PostgreSQL Logical Replication

Logical replication is a method of replicating data objects and their changes, based upon their replication identity (usually a primary key). We use the term logical in contrast to physical replication, which uses exact block addresses and byte-by-byte replication. PostgreSQL supports both mechanisms concurrently. Logical replication allows fine-grained control over both data replication and security.

Logical replication uses a publish and subscribe model with one or more subscribers subscribing to one or more publications on a publisher node. Subscribers pull data from the publications they subscribe to and may subsequently re-publish data to allow cascading replication or more complex configurations.

Logical replication of a table typically starts with taking a snapshot of the data on the publisher database and copying that to the subscriber. Once that is done, the changes on the publisher are sent to the subscriber as they occur in real-time. The subscriber applies the data in the same order as the publisher so that transactional consistency is guaranteed for publications within a single subscription. This method of data replication is sometimes referred to as transactional replication.

The typical use-cases for logical replication are:

  • Sending incremental changes in a single database or a subset of a database to subscribers as they occur.
  • Firing triggers for individual changes as they arrive on the subscriber.
  • Consolidating multiple databases into a single one (for example for analytical purposes).
  • Replicating between different major versions of PostgreSQL.
  • Replicating between PostgreSQL instances on different platforms (for example Linux to Windows)
  • Giving access to replicated data to different groups of users.
  • Sharing a subset of the database between multiple databases.

The subscriber database behaves in the same way as any other PostgreSQL instance and can be used as a publisher for other databases by defining its own publications. When the subscriber is treated as read-only by application, there will be no conflicts from a single subscription. On the other hand, if there are other writes done either by an application or by other subscribers to the same set of tables, conflicts can arise.

Publication

publication can be defined on any physical replication master. The node where a publication is defined is referred to as publisher. A publication is a set of changes generated from a table or a group of tables, and might also be described as a change set or replication set. Each publication exists in only one database.

Publications are different from schemas and do not affect how the table is accessed. Each table can be added to multiple publications if needed. Publications may currently only contain tables. Objects must be added explicitly, except when a publication is created for ALL TABLES.

Publications can choose to limit the changes they produce to any combination of INSERTUPDATEDELETE, and TRUNCATE, similar to how triggers are fired by particular event types. By default, all operation types are replicated.

A published table must have a “replica identity” configured in order to be able to replicate UPDATE and DELETE operations, so that appropriate rows to update or delete can be identified on the subscriber side. By default, this is the primary key, if there is one. Another unique index (with certain additional requirements) can also be set to be the replica identity. If the table does not have any suitable key, then it can be set to replica identity “full”, which means the entire row becomes the key. This, however, is very inefficient and should only be used as a fallback if no other solution is possible. If a replica identity other than “full” is set on the publisher side, a replica identity comprising the same or fewer columns must also be set on the subscriber side. See REPLICA IDENTITY for details on how to set the replica identity. If a table without a replica identity is added to a publication that replicates UPDATE or DELETE operations then subsequent UPDATE or DELETE operations will cause an error on the publisher. INSERT operations can proceed regardless of any replica identity.

Every publication can have multiple subscribers.

A publication is created using the CREATE PUBLICATION command and may later be altered or dropped using corresponding commands.

The individual tables can be added and removed dynamically using ALTER PUBLICATION. Both the ADD TABLE and DROP TABLE operations are transactional; so the table will start or stop replicating at the correct snapshot once the transaction has committed.

Subscription

subscription is the downstream side of logical replication. The node where a subscription is defined is referred to as the subscriber. A subscription defines the connection to another database and set of publications (one or more) to which it wants to subscribe.

The subscriber database behaves in the same way as any other PostgreSQL instance and can be used as a publisher for other databases by defining its own publications.

A subscriber node may have multiple subscriptions if desired. It is possible to define multiple subscriptions between a single publisher-subscriber pair, in which case care must be taken to ensure that the subscribed publication objects don’t overlap.

Each subscription will receive changes via one replication slot. Additional temporary replication slots may be required for the initial data synchronization of pre-existing table data.

A logical replication subscription can be a standby for synchronous replication. The standby name is by default the subscription name. An alternative name can be specified as application_name in the connection information of the subscription.

Subscriptions are dumped by pg_dump if the current user is a superuser. Otherwise a warning is written and subscriptions are skipped, because non-superusers cannot read all subscription information from the pg_subscription catalog.

The subscription is added using CREATE SUBSCRIPTION and can be stopped/resumed at any time using the ALTER SUBSCRIPTION command and removed using DROP SUBSCRIPTION.

When a subscription is dropped and recreated, the synchronization information is lost. This means that the data has to be resynchronized afterwards.

The schema definitions are not replicated, and the published tables must exist on the subscriber. Only regular tables may be the target of replication. For example, you can’t replicate to a view.

The tables are matched between the publisher and the subscriber using the fully qualified table name. Replication to differently-named tables on the subscriber is not supported.

Columns of a table are also matched by name. The order of columns in the subscriber table does not need to match that of the publisher. The data types of the columns do not need to match, as long as the text representation of the data can be converted to the target type. For example, you can replicate from a column of type integer to a column of type bigint. The target table can also have additional columns not provided by the published table. Any such columns will be filled with the default value as specified in the definition of the target table.

Replication Slot Management

As mentioned earlier, each (active) subscription receives changes from a replication slot on the remote (publishing) side. Normally, the remote replication slot is created automatically when the subscription is created using CREATE SUBSCRIPTION and it is dropped automatically when the subscription is dropped using DROP SUBSCRIPTION. In some situations, however, it can be useful or necessary to manipulate the subscription and the underlying replication slot separately. Here are some scenarios:

  • When creating a subscription, the replication slot already exists. In that case, the subscription can be created using the create_slot = false option to associate with the existing slot.
  • When creating a subscription, the remote host is not reachable or in an unclear state. In that case, the subscription can be created using the connect = false option. The remote host will then not be contacted at all. This is what pg_dump uses. The remote replication slot will then have to be created manually before the subscription can be activated.
  • When dropping a subscription, the replication slot should be kept. This could be useful when the subscriber database is being moved to a different host and will be activated from there. In that case, disassociate the slot from the subscription using ALTER SUBSCRIPTION before attempting to drop the subscription.
  • When dropping a subscription, the remote host is not reachable. In that case, disassociate the slot from the subscription using ALTER SUBSCRIPTION before attempting to drop the subscription. If the remote database instance no longer exists, no further action is then necessary. If, however, the remote database instance is just unreachable, the replication slot should then be dropped manually; otherwise it would continue to reserve WAL and might eventually cause the disk to fill up. Such cases should be carefully investigated.

Conflicts

Logical replication behaves similarly to normal DML operations in that the data will be updated even if it was changed locally on the subscriber node. If incoming data violates any constraints the replication will stop. This is referred to as a conflict. When replicating UPDATE or DELETE operations, missing data will not produce a conflict and such operations will simply be skipped.

A conflict will produce an error and will stop the replication; it must be resolved manually by the user. Details about the conflict can be found in the subscriber’s server log.

The resolution can be done either by changing data on the subscriber so that it does not conflict with the incoming change or by skipping the transaction that conflicts with the existing data. The transaction can be skipped by calling the pg_replication_origin_advance() function with a node_name corresponding to the subscription name, and a position. The current position of origins can be seen in the pg_replication_origin_status system view.

Restrictions

Logical replication currently has the following restrictions or missing functionality. These might be addressed in future releases.

  • The database schema and DDL commands are not replicated. The initial schema can be copied by hand using pg_dump --schema-only. Subsequent schema changes would need to be kept in sync manually. (Note, however, that there is no need for the schemas to be absolutely the same on both sides.) Logical replication is robust when schema definitions change in a live database: When the schema is changed on the publisher and replicated data starts arriving at the subscriber but does not fit into the table schema, replication will error until the schema is updated. In many cases, intermittent errors can be avoided by applying additive schema changes to the subscriber first.
  • Sequence data is not replicated. The data in serial or identity columns backed by sequences will of course be replicated as part of the table, but the sequence itself would still show the start value on the subscriber. If the subscriber is used as a read-only database, then this should typically not be a problem. If, however, some kind of switchover or failover to the subscriber database is intended, then the sequences would need to be updated to the latest values, either by copying the current data from the publisher (perhaps using pg_dump) or by determining a sufficiently high value from the tables themselves.
  • Replication of TRUNCATE commands is supported, but some care must be taken when truncating groups of tables connected by foreign keys. When replicating a truncate action, the subscriber will truncate the same group of tables that was truncated on the publisher, either explicitly specified or implicitly collected via CASCADE, minus tables that are not part of the subscription. This will work correctly if all affected tables are part of the same subscription. But if some tables to be truncated on the subscriber have foreign-key links to tables that are not part of the same (or any) subscription, then the application of the truncate action on the subscriber will fail.
  • Large objects are not replicated. There is no workaround for that, other than storing data in normal tables.
  • Replication is only supported by tables, including partitioned tables. Attempts to replicate other types of relations, such as views, materialized views, or foreign tables, will result in an error.
  • When replicating between partitioned tables, the actual replication originates, by default, from the leaf partitions on the publisher, so partitions on the publisher must also exist on the subscriber as valid target tables. (They could either be leaf partitions themselves, or they could be further subpartitioned, or they could even be independent tables.) Publications can also specify that changes are to be replicated using the identity and schema of the partitioned root table instead of that of the individual leaf partitions in which the changes actually originate.

Architecture

Logical replication starts by copying a snapshot of the data on the publisher database. Once that is done, changes on the publisher are sent to the subscriber as they occur in real time. The subscriber applies data in the order in which commits were made on the publisher so that transactional consistency is guaranteed for the publications within any single subscription.

Logical replication is built with an architecture similar to physical streaming replication. It is implemented by “walsender” and “apply” processes. The walsender process starts logical decoding of the WAL and loads the standard logical decoding plugin (pgoutput). The plugin transforms the changes read from WAL to the logical replication protocol  and filters the data according to the publication specification. The data is then continuously transferred using the streaming replication protocol to the apply worker, which maps the data to local tables and applies the individual changes as they are received, in correct transactional order.

The apply process on the subscriber database always runs with session_replication_role set to replica, which produces the usual effects on triggers and constraints.

The logical replication apply process currently only fires row triggers, not statement triggers. The initial table synchronization, however, is implemented like a COPY command and thus fires both row and statement triggers for INSERT.

Initial Snapshot

The initial data in existing subscribed tables are snapshotted and copied in a parallel instance of a special kind of apply process. This process will create its own temporary replication slot and copy the existing data. Once existing data is copied, the worker enters synchronization mode, which ensures that the table is brought up to a synchronized state with the main apply process by streaming any changes that happened during the initial data copy using standard logical replication. Once the synchronization is done, the control of the replication of the table is given back to the main apply process where the replication continues as normal.

Monitoring

Because logical replication is based on a similar architecture as physical streaming replication, the monitoring on a publication node is similar to monitoring of a physical replication master.

The monitoring information about subscription is visible in pg_stat_subscription. This view contains one row for every subscription worker. A subscription can have zero or more active subscription workers depending on its state.

Normally, there is a single apply process running for an enabled subscription. A disabled subscription or a crashed subscription will have zero rows in this view. If the initial data synchronization of any table is in progress, there will be additional workers for the tables being synchronized.

Security

A user able to modify the schema of subscriber-side tables can execute arbitrary code as a superuser. Limit ownership and TRIGGER privilege on such tables to roles that superusers trust. Moreover, if untrusted users can create tables, use only publications that list tables explicitly. That is to say, create a subscription FOR ALL TABLES only when superusers trust every user permitted to create a non-temp table on the publisher or the subscriber.

The role used for the replication connection must have the REPLICATION attribute (or be a superuser). If the role lacks SUPERUSER and BYPASSRLS, publisher row security policies can execute. If the role does not trust all table owners, include options=-crow_security=off in the connection string; if a table owner then adds a row security policy, that setting will cause replication to halt rather than execute the policy. Access for the role must be configured in pg_hba.conf and it must have the LOGIN attribute.

In order to be able to copy the initial table data, the role used for the replication connection must have the SELECT privilege on a published table (or be a superuser).

To create a publication, the user must have the CREATE privilege in the database.

To add tables to a publication, the user must have ownership rights on the table. To create a publication that publishes all tables automatically, the user must be a superuser.

To create a subscription, the user must be a superuser.

The subscription apply process will run in the local database with the privileges of a superuser.

Privileges are only checked once at the start of a replication connection. They are not re-checked as each change record is read from the publisher, nor are they re-checked for each change when applied.

Configuration Settings

Logical replication requires several configuration options to be set.

On the publisher side, wal_level must be set to logical, and max_replication_slots must be set to at least the number of subscriptions expected to connect, plus some reserve for table synchronization. And max_wal_senders should be set to at least the same as max_replication_slots plus the number of physical replicas that are connected at the same time.

The subscriber also requires the max_replication_slots be set to configure how many replication origins can be tracked. In this case it should be set to at least the number of subscriptions that will be added to the subscriber. max_logical_replication_workers must be set to at least the number of subscriptions, again plus some reserve for the table synchronization. Additionally the max_worker_processes may need to be adjusted to accommodate for replication workers, at least (max_logical_replication_workers + 1). Note that some extensions and parallel queries also take worker slots from max_worker_processes.

Quick Setup

First set the configuration options in postgresql.conf:

wal_level = logical

The other required settings have default values that are sufficient for a basic setup.

pg_hba.conf needs to be adjusted to allow replication (the values here depend on your actual network configuration and user you want to use for connecting):

host     all     repuser     0.0.0.0/0     md5

Then on the publisher database:

CREATE PUBLICATION mypub FOR TABLE users, departments;

And on the subscriber database:

CREATE SUBSCRIPTION mysub CONNECTION 'dbname=foo host=bar user=repuser' PUBLICATION mypub;

The above will start the replication process, which synchronizes the initial table contents of the tables users and departments and then starts replicating incremental changes to those tables.