Skip to main content

How to Implement a New Database Driver

First, read through the document.

Make sure data-diff is set up for development, and that all the tests pass (try to at least set it up for mysql and postgresql).

Look at the other database drivers for examples and inspiration.

1. Add dependencies to pyproject.toml

Most new drivers will require a 3rd party library in order to connect to the database.

These dependencies should be specified in the pyproject.toml file, in [tool.poetry.extras].


postgresql = ["psycopg2"]

This way, users can install the dependencies needed for your database driver, with pip install 'data-diff[postgresql]'.

This way, data-diff can support a wide variety of drivers, without requiring our users to install libraries that they won't use.

2. Implement the database module

New database modules belong in the data_diff/databases directory.

A database module consists of:

  1. Dialect (Class responsible for normalizing/casting fields. e.g. Numbers/Timestamps)
  2. Database class that handles connecting to the DB, querying (if the default doesn't work), closing connectiosn and etc.

Choosing a base class, based on threading Model

You can choose to inherit from either base.Database or base.ThreadedDatabase.

Usually, databases with cursor-based connections, like MySQL or PostgreSQL, only allow connections to be used by the thread that created them. In order to support multithreading, we implement them by inheriting from ThreadedDatabase, which holds a pool of worker threads, and creates a new connection per thread.

Usually, cloud databases, such as Snowflake and BigQuery, open a new connection per request, and support simultaneous queries from any number of threads. In other words, they already support multithreading, so we can implement them by inheriting directly from Database.

Import on demand

Database drivers should not import any 3rd party library at the module level.

Instead, they should be imported and initialized within a function. Example:

    from .base import import_helper

def import_postgresql():
import psycopg2
import psycopg2.extras

return psycopg2

We use the import_helper() decorator to provide a uniform and informative error. The string argument should be the name of the package, as written in pyproject.toml.


All queries to the database pass through _query(). It takes SQL code, and returns a list of rows. Here is its signature:

    def _query(self, sql_code: str) -> list: ...

For standard cursor connections, it's sufficient to implement it with a call to base._query_conn(), like:

    return _query_conn(self._conn, sql_code)

select_table_schema() / query_table_schema()

If your database does not have a information_schema.columns table, or if its structure is unusual, you may have to implement your own select_table_schema() function, which returns the query needed to return column information in the form of a list of tuples, where each tuple is _column_name, data_type, datetime_precision, numeric_precision, numeric_scale_.

If such a query isn't possible, you may have to implement query_table_schema() yourself, which extracts this information from the database, and returns it in the proper form.

If the information returned from query_table_schema() requires slow or error-prone post-processing, you may delay that post-processing by overriding _process_table_schema() and implementing it there.
The method _process_table_schema() only gets called for the columns that will be diffed.


  • data_diff.databases.database_types.AbstractDatabase.select_table_schema()

  • data_diff.databases.database_types.AbstractDatabase.query_table_schema()


Each database class must have a TYPE_CLASSES dictionary, which maps between the string data-type, as returned by querying the table schema, into the appropriate data-diff type class, i.e. a subclass of database_types.ColType.


When providing a datetime or a timestamp to a database, the database may lower its precision to correspond with the target column type.

Some databases will lower precision of timestamp/datetime values by truncating them, and some by rounding them.

ROUNDS_ON_PREC_LOSS should be True if this database rounds, or False if it truncates.

__init__(), create_connection()

The options for the database connection will be given to the __init__() method as keywords.

If you inherit from Database, your __init__() method may create the database connection.

If you inherit from ThreadedDatabase, you should instead create the connection in the create_connection() method.


If you inherit from Database, you will need to implement this method to close the connection yourself.

If you inherit from ThreadedDatabase, you don't have to implement this method.


  • data_diff.databases.database_types.AbstractDatabase.close()

quote() and to_string()

These methods are used when creating queries, to quote a value, or cast it to STRING/VARCHAR.

They accept an SQL code fragment, and returns a new code fragment representing the appropriate computation.

For more information, read their docs:

  • data_diff.databases.database_types.AbstractDatabase.quote()

  • data_diff.databases.database_types.AbstractDatabase.to_string()

normalize_number(), normalize_timestamp() and md5_to_int()

Because comparing data between 2 databases requires both the data to be in the same format - we have normalization functions.

Databases can have the same data in different formats, e.g. DECIMAL vs FLOAT vs VARCHAR, with different precisions. DataDiff works by converting the values to VARCHAR and comparing it. Your normalize_number/normalize_timestamp functions should account for differing precisions between columns.

These functions accept an SQL code fragment, and returns a new code fragment representing the appropriate computation.


This is used to determine types which the system cannot effectively detect.
DECIMAL(10,3) needs to be parsed by a custom algorithm. You'd be using regex to split it into Field name + Width + Scale.

3. Add tests

Add your new database to the DATABASE_TYPES dict in tests/

The key is the class itself, and the value is a dict of {category: [type1, type2, ...]}

Categories supported are: int, datetime, float, boolean, and uuid.


db.PostgreSQL: {
"int": [ "int", "bigint" ],
"datetime": [
"timestamp(6) without time zone",
"timestamp(3) without time zone",
"timestamp(0) without time zone",
"timestamp with time zone",

Then run the tests and make sure your database driver is being tested.

You can run the tests with unittest.

To save time, we recommend running them with unittest-parallel.

When debugging, we recommend using the -f flag, to stop on error. Also, use the -k flag to run only the individual test that you're trying to fix.

4. Debugging

You can enable debug logging for tests by setting the logger level to DEBUG, via the environment variable LOG_LEVEL, or the LOG_LEVEL variable in /tests/ This will display all the queries ran, and display the type detected for each column.

5. Create a Pull-Request

Open a pull request on github, and we'll take it from there!