pgloader loads data into PostgreSQL. It knows how to load data from several kinds of source:
- Files
- Live Databases
Let’s dive into those two different behaviours of pgloader, starting with its support for migrating a whole database from a connection string.
Loading data from Files
You might have to process huge CSV files from time to time. For example if you’re in the telco business, those might be Call Detail Records. When using PostgreSQL the best way to load CSV files is the COPY command and protocol, as read in the article entitled Faster bulk loading in Postgres with copy from Craig Kerstiens.
pgloader implements a retry policy around the data it loads, so that if your COPY transaction fails to load all the data, the rejected data is separated away from the rest of the file, and the data that is accepted by PostgreSQL makes it to the database server.
As simple way to describe pgloader is to say that it implement a retry-wrapper on-top of the PostgreSQL COPY command.
Database Migration
Migrating the data should be easy. After all, your source database is already a relational database server and implements the same SQL basics as any other one: relations, tables, views, and the expected column data types, or attribute and attribute domains.
It is actually quite impressive to realize how many differences exist between different technologies when it comes to implementing data types, default values, and other behaviors.
When using pgloader, all the knowledge when it comes to those differences is integrated into an open source product.
General Migration Behavior
pgloader follows those steps when migrating a whole database to PostgreSQL:
Fetch meta data and catalogs
pgloader fetches the source database meta data by querying the live database catalogs. Retrieving structured data makes it easier to process the complex level of information. Catalog queries that are used can be seen in the pgloader source code at GitHub.
In cases in which a database schema merge is required — that’s called the ORM compatibility mode, where the schema is created by a tool other than pgloader — then pgloader also fetches the target PostgreSQL metadata and compares it to the source catalogs.
The information retrieved by pgloader is handled internally in its own in-memory catalog format, and then used throughout the migration tasks. It is possible to implement some catalog mapping in these internal pgloader catalogs, as seen later in this chapter.
Prepare the target PostgreSQL database
Unless the database has already been created by another tool, pgloader creates the target PostgreSQL database schema by translating the source catalogs into PostgreSQL specific catalogs.
This steps involves casting rules and the advanced command language included in pgloader allows the user to define custom casting rules.
The casting rules are applied to the schema definition (column types, or attribute domains) and default values.
Copy table data using PostgreSQL copy protocol
For each table fetched in the first step, pgloader starts two threads:
The reader thread issues a
select
statement that retrieves all the data from the source database, sometimes already with some data processing.The writer thread receives the data read from the other thread thanks to a shared queue, and batches the data to be sent over to PostgreSQL.
While arranging a batch of data, pgloader also applies transformation functions to the data. The transformations applied depend on the casting rules. Examples include transforming a tinyint value into a Boolean value, or replacing date time entries
0000-00-00 00:00:00
with NULL values.
It’s possible to setup how many threads are allowed to be active at the same time on a pgloader run, and the data migration timing is influenced by that setup.
Create the indexes in parallel
For each table, as soon as the entire data set has made it to PostgreSQL, all the indexes defined on the table are created in parallel. The reason why pgloader creates all the indexes in parallel is PostgreSQL’s implementation of synchronize_seqscans:
This allows sequential scans of large tables to synchronize with each other, so that concurrent scans read the same block at about the same time and hence share the I/O workload. When this is enabled, a scan might start in the middle of the table and then “wrap around” the end to cover all rows, so as to synchronize with the activity of scans already in progress. This can result in unpredictable changes in the row ordering returned by queries that have no ORDER BY clause. Setting this parameter to off ensures the pre-8.3 behavior in which a sequential scan always starts from the beginning of the table. The default is on.
For a large table, having to scan it sequentially only once to create many indexes is a big win of course. Actualy, it’s still a win with smaller tables.
A primary key index is created with the
ALTER TABLE
command though, and PostgreSQL issues an exclusive lock, which prevents those indexes from being created in parallel with the other ones. That’s why pgloader creates primary key indexes in two steps. In this step, pgloader creates a unique index over not-null attributes.Complete the target PostgreSQL database
Once all the data is transfered and all indexes are created, pgloader completes the PostgreSQL database schema by installing the constraints and comments.
Constraints include turning the candidate unique not-null indexes into primary keys, then installing foreign keys as defined in the source database schema.
Once all those steps are done, your database has been converted to PostgresQL and it is ready to use!
As explained in the previous part about Continuous Migration, the main idea behind pgloader is to implement fully automated database migrations so that you can run this as a nightly task for the whole duration of your migration project.
Sample Output
It’s possible to run a whole database migration with a single command line, using only pgloader defaults, like this:
$ pgloader mysql://root@localhost/f1db pgsql:///f1db
Here’s a sample output of that command:
table name errors rows bytes total time
------------------------- --------- --------- --------- --------------
fetch meta data 0 33 0.231s
Create Schemas 0 0 0.009s
Create SQL Types 0 0 0.007s
Create tables 0 26 0.095s
Set Table OIDs 0 13 0.005s
------------------------- --------- --------- --------- --------------
f1db.constructorresults 0 11142 186.2 kB 0.123s
f1db.circuits 0 73 8.5 kB 0.020s
f1db.constructors 0 208 15.0 kB 0.026s
f1db.drivers 0 842 79.8 kB 0.059s
f1db.laptimes 0 426633 11.2 MB 2.760s
f1db.constructorstandings 0 11896 249.3 kB 0.184s
f1db.driverstandings 0 31726 719.1 kB 0.533s
f1db.pitstops 0 6251 209.6 kB 0.234s
f1db.races 0 997 100.6 kB 0.200s
f1db.seasons 0 69 3.9 kB 0.194s
f1db.qualifying 0 7516 286.4 kB 0.146s
f1db.results 0 23777 1.3 MB 0.364s
f1db.status 0 134 1.7 kB 0.008s
------------------------- --------- --------- --------- --------------
COPY Threads Completion 0 4 3.193s
Create Indexes 0 20 2.033s
Index Build Completion 0 20 1.274s
Reset Sequences 0 10 0.061s
Primary Keys 0 13 0.014s
Create Foreign Keys 0 0 0.000s
Create Triggers 0 0 0.000s
Set Search Path 0 1 0.002s
Install Comments 0 0 0.000s
------------------------- --------- --------- --------- --------------
Total import time ✓ 521264 14.3 MB 6.577s
The output clearly shows three sections: prepare the target database schema, transfer the data and complete the PostgreSQL database schema. Performance and timning will vary depending on hardware used. This sample is using a laptop optimized for traveling light.
As a result, don’t read too much into the numbers shown here. Best practices for doing the database migration include having different source hardware and target machines in order to maximize the disk IO capacity, and also a high bandwitdth network in between the servers. Using the same disk both as a source for reading and a target for writing is quite bad in term of performance.
Catalog Mapping
It is possible to change some catalog entries on the fly when doing a database migration. Supported changes include:
- Limiting tables to take care of,
- Migrating from one schema name to another one,
- Renaming tables.
It is also possible (only with MySQL as a source at the moment) to migrate data from a view definition rather than a table, allowing schema redesign on the fly.
Data Type Casting Rules
In order for the database migration to be fully automated, pgloader allows
users to define their own casting rules. Here’s a sample.load
command file
implementing this idea:
load database
from mysql://root@unix:/tmp/mysql.sock:3306/pgloader
into postgresql://dim@localhost/pgloader
alter schema 'pgloader' rename to 'mysql'
CAST column base64.id to uuid drop typemod drop not null,
column base64.data to jsonb using base64-decode,
type decimal when (and (= 18 precision) (= 6 scale))
to "double precision" drop typemod
before load do $$ create schema if not exists mysql; $$;
Custom casting rules may target either specific columns in your source database schema, or a whole type definition.
pgloader Features
Many users already use pgloader happily in their production environment and are quite satisfied with it. The software comes with a long list of features and is well documented at http://pgloader.readthedocs.io/en/latest/.
Supported RDBMS
pgloader already has support for migrating the following RDBMS in a fully automated way:
- DBF
- SQLite
- MySQL
- MS SQL Server
More sources for migrations are possible, such as Oracle™, Sybase™ or IBM Db2, see the pgloader Road Map document for details.