions the user needs to perform replications, but you might also need to grant other permissions.

CREATE ROLE <name> REPLICATION LOGIN;

Check replica identity and choose a publication mode

Before you start RDI, check the tables that RDI will capture and choose how PostgreSQL should publish their changes.

PostgreSQL needs a replica identity to identify the old row when it publishes an UPDATE or DELETE. By default, PostgreSQL uses the table's primary key. A table without a primary key has no usable replica identity unless you configure one. Inserts do not need a replica identity.

Run the following query while connected to the database that RDI uses. You do not put the database name in the query; select that database in your SQL editor or connection settings. The query checks all non-system schemas. To check only one schema, uncomment the marked line and replace <schema_name> with your schema name.

SELECT n.nspname AS schema_name,
       c.relname AS table_name,
       CASE c.relreplident
           WHEN 'd' THEN 'default'
           WHEN 'n' THEN 'nothing'
           WHEN 'f' THEN 'full'
           WHEN 'i' THEN 'index'
       END AS replica_identity
FROM pg_class AS c
JOIN pg_namespace AS n ON n.oid = c.relnamespace
WHERE c.relkind = 'r'
  AND n.nspname NOT IN ('pg_catalog', 'information_schema')
  -- Optional: uncomment and replace <schema_name> to check one schema only.
  -- AND n.nspname = '<schema_name>'
  AND (
      c.relreplident = 'n'
      OR (
          c.relreplident IN ('d', 'i')
          AND NOT EXISTS (
              SELECT 1
              FROM pg_index AS i
              WHERE i.indrelid = c.oid
                AND i.indislive
                AND i.indisvalid
                AND i.indisunique
                AND i.indimmediate
                AND i.indpred IS NULL
                AND CASE WHEN c.relreplident = 'd'
                         THEN i.indisprimary
                         ELSE i.indisreplident
                    END
          )
      )
  )
ORDER BY n.nspname, c.relname;

If this query returns a table that will be published, add a primary key that is not DEFERRABLE where possible. A deferrable primary key cannot serve as the replica identity. You can also configure a suitable unique index as the replica identity:

ALTER TABLE <table_name> REPLICA IDENTITY USING INDEX <unique_index_name>;

As a fallback, you can record the full old row in the write-ahead log:

ALTER TABLE <table_name> REPLICA IDENTITY FULL;

REPLICA IDENTITY FULL can increase write-ahead log volume and reduce performance. Use it when no suitable key is available or when RDI needs the full old row, such as denormalizing child tables with nesting. The keys setting in an RDI table definition creates a key for the Redis record. It does not create a PostgreSQL primary key or replica identity.

Understand which row values are available

Replica identity controls the old row values available to logical decoding for updates and deletes. With the default identity, PostgreSQL uses the primary key; it does not provide a complete old row. REPLICA IDENTITY FULL records all old column values.

Update events are not generally limited to changed fields. However, PostgreSQL can omit unchanged large values stored through The Oversized-Attribute Storage Technique (TOAST) when they are not part of the replica identity. Debezium represents unavailable values with a placeholder. With FULL, these values are available in the event's old and new row images. See Debezium's TOAST guidance.

For example, nesting a child record requires its parent_key to identify the parent Redis document, including when the child is deleted. Set REPLICA IDENTITY FULL on PostgreSQL child tables as required by the nesting instructions.

Choose a publication mode

The publication.autocreate.mode setting controls how RDI and Debezium manage the PostgreSQL publication:

The all_tables mode can publish changes for tables that are not listed in the RDI pipeline, including tables created in the future. If an application later updates or deletes a row in a published table without a replica identity, PostgreSQL can reject that application write. For this reason, filtered is the recommended mode for a pipeline that captures only selected tables.

The filtered mode requires permission to create the publication and ownership rights on the tables it manages. Use a separate publication for each source with different table filters, so one source does not change another source's publication.

For a customer-managed publication, create an explicit table list and disable automatic publication management. Replace the example table names with the tables you want to capture:

CREATE PUBLICATION rdi_publication
FOR TABLE public.customers, public.orders;

Then set the following properties under the PostgreSQL source in your RDI config.yaml:

sources:
  <source_name>:
    advanced:
      source:
        publication.name: rdi_publication
        publication.autocreate.mode: disabled

For RDI on Redis Cloud, enter these properties under Collector properties in the source configuration. See Configure PostgreSQL publications on Redis Cloud. With disabled, update the publication's table list yourself when changing the source's selected tables. Selecting a table in RDI does not add it to a customer-managed publication.

Replace an existing all-tables publication

Changing the RDI table selection does not narrow a FOR ALL TABLES publication. Changing publication.autocreate.mode to filtered also does not convert it: PostgreSQL rejects attempts to set, add, or remove individual tables in that publication.

Have your database administrator plan the change with the owners of every consumer of the publication:

1. Check existing publications in the source database:

   SELECT pubname, puballtables, pubupdate, pubdelete
   FROM pg_publication;
   

2. Identify every consumer of the broad publication and coordinate a maintenance window.

3. Stop the affected RDI sources before replacing the publication.

4. Replace the broad publication with an explicit table-list publication. You can recreate it under the same name, or create a new publication and update each consumer's publication name. For RDI, use publication.autocreate.mode: disabled for an administrator-managed publication. On Redis Cloud, contact support if the running source's collector properties cannot be edited in the console.

5. Remove the old broad publication only after coordinating all consumers. Leaving it in place still subjects its tables to replica-identity requirements, even if RDI uses a different publication or is stopped.

6. Resume the affected sources and verify inserts, updates, and deletes reach Redis.

Keep the replication slots and saved RDI offsets during this change. Do not reset the pipeline or drop its slot as a publication repair. If other consumers still need the broad publication, configure replica identity on its affected tables instead of removing it.

PostgreSQL publication error FAQ

I get this PostgreSQL error. What should I do?

You may see the following error when an application updates a published table without a replica identity:

SQLSTATE 55000
ERROR: cannot update table "example_table" because it does not have a replica identity and publishes updates
HINT: To enable updating the table, set REPLICA IDENTITY using ALTER TABLE.

Check whether the table should be captured by RDI. If it should, add a primary key or configure a suitable replica identity. As a fallback, run:

ALTER TABLE <table_name> REPLICA IDENTITY FULL;

If the table should not be captured, remove it from an explicit table-list publication or use filtered with the required permissions. For an existing FOR ALL TABLES publication, follow Replace an existing all-tables publication. Check all publications containing the table: another publication that publishes updates or deletes can still require replica identity. After changing RDI configuration, redeploy the pipeline so RDI applies it.

Set privileges for Debezium to create PostgreSQL publications with pgoutput

The Debezium user needs specific permissions to work with the pgoutput plug-in. The plug-in captures change events from the publications that PostgreSQL produces for your chosen source tables. A publication contains change events from one or more tables that are filtered using criteria from a publication specification.

If you have administrator privileges, you can create the publication specification manually or you can grant the Debezium user the privileges to create the specification automatically. The required privileges are:

• Replication privileges in the database to add the table to a publication.
• CREATE privileges on the database to add publications.
• SELECT privileges on the tables to copy the initial table data. Table owners automatically have SELECT permission for the table.

To add a table to a publication, the user must be an owner of the table. However, in this case, the source table already exists, so you must use a PostgreSQL replication group to share ownership between the Debezium user and the original owner. Configure the replication group using the following commands:

1. Create the replication group (the name replication_group here is just an example):

   CREATE ROLE replication_group;
   

2. Add the original owner of the table to the group:

   GRANT replication_group TO original_owner;
   

3. Add the Debezium replication user to the group:

   GRANT replication_group TO replication_user;
   

4. Transfer ownership of the table to replication_group:

   ALTER TABLE table_name OWNER TO replication_group;
   

If you want Debezium to create or update a publication for the tables configured in the RDI source, set publication.autocreate.mode to filtered. See the Debezium documentation for publication.autocreate.mode to learn more about this setting. If a database administrator manages an existing publication, use disabled instead, as described in Choose a publication mode.

Configure PostgreSQL for replication with the Debezium connector host

You must configure the database to allow replication with the host that runs the PostgreSQL Debezium connector. To do this, add an entry to the host-based authentication file, pg_hba.conf, for each client that needs to use replication. For example, to enable replication for <youruser> locally, on the server machine, you would add a line like the following:

local   replication     <youruser>                          trust

To allow <youruser> on localhost to receive replication changes using IPV4, add the line:

host    replication     <youruser>  127.0.0.1/32            trust

To allow <youruser> on localhost to receive replication changes using IPV6, add the line:

host    replication     <youruser>  ::1/128                 trust 

Find out more from the PostgreSQL pages about pg_hba.conf and network address types.

Supported PostgreSQL topologies

You can use the Debezium PostgreSQL connector with a standalone PostgreSQL server or with a cluster of servers. For versions 12 and below, PostgreSQL supports logical replication slots on only primary servers. This means that Debezium can only connect to a primary server for CDC and the connection will stop if this server fails. If the same server is promoted to primary when service resumes then you can simply restart the Debezium connector. However, if a different server is promoted to primary, then you must reconfigure Debezium to use the new server before restarting. Also, make sure the new server has the correct plug-in and configuration for Debezium.