pgcat - Enhanced postgresql logical replication

• Why pgcat?
• Architecture
• Build from source
• Run
• Conflict handling
• Table mapping
• Replication identity
• Limitations
• Conflict resolution

---

Note that some facts have changed in postgresql offical over the years since my last commit to pgcat:

1. Since version 13:

Allow partitioned tables to be logically replicated via publications (Amit Langote)

2. Since version 16:

Allow logical replication subscribers to process only changes that have no origin (Vignesh C, Amit Kapila)

This can be used to avoid replication loops. This is controlled by the new CREATE SUBSCRIPTION ... ORIGIN option.

3. Since version 17

Add logical replication failover control to CREATE/ALTER SUBSCRIPTION (Shveta Malik, Hou Zhijie, Ajin Cherian)

https://www.postgresql.org/docs/17/logical-replication-failover.html

Allow partitioned tables to have identity columns (Ashutosh Bapat)

However, it doesn't cover all the features of pgcat, so pgcat is still meaningful.

---

Why pgcat?

The built-in logicial replication has below shortages:

• only support base table as replication target (also support paritioned table since postgresql 13)
• do not filter any origin, which will cause bi-directional dead loop
• could not do table name mapping
• no conflict resolution

pgcat makes below enhancements:

• supports any table type as replication target e.g. view, fdw, partitioned table, citus distributed table
• only replicates local changes so that you could make bi-directional replication, e.g. replicates data between two datacenter
• table name mapping
• optional lww (last-writer-win) conflict resolution
• save replication progress in table, so that it would be logged when subscriber failovers, it would retain the progress. In contrast, the built-in logical replication of pg saves the progress in non-logged file.

Architecture

pgcat is based on logical decoding, and reuses the publication part and pgoutput output plugin of the pg built-in logical replication.

Instead of worker processes and low-level row ingression, pgcat uses sql templates to apply high-level sql commands, so it could make largest compatibility on target table type. It is written in golang and runs in separate process.

The lww uses one additional jsonb column to store meta info, e.g. timestamp. It supports timestamp-override columns or cassandra-like counter columns.

Build from source

pgcat has two parts, pgcat binary and postgresql extension.

Dependencies

• golang >= 1.12
• postgressql >= 11

Build pgcat

cd cmd/pgcat
go build
cd ../pgcat_setup_lww
go generate
go build

Build pgcat-pgxs

https://github.com/kingluo/pgcat-pgxs

Add pgcat user

• Create pgcat user with replication attribute on the publisher database:

-- adjust your password here
CREATE USER pgcat with REPLICATION PASSWORD 'pgcat';

• Create pgcat user on the subscriber database:

-- adjust your password here
CREATE USER pgcat with PASSWORD 'pgcat';

Note: if you do bi-directional replication, create pgcat user with replication attribute on both publisher and subscriber databases.

Create extension

Use superuser role to create extension on both the publisher and subscriber databases.

CREATE EXTENSION IF NOT EXISTS pgcat;

This command would create extension under the pgcat schema.

Configure the publisher database

• set logical wal level

postgresql.conf

wal_level = logical

• Allow replication connection to the publisher database

pg_hba.conf

host     all     pgcat     0.0.0.0/0     md5

Restart pg.

pg_ctl restart

Run

Setup the table

Grant pgcat

On subscriber database, pgcat needs to read/write the table.

grant select,insert,update,delete,truncate on foobar to pgcat;

If you configure to copy the table in the subscription, then on publisher database, pgcat needs to select the table.

grant select on foobar to pgcat;

Setup publication

On publisher database:

CREATE PUBLICATION foobar FOR TABLE foobar;
alter publication foobar add table foobar;

Setup lww (optional)

If you need last-writer-win conflict resolution, then run pgcat_setup_lww on all pg instances.

pgcat_setup_lww -c lww.yml

Check lww.yml for configuration file example.

Setup subscription

On subscriber database:

INSERT INTO pgcat.pgcat_subscription(name, hostname, port, username, password,
dbname, publications, copy_data, enabled) VALUES ('foobar', '127.0.0.1', 5433,
'pgcat', 'pgcat', 'tmp', '{foobar}', true, true);

The pgcat would check pgcat_subscription, if it changes, pgcat would apply the changes.

Run pgcat

cd cmd/pgcat
# modify pgcat.yml to fit your need
pgcat -c pgcat.yml

pgcat uses golang proxy dialer, so if you need to access your database via proxy, you could run below command:

all_proxy=socks5h://127.0.0.1:20000 pgcat -c pgcat.yml

If you need to run pgcat in daemon on Linux, just use setsid command:

setsid pgcat -c pgcat.yml &>/dev/null

Admin HTTP API:

If you configure admin_listen_address, e.g. admin_listen_address: 127.0.0.1:30000, then you could run below commands to admin the pgcat process:

# rotate the log file
# you could use logrotate to rotate pgcat log files as you need
curl http://127.0.0.1:30000/rotate

# reload the yaml config file, e.g. you could add new databases to replicate
curl http://127.0.0.1:30000/reload

Conflict handling

If not using lww, it's likely to have conflict, e.g. unique violation, especially for bi-directional replication. When conflict happens, pgcat would panic this subscription, and restarts it in 1 minute, and so on.

How to handle conflict? As pg doc said:

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.

It also works for pgcat. But since pgcat substitutes the subscriber part, so to skip transaction, you need to use below sql command:

update pgcat.pgcat_subscription_progress set lsn='0/27FD9B0';

The lsn is the lsn of the commit record of this transaction, you could find it in pgcat log when conflict happens:

dml failed, commit_lsn=0/27FD9B0, err=...

Table mapping

You could map publisher table name to different subscriber table name.

You could map multiple tables into one table, gathering multiple data source into one target, e.g. partition tables, citus shards. Here the target could be partitioned table, view, or citus main table, so that you could have heterogeneous layout at different database and do easy replication.

For example, at database1, you have tables foobar2 and foobar3, and you need to configure them be put into table foobar2 at database2.

In database2 (subscriber database), run below sql command via superuser:

insert into pgcat.pgcat_table_mapping(subscription,priority,src,dst)
	values('foobar',1,'^public.foobar[2-3]$','public.foobar2');

Note that the regexp and table name should be full qualified, i.e. with schema prefix. And the regexp is better to be surrounded with ^ and $, otherwise the matching is error prone.

Replication identity

View and foreign table does not have replica ident, so it needs to configure them in pgcat. The configuration table is pgcat.pgcat_replident.

For example, I need to set id1 and id2 columns as replica ident of view foobar2_view.

insert into pgcat.pgcat_replident values('public.foobar2_view', '{"id1", "id2"}');

Note that the table name should be full qualified, i.e. with schema prefix.

Limitations

• if the target is view, set copy_data to false in the subscription, and do not use instead of trigger on view, please use view rule instead. copy to view requires instead of trigger attached, but the current pg has no way to set always role on the view instead of trigger. On the other hand, pgcat sets its role to replica, so instead of trigger would not be called, and then copy to view would be no-op.

Conflict resolution

Conflict resolution is necessary for logical replication.

Logical replication is normally used for loose-coupling different pg deployments (or different pg HA deployments, one HA deployment consists of one master and multiple slaves, where they are connected via physical replication), especially for different data centers, where it does not require real-time data consistence).

For example, we have two groups of pg:

pg1 consists of pg1-master and pg1-slave, run in datacenter1.

pg2 consists of pg2-master, pg2-slave1 and pg2-slave2, run in datacenter2.

They replicate data changes to each other, but not in real time, and the network between data centers may be broken for time to time.

Each data center changes data independently, which would normally involves data with same identity, e.g. same primary key. Such data needs some sort of policy to keep consistent in both data centers.

For example, pg1 writes "foo" to row foo, before this change is applied to pg2, pg2 writes "bar" to row foo. Then which would be the final version of value?

lww (Last-writer-win)

Similar to Cassandra, it could use write timestamp to resolve the conflict. The write with latest timestamp would be the last version of data value. The pg deployment would keep time in sync, e.g. via NTP.

The pgcat_setup_lww command is used to setup the table for used by lww. Note that conflict resolution is optional, so if you're sure the data changes are consistent by nature or by design, you do not need to run this command on your tables.

How does it work?

It would category the columns as:

• Inline columns Inline columns have the same timestamp as the row.
• Individual columns Individual columns have their own timestamp associated.
• Counter columns Each pg deployment would have their own individual copy of counter value. But when you read it, it would sum up all copys as the final value.

When you run pgcat_setup_lww, it would:

• Add a new column __pgcat_lww to the target table, which is in jsonb type, used to record the meta info of the row and columns.
• for existing rows, it would populate __pgcat_lww concurrently (i.e. would not block concurrent data r/w when the table setup is processing).
• Create a trigger, which would
  • for local change, get the current timestamp as the row timestamp
  • for remote change, get the row timestamp in __pgcat_lww
  • for inline columns, compare the row timestamp
  • for each individual column, compare its own timestamp
  • for counter column, store the remote change in __pgcat_lww, use system identifier from pg_control_system() to identify different remote peers
• Create a view to filter the __pgcat_lww so that it would not be exported to application level. For example, if you have table foobar, then it would create a view foobar_pgcat_lww_view.
• Create a helper function pgcat_lww_counter() used to sum up counter column. You could call pgcat_lww_counter(__pgcat_lww, 'foobar') to get the value of column foobar.
• When you delete a row, it would not be really deleted, instead, it would be marked as tombstone in __pgcat_lww, so pgcat_setup_lww would create index for tombstone rows and create a helper function to vacuum them whenever you need, e.g. remove tombstones older than 3 days: foobar_pgcat_lww_vaccum(interval '3 days') .