Pglogical DDL Deployment (pgl_ddl_deploy)

Transparent DDL replication for Postgres 9.5+

Overview

• Release Notes
• High Level Description
• Features
• A Full Example
• Installation

Setup and Deployment

• Configuration
• Permissions
• Deployment
• Monitoring and Administration

Limitations and Restrictions

• DDL involving multiple tables
• Unsupported Commands
• Multi-Statement Client SQL Limitations

Resolving DDL Replication Issues

• Resolving Failed DDL on Subscribers
• Resolving Unhandled DDL

For Developers

• Help Wanted Features
• Regression testing

Overview

Read the Release Summary: https://innovation.enova.com/pursuing-postgres-ddl-replication/

Release Notes

Release 1.3

Summary of changes:

• Fix missed renaming of columns or constraints for include_only_repset_tables configs
• Added undeploy functions which allow you to drop DDL replication for only a given config
• Add is_deployed column to event_trigger_schema to easily tell which configs are deployed.

Release 1.2

Summary of changes:

• Fix a bug where unsupported event triggers are created for include_only_repset_tables configs, which only apply to include_schema_regex configs
• Add grant to pglogical.replication_set to allow any user to fire event trigger function even for temp tables

Release 1.1

Summary of changes:

• Allow a customized set of command tags for event triggers in set_configs - defaulted to the original tags in 1.0.
• Add DDL replication by the specific set of tables (include_only_repset_tables) in a replication set, instead of schema-only regex. This option only supports ALTER TABLE events
• Add option to allow subscriber DDL to fail and be queued without breaking replication - useful for VIEWs and such which we want to replicate but we don't want failures to stop data replication - queue_subscriber_failures
• Allow to mark unhandled events as resolved for monitoring purposes

High Level Description

With any current logical replication technology for Postgres, we normally have excellent ways to replicate DML events (INSERT, UPDATE, DELETE), but are left to figure out propagating DDL changes on our own. That is, when we create new tables, alter tables, and the like, we have to manage this separately in our application deployment process in order to make those same changes on logical replicas, and add such tables to replication.

As of Postgres 10.0, there is no native way to do "transparent DDL replication" to other Postgres clusters alongside any logical replication technology, built on standard Postgres.

This project is an attempt to do just that. The framework is built on the following concepts:

• Event triggers always fire on DDL events, and thus give us immediate access to what we want
• Event triggers gives us access (from 9.5+) to what objects are being altered
• We can see what SQL the client is executing within an event trigger
• We can validate and choose to propagate that SQL statement to subscribers
• We can add new tables to replication at the point of creation, prior to any DML execution

In many environments, this may cover most if not all DDL statements that are executed in an application environment. We know this doesn't cover 100% of edge cases, but we believe the functionality and robustness is significant enough to add great value in many Postgres environments. It will work especially well in these two use cases:

• When you want to replicate all user tables
• When you want to replicate only a subset of tables in a schema that will not have any foreign key dependencies to other schemas

We also think it's possible to expand this concept by further leveraging the Postgres parser. There is much detail below on what the Limitations and Restrictions are of this framework.

NOTE: The concept implemented here could be extended with any replication framework that can propagate SQL to subscribers, i.e. skytools and Postgres' built-in logical replication starting at 10.0. The reason this has not already been done is because of time constraints, and because there are a lot of specifics related to each replication technology. We would welcome a project to extend this to work with any replication technology.

Features

• Any DDL SQL statement can be propagated directly to subscribers without your developers needing to overhaul their migration process or know the intricacies of replication.

• Tables will also be automatically added to replication upon creation.

• Filtering by schema (regular expression) is supported. This allows you to selectively replicate only certain schemas within a replication set.

• Filtering by a specific set of tables is supported, which is most useful to replicate a small set of tables and maintain things like columns added/dropped

• There is an option to deploy in a lock-safe way on subscribers. Note that this means replication will lag until all blockers finish running or are terminated.

• There is an option to fail certain events on the subscriber to be retried later. This is useful for example if you are replicating VIEW DDL but do not want that to block replication on failure.

• In some edge cases, alerting can be built around provided logging for the DBA to then handle possible manual deployments

A Full Example

Since we always look for documentation by example, we show this first. Assuming pglogical is already setup, and given these replication sets:

• default - replicate every event
• insert_update - replicate only inserts and updates

Provider:

CREATE EXTENSION pgl_ddl_deploy;

--Setup permissions
SELECT pgl_ddl_deploy.add_role(oid) FROM pg_roles WHERE rolname = 'app_owner';

--Setup configs
INSERT INTO pgl_ddl_deploy.set_configs
(set_name,
include_schema_regex,
lock_safe_deployment,
allow_multi_statements)
VALUES ('default',
  '.*',
  true,
  true),
  ('insert_update',
  '.*happy.*',
  true,
  true);

Subscribers (run on both default and insert_update subscribers):

CREATE EXTENSION pgl_ddl_deploy;

--Setup permissions for the same role on subscriber to have DDL permissions 
CREATE ROLE app_owner WITH NOLOGIN;
SELECT pgl_ddl_deploy.add_role(oid) FROM pg_roles WHERE rolname = 'app_owner';

--Be sure that on the subscriber, app_owner role has the following perms:
GRANT CREATE ON DATABASE :DBNAME TO app_owner;

--If schemas already exist on subscriber that you want the app_owner
--role to be able to modify with any DDL, then you must do this: 
ALTER TABLE foo OWNER TO app_owner; --...etc

Here is a way to fix the subscriber table owner based on the tables already in replication on provider for the same replication set and ddl configuration you just setup on provider (shell command):

PGSERVICE=provider_cluster psql provider_db << EOM | PGSERVICE=subscriber_cluster psql subscriber_db
COPY
(
SELECT 'ALTER TABLE '||quote_ident(n.nspname)||'.'||quote_ident(c.relname)||' OWNER TO app_owner;'
FROM pglogical.replication_set rs
INNER JOIN pgl_ddl_deploy.set_configs sc
  ON sc.set_name = rs.set_name
INNER JOIN pgl_ddl_deploy.rep_set_table_wrapper rsr
  ON rsr.set_id = rs.set_id
INNER JOIN pg_class c ON c.oid = rsr.set_reloid
INNER JOIN pg_namespace n ON n.oid = c.relnamespace
WHERE rs.set_name = 'insert_update'
  AND n.nspname ~* sc.include_schema_regex
  AND n.nspname !~* pgl_ddl_deploy.exclude_regex()
)
TO STDOUT;
EOM

Provider:

--Deploy DDL replication
SELECT pgl_ddl_deploy.deploy(set_name)
FROM pgl_ddl_deploy.set_configs;

--App deployment role
SET ROLE app_owner;

--Let's make some data!
CREATE TABLE foo(id serial primary key);
ALTER TABLE foo ADD COLUMN bla INT;
INSERT INTO foo (bla) VALUES (1),(2),(3);

CREATE SCHEMA happy;
CREATE TABLE happy.foo(id serial primary key);
ALTER TABLE happy.foo ADD COLUMN bla INT;
INSERT INTO happy.foo (bla) VALUES (1),(2),(3);
DELETE FROM happy.foo WHERE bla = 3;

Subscriber to default:

SELECT * FROM foo;
 id | bla
----+-----
  1 | 1
  2 | 2
  3 | 3
(3 rows)

SELECT * FROM happy.foo;
 id | bla
----+-----
  1 | 1
  2 | 2
(3 rows)

Note that both tables are replicated based on configuration, as are all events (inserts, updates, and deletes).

Subscriber to insert_update:

SELECT * FROM foo;
ERROR:  relation "foo" does not exist
LINE 1: SELECT * FROM foo;

SELECT * FROM happy.foo;
 id | bla
----+-----
  1 | 1
  2 | 2
  3 | 3
(3 rows)

Note that the foo table (in public schema) was not replicated. Also, because we are not replicating deletes here, happy.foo still has all data.

Installation

The functionality of this requires postgres version 9.5+ and a working install of pglogical. Packages will be available soon. To build from source:

make
make install
make installcheck # run regression suite

See the notes below on requirements to run the regression suite.

This extension requires pglogical to be installed before you can create the extension in any database. Then the extension can be deployed as any postgres extension:

CREATE EXTENSION pgl_ddl_deploy;

This extension needs to be installed on provider and all subscribers.

NOTE: If you upgrade pglogical from version 1.* to 2.* while pgl_ddl_deploy is installed, you will need to re-run function pgl_ddl_deploy.dependency_update() to update pglogical-specific dependencies. If you indeed were to upgrade and do nothing, you will likely start to see WARNING level logs indicating a problem. DDL statements should not fail because of the way we are allowing exceptions in this deployment framework to proceed with only a WARNING level log.

Setup and Deployment

Configuration

DDL replication is configured on a per-replication set basis, in terms of pglogical.replication_set.

Add rows to pgl_ddl_deploy.set_configs in order to configure (but not yet deploy) DDL replication for a particular replication set. For example:

--Only some options are shown.  See below for all options

--This type of configuration will DDL replicate all user schemas, and auto-add
--new matching schemas/tables to replication:
INSERT INTO pgl_ddl_deploy.set_configs (set_name, include_schema_regex)
VALUES ('default', '.*');

--This type of configuration will maintain only the specific set of tables
--in the given replication set for any `ALTER TABLE` statements:
INSERT INTO pgl_ddl_deploy.set_configs (set_name, include_only_repset_tables)
VALUES ('my_special_tables', TRUE);

The relevant settings:

• set_name: pglogical replication_set name
• include_schema_regex: a regular expression for which schemas to include in DDL replication. This can be used to auto-add new tables to replication. This option is incompatible with include_only_repset_tables.
• lock_safe_deployment: if true, DDL will execute in a low lock_timeout loop on subscriber
• allow_multi_statements: if true, multiple SQL statements sent by client can be propagated under certain conditions. See below for more details on caveats and edge cases. If false, only a single SQL statement (technically speaking - a SQL statement with a single node parsetree) will be eligible for propagation.
• include_only_repset_tables: if true, only tables that are in replication will be maintained by DDL replication. Thus only ALTER TABLE statements are permitted here. This option is incompatible with include_schema_regex.
• queue_subscriber_failures: if true, DDL will be allowed to fail on subscriber without breaking replication, and queued for retry using function pgl_ddl_deploy.retry_all_subscriber_logs(). This is useful for example if you are replicating VIEW DDL but do not want failures to block data replication. It is NOT recommendeded that you use this with any TABLE replication, since those events are likely to break data replication.
• create_tags: the set of command tags for which the create event triggers will fire. Change with caution. These are defaulted to the appropriate default set for either include_schema_regex or include_only_repset_tables.
• drop_tags: the set of command tags for which the drop event triggers will fire. Change with caution. These are defaulted to the appropriate default set for either include_schema_regex or include_only_repset_tables.

There is already a pattern of schemas excluded always that you need not worry about. You can view them in this function:

SELECT pgl_ddl_deploy.exclude_regex();

You can use this query to check what your current configs will pull in for schemas:

SELECT sc.set_name, n.nspname
FROM pg_namespace n
INNER JOIN pgl_ddl_deploy.set_configs sc
  ON nspname !~* pgl_ddl_deploy.exclude_regex()
  AND n.nspname ~* sc.include_schema_regex
ORDER BY sc.set_name, n.nspname;

You can use this query to test a new regex for what existing schemas it matches:

SELECT n.nspname
FROM pg_namespace n
WHERE nspname !~* pgl_ddl_deploy.exclude_regex()
  AND n.nspname ~* 'test'
ORDER BY n.nspname;

There are no stored procedures to insert/update set_configs, which we don't think would add much value at this point. There is a check constraint in place to ensure the regex is valid.

Permissions

It is important to consider which role will be allowed to run DDL in a given provider. As it stands, this role will need to exist on the subscriber as well, because this same role will be used to try to deploy on the subscriber. pgl_ddl_deploy provides a function to provide permissions needed for a given role to use DDL deployment. This needs to run provider and all subscribers:

SELECT pgl_ddl_deploy.add_role(oid)
FROM pg_roles
WHERE rolname IN('app_owner_role');

Deployment of Automatic DDL Replication

To deploy (meaning activate) DDL replication for a given replication set, run:

--Deploy any set_configs with given set_name:
SELECT pgl_ddl_deploy.deploy(set_name);

--Deploy only a single set_config_id:
SELECT pgl_ddl_deploy.deploy(set_config_id);

• From this point on, the event triggers are live and will fire on the following events (by default, unless you have customized create_tags or drop_tags):

   command_tag
-----------------
 ALTER FUNCTION
 ALTER SEQUENCE
 ALTER TABLE
 ALTER TYPE
 ALTER VIEW
 CREATE FUNCTION
 CREATE SCHEMA
 CREATE SEQUENCE
 CREATE TABLE
 CREATE TABLE AS
 CREATE TYPE
 CREATE VIEW
 DROP FUNCTION
 DROP SCHEMA
 DROP SEQUENCE
 DROP TABLE
 DROP TYPE
 DROP VIEW
 SELECT INTO

• Not all of these events are handled in the same way - see Limitations and Restrictions below
• Note that if, based on your configuration, you have tables that should be added to replication already, but are not, you will not be allowed to deploy. This is because DDL replication should only be expected to automatically add new tables to replication. To override this, add the tables to replication manually and sync as necessary.

DDL replication can be disabled/enabled (this will disable/enable event triggers):

--By set_name
SELECT pgl_ddl_deploy.disable(set_name);
SELECT pgl_ddl_deploy.enable(set_name);

--By set_config_id
SELECT pgl_ddl_deploy.disable(set_name);
SELECT pgl_ddl_deploy.enable(set_name);

You can also undeploy DDL replication, which means dropping all event triggers and functions for a given config:

SELECT pgl_ddl_deploy.undeploy(set_name);
SELECT pgl_ddl_deploy.undeploy(set_config_id);

If you want to change the configuration in set_configs, you can re-deploy by again running pgl_ddl_deploy.deploy on the given set_name. There is currently no enforcement/warning if you have changed configuration but not deployed, but should be easy to add such a feature.

Note that you are able to override the event triggers completely, for example, if you are an administrator who wants to run DDL and you know you don't want that propagated to subscribers. You can do this with SESSION_REPLICATION_ROLE, i.e.:

SET SESSION_REPLICATION_ROLE TO REPLICA;

--I don't care to send this to subscribers
ALTER TABLE foo SET (autovacuum_vacuum_threshold = 1000);

RESET SESSION_REPLICATION_ROLE;

Monitoring and Administration

This framework will log all DDL changes that attempt to be propagated. It is also generous in allowing DDL replication procedures to fail in order not to prevent application deployments of DDL. An exception will never be allowed to block a DDL statement. The most that will happen is log_level WARNING will be raised. This feature is based on the assumption that we do not want replication issues to ever block client-facing application functionality.

Several tables are setup to manage DDL replication and log exceptions, in addition to server log warnings raised at WARNING level in case of issues:

• events - Logs replicated DDL events on the provider
• subscriber_logs - Logs replicated DDL events executed on subscribers
• commands - Logs detailed output from pg_event_trigger_ddl_commands() and pg_event_trigger_dropped_objects()
• unhandled - Any DDL that is captured but cannot be handled by this framework (see details below) is logged here.
• exceptions - Any unexpected exception raise by the event trigger functions are logged here

There are resolved fields on the unhandled and exceptions tables that can be marked so that monitoring will show only new problems based on this table by using functions:

• pgl_ddl_deploy.resolve_unhandled(unhandled_id INT, notes TEXT = NULL)
• pgl_ddl_deploy.resolve_exception(exception_id INT, notes TEXT = NULL)

Limitations and Restrictions

DDL involving multiple tables

A single DDL SQL statement which alters tables both replicated and non-replicated cannot be supported. For example, if I have include_schema_regex which includes only the regex '^replicated.*', this is unsupported:

DROP TABLE replicated.foo, notreplicated.bar;

Likewise, the following can be problematic if you are using filtered replication:

ALTER TABLE replicated.foo ADD COLUMN foo_id INT REFERENCES unreplicated.foo (id);

Depending on your environment, such cases may be very rare, or possibly common. For example, such edge cases are far less likely when you want to do 1:1 replication of just about all tables in your application database. Also, if you are not likely to have relationships between schemas you both are and are not replicating, then edge cases will be unlikely. As mentioned above, this framework will work best with these two use cases:

• When you want to replicate all user tables
• When you want to replicate only a subset of tables in a schema that will not have any foreign key dependencies to other schemas

In this case, the DDL statement could fail on the subscriber. To resolve this, see Resolving Failed DDL on Subscribers.

Unsupported Commands

CREATE TABLE AS and SELECT INTO are not supported to replicate DDL due to limitations on transactional consistency. That is, if a table is created from a set of data on the provider, to run the same SQL on the subscriber will in no way guarantee consistent data. For example:

CREATE TABLE foo AS
SELECT field_1, field_2, now() AS refreshed_at
FROM table_1;

Not only is it possible that table_1 doesn't even exist on the subscriber, even if it does it may not be fully up to date with the provider, in which case the data created in the table on the subscriber would not match. Worse, is that the now() function is basically guaranteed to be different on the subscriber.

It is recommended instead that the DDL statement creating the table and the DML inserting into the table are separated. Continuing the above example:

CREATE TABLE foo (field_1 INT PRIMARY KEY, field_2 TEXT, refreshed_at TIMESTAMPTZ);
INSERT INTO foo (field_1, field_2, refreshed_at)
SELECT field_1, field_2, now() AS refreshed_at
FROM table_1;

The above is completely supported by this framework, bearing in mind some of the edge cases with multi-statements. The CREATE TABLE will automatically be replicated by this framework, and the table will be added to replication since it has a primary key. Then the INSERT will be replicated by normal pglogical replication.

NOTE that temp tables are not affected by this limitation, since temp objects are always excluded from DDL replication anyway.

To resolve these, see Resolving Unhandled DDL.

Multi-Statement Client SQL Limitations

It is important to understand that limitations on multi-statement client SQL has nothing to do with executing multiple SQL statements at once, or in one transaction. Of course, it is assumed that will often or even usually be the case, and this framework can handle that just fine.

The complexities and limitations come when the client sends all SQL statements as one single string to Postgres. Assume the following SQL statements:

CREATE TABLE foo (id serial primary key, bla text);
INSERT INTO foo (bla) VALUES ('hello world');

If this was in a file that I called via psql, it would run as two separate SQL command strings.

However, if in python or ruby's ActiveRecord I create a single string as above and execute it, then it would be sent to Postgres as 1 single SQL command string. In such a case, this framework is aware that a multi-statement is being executed by using Postgres' parser to get the command tags of the full SQL statement. You have a little freedom here with the allow_multi_statements option:

• If false, pgl_ddl_deploy will only auto-replicate a client SQL statement that contains 1 command tag that matches the event trigger command tag. That's really safe, but it means you may have a lot more unhandled deployments.
• If true, pgl_ddl_deploy will only auto-replicate DDL that contains safe command tags to propagate. For example, mixed DDL and DML is forbidden, because executing such a statement would in effect double-execute DML on both provider and subscriber. However, if you have a CREATE TABLE and ALTER TABLE in one command, assuming the table should be included in replication based on your configuration, then such a statement will be sent to subscribers. But of course, we can't guarantee that all DDL statements in this command are on the same table - so there could be edge cases.

In any case that a SQL statement cannot be automatically run on the subscriber based on these analyses, instead it will be logged as a WARNING and put into the unhandled table for manual processing.

To resolve these, see Resolving Unhandled DDL.

The regression suite in the sql folder has examples of several of these cases.

Thus, limitations on multi-statement SQL is largely based on how your client sends its messages in SQL to Postgres, and likewise how your developers tend to write SQL. The good thing about this is that it ought to be much easier to train developers to logically separate SQL in their migrations, as opposed to far more work we need to do when we don't have any kind of transparent DDL replication.

These limitations obviously have to be weighed against the cost of not using a framework like this at all in your environment.

The unhandled table and WARNING logs are designed to be leveraged with monitoring to create alerting around when manual intervention is required for DDL changes.

Resolving DDL Replication Issues

Resolving Failed DDL on Subscribers

In some cases, you may propagate DDL that fails on the subscriber, and replication will break, unless you have enabled queue_subscriber_failures. You will then need to:

• Manually deploy with the same SQL statement modified so that it excludes the failing portion. For the example above of both adding a column and adding a foreign key, assuming we don't have that unreplicated table, then you would run:

ALTER TABLE replicated.foo ADD COLUMN foo_id INT;

• Consume the change in affected replication slot using pg_logical_slot_get_changes up to specific LSN of the transaction which included the DDL statement to get replication working again.
• You could also "trick" the DDL into applying - for example, by creating a dummy object that will allow the DDL to apply even if you don't need it, then discarding those objects once replication is working again.
• Re-enable replication for affected subscriber(s)

If you are using queue_subscriber_failures, any DDL failures for your given configuration will be logged as failed in pgl_ddl_deploy.subscriber_logs. You can resolve issues, and attempt to re-deploy by using functions:

--Retry all failed, in transactional order
SELECT pgl_ddl_deploy.retry_all_subscriber_logs();

--Retry only a single failed log
SELECT pgl_ddl_deploy.retry_subscriber_log(subscriber_log_id);

Resolving Unhandled DDL

At present, an unhandled DDL deployment may not break replication by itself. If the DDL statement that could not be deployed doesn't actually affect any data being replicated (for example, a brand new table is added), then replication will continue. However, the situation should be resolved ASAP because it is assumed that whatever table(s) that were involved in the DDL should be propagated to subscribers. It is also possible that replication will break immediately. For example, if a column is added to the table, and data is replicating, it will fail immediately because of a column mismatch. In such cases you will need to:

• Manually deploy the unhandled SQL statement modified so that it excludes the unhandled portion. For the example of a multi-statement SQL above, you would need to exclude the INSERT portion of the SQL and run:

CREATE TABLE foo (id serial primary key, bla text);

• If a new table is involved, in this case you also will need to manually add the table to replication using pglogical.replication_set_add_table
• If a new table is involved, you may need to resynchronize the table if data has been replicating for it
• If a new table is NOT involved (for example ALTER TABLE ADD COLUMN), then replication will simply continue where it broke
• Re-enable replication for affected subscriber(s)
• Mark your unhandled records as resolved using function:

SELECT pgl_ddl_deploy.resolve_unhandled(unhandled_id INT, notes TEXT = NULL);

To be more conservative, we may want a feature that forces replication to break if there is an unhandled deployment, for example, by sending an exception through replicate_ddl_command. But such a feature may cause additional and unnecessary administration overhead, since it is likely the strictness of replication will cause the system to break when it should.

For Developers

Help Wanted Features

We are currently using the parser only to get the list of command tags. One big advantage to this is that it isn't going to be difficult to maintain as Postgres develops. One disadvantage is that it tells us nothing about the actual objects being modified, nor the type of modification.

We believe it is feasible to use the parser to actually only process the parts of a multi-statement SQL command that we want to, but this far more ambitious, and would be more work to maintain. We would need to leverage the different structures of each DDL command's parsetree to determine if the table is in a schema we care to replicate. Then, we would need to use the parser's lex code to take out the piece(s) we want.

Theoretically speaking, don't we have all that we need in the SQL statement + the parser to programatically use only what we want and send it to subscribers? I think we do, but lots of work would be required, and we would welcome those with more comfort in the parser code to help if interested.

SQL files

To build the higher version SQL files (i.e. 1.1) from the lower versions + the upgrade patch SQL files, run pgl_ddl_deploy-sql-maker.sh.

Regression testing

You can run the regression suite, which must be on a server that has pglogical packages available, and a cluster that is configured to allow creating the pglogical extension (i.e. adding it to shared_preload_libraries). Note that the regression suite does not do any cross-server replication testing, but it does cover a wide variety of replication cases and the core of what is needed to verify DDL replication.

As with any Postgres extension:

make install
make installcheck

# There is support for testing both 1.1 and the upgraded path from 1.0 to 1.1
FROMVERSION=1.0 make installcheck   # Test from 1.0 upgrade to 1.1
FROMVERSION=1.1 make installcheck