PostgreSQL: Difference between pg_dump + psql versus template create?

Question

I know there are two ways of making a copy of a database.

One is to export the database as a giant SQL file, then load it as a separate database:

pg_dump <database> | psql <new database>

Another way is to pass the database name as a template to a database creation argument:

createdb -T <database> <new database>

What is the difference between these two methods, if any? Are there any benefits of using one over another, such as performance?

Answer 1

Using CREATE DATABASE / createdb with a template makes a directory copy, whereas pg_dump + psql has to serialize and deserialize the whole database, send them on a round-trip to the client, and has to run everything through the transaction and write-ahead logging machinery. So the former method should be much faster.

The disadvantage is that CREATE DATABASE locks the template database while it's being copied. So if you want to create copies of a live database, that won't work so well. But if you want to quickly make copies of an inactive/template database, then using CREATE DATABASE is probably the right solution.

Answer 2

According to the current docs

Although it is possible to copy a database other than template1 by specifying its name as the template, this is not (yet) intended as a general-purpose "COPY DATABASE" facility. The principal limitation is that no other sessions can be connected to the template database while it is being copied. CREATE DATABASE will fail if any other connection exists when it starts; otherwise, new connections to the template database are locked out until CREATE DATABASE completes.

Apart from that mild warning, which goes back to at least version 8.2, you can make certain kinds of changes by using createdb--things like changing collation, encoding, etc. (Within limits.)

Personally, I'd have a hard time justifying the use of createdb, which takes a full database lock, to copy a production database.

I think the other main difference is that "dump and load" is a fully supported way of copying a database. Also, you can carry a copy of the dump to an isolated development computer for testing if you need to. (The createdb utility has to have access to both the source and the target at the same time.) But I have not used createdb to make copies, so I could be wrong.