pgcopydb

pgcopydb

Introduction

pgcopydb is a tool that automates running pg_dump | pg_restore between two running Postgres servers. To make a copy of a database to another server as quickly as possible, one would like to use the parallel options of pg_dump and still be able to stream the data to as many pg_restore jobs.

The idea would be to use pg_dump --jobs=N --format=directory postgres://user@source/dbname | pg_restore --jobs=N --format=directory -d postgres://user@target/dbname in a way. This command line can't be made to work, unfortunately, because pg_dump --format=directory writes to local files and directories first, and then later pg_restore --format=directory can be used to read from those files again.

Given that, pgcopydb then uses pg_dump and pg_restore for the schema parts of the process, and implements its own data copying multi-process streaming parts. Also, pgcopydb bypasses pg_restore index building and drives that internally so that all indexes may be built concurrently.

Base Copy and Change Data Capture

pgcopydb implements both the base copy of a database and also Change Data Capture to allow replay of changes from the source database to the target database. The Change Data Capture facility is implemented using Postgres Logical Decoding infrastructure. The default plugin is pgoutput, which is built into PostgreSQL core (since version 10) and requires no extension installation. The test_decoding and wal2json plugins are also supported.

WAL changes are stored in structured SQLite databases (output.db and replay.db) in the pgcopydb work directory, enabling reliable processing across process restarts, size-based file rotation, and a stream prune command to reclaim disk space.

The pgcopydb follow command implements a logical replication client for the pgoutput logical decoding plugin (built into PostgreSQL core), as well as test_decoding and wal2json.

The pgcopydb clone --follow command implements a full solution for online migration. Beware that online migrations involve a lot more complexities when compared to offline migration. It is always a good idea to first implement offline migration first. The command pgcopydb clone is used to implement the offline migration approach.

The pgcopydb clone --all-databases command copies an entire Postgres instance (all databases) from source to target in a single operation.

Documentation

Full documentation is available online, including manual pages of all the pgcopydb sub-commands. Check out https://pgcopydb.readthedocs.io/.

…

Example

When using pgcopydb it is possible to achieve the result outlined before with this simple command line:

$ export PGCOPYDB_SOURCE_PGURI="postgres://[email protected]/dbname"
$ export PGCOPYDB_TARGET_PGURI="postgres://[email protected]/dbname"

$ pgcopydb clone --table-jobs 8 --index-jobs 2

A typical output from the command would contain lots of lines of logs, and then a table summary with a line per table and some information (timing for the table COPY, cumulative timing for the CREATE INDEX commands), and then an overall summary that looks like the following:

…

Installing pgcopydb

See our documentation.

Design Considerations (why oh why)

The reason why pgcopydb has been developed is mostly to allow two aspects that are not possible to achieve directly with pg_dump and pg_restore, and that requires just enough fiddling around that not many scripts have been made available to automate around.

Bypass intermediate files for the TABLE DATA

First aspect is that for pg_dump and pg_restore to implement concurrency they need to write to an intermediate file first.

The docs for pg_dump say the following about the --jobs parameter:

You can only use this option with the directory output format because this is the only output format where multiple processes can write their data at the same time.

The docs for pg_restore say the following about the --jobs parameter:

Only the custom and directory archive formats are supported with this option. The input must be a regular file or directory (not, for example, a pipe or standard input).

So the first idea with pgcopydb is to provide the --jobs concurrency and bypass intermediate files (and directories) altogether, at least as far as the actual TABLE DATA set is concerned.

The trick to achieve that is that pgcopydb must be able to connect to the source database during the whole operation, when pg_restore may be used from an export on-disk, without having to still be able to connect to the source database. In the context of pgcopydb requiring access to the source database is fine. In the context of pg_restore, it would not be acceptable.

For each table, build all indexes concurrently

The other aspect that pg_dump and pg_restore are not very smart about is how they deal with the indexes that are used to support constraints, in particular unique constraints and primary keys.

Those indexes are exported using the ALTER TABLE command directly. This is fine because the command creates both the constraint and the underlying index, so the schema in the end is found as expected.

That said, those ALTER TABLE ... ADD CONSTRAINT commands require a level of locking that prevents any concurrency. As we can read on the docs for ALTER TABLE:

Although most forms of ADD table_constraint require an ACCESS EXCLUSIVE lock, ADD FOREIGN KEY requires only a SHARE ROW EXCLUSIVE lock. Note that ADD FOREIGN KEY also acquires a SHARE ROW EXCLUSIVE lock on the referenced table, in addition to the lock on the table on which the constraint is declared.

The trick is then to first issue a CREATE UNIQUE INDEX statement and when the index has been built then issue a second command in the form of ALTER TABLE ... ADD CONSTRAINT ... PRIMARY KEY USING INDEX ..., as in the following example taken from the logs of actually running pgcopydb:

…

This trick is worth a lot of performance gains on its own, as has been discovered and experienced and appreciated by pgloader users already.

Dependencies

At run-time pgcopydb depends on the pg_dump and pg_restore tools being available in the PATH. The tools version should match the Postgres version of the target database.

When you have multiple versions of Postgres installed, consider exporting the PG_CONFIG environment variable to the version you want to use. pgcopydb then uses the PG_CONFIG from the path and runs ${PG_CONFIG} --bindir to find the pg_dump and pg_restore binaries it needs.

Manual Steps

The pgcopydb command line also includes entry points that allows implementing any step on its own.

1. pgcopydb snapshot &
2. pgcopydb dump schema
3. pgcopydb restore pre-data
4. pgcopydb copy table-data
5. pgcopydb copy blobs
6. pgcopydb copy sequences
7. pgcopydb copy indexes
8. pgcopydb copy constraints
9. pgcopydb restore post-data
10. kill %1

Using individual commands fails to provide the advanced concurrency capabilities of the main pgcopydb clone command, so it is strongly advised to prefer that main command.

Also when using separate commands, one has to consider the --snapshot option that allows for consistent operations. A background process should then export the snapshot and maintain a transaction opened for the duration of the operations. See documentation for pgcopydb snapshot.

Authors

• Dimitri Fontaine

License

Copyright (c) The PostgreSQL Global Development Group.

This project is licensed under the PostgreSQL License, see LICENSE file for details.

This project includes bundled third-party dependencies, see NOTICE file for details.

Maintenance & Support

This project is maintained through OSS Maintenance — a funded model designed for enterprises running these tools in production.

• Priority bug fixes and improvements via a paid queue
• Quarterly releases with predictable delivery
• Upstream-first: fixes land in the project itself

PostgreSQL consultancies can also subscribe at the Partner tier to cover their entire client base under a single plan.

View support plans