Subsetter is a Python utility that can be used for subsetting portions of relational databases. Subsetting is the action extracting a smaller set of rows from your database that still maintain expected foreign-key relationships between your data. This can be useful for testing against a small but realistic dataset or for generating sample data for use in demonstrations. This tool also supports filtering that allows you to remove/anonymize rows that may contain sensitive data.

Similar tools include Tonic.ai's platform and condenser. This is meant to be a simple CLI tool that overcomes many of the difficulties in using condenser.

You can use subsetter by installing it through pip:

pip install subsetter

Or by using the published msg555/subsetter image:

docker run --rm -v "./subsetter.yaml:/tmp/subsetter.yaml" msg555/subsetter -c /tmp/subsetter.yaml subset

The subsetter tool takes an approach of "one table, one query". This means that the subsetter will sample each table using only a single query. It cannot support calculating a full transitive closure of foreign key relationships for schemas that contain cycles. In general, as long as your schema contains no foreign key cycles and no target is reachable from another target, the subsetter will be able to automatically generate a plan that can sample your data.

The first step in subsetting a database is to generate a sampling plan. A sampling plan defines both the direct targets of the subsetter and what tables should be brought in through indirect foreign key references. You'll want to create a configuration file similar to subsetter.example.yaml, making sure to fill out the planner section to tell the planner what tables you want to sample any additional constraints that should be considered. Then you can create a plan with the below command:

subsetter -c my-config.yaml plan > plan.yaml

If you inspect the generated plan YAML document you will see a syntax tree that defines how each table will be sampled, potentially referencing other tables. Queries can reference either source tables or previously sampled tables. If you need to customize the way that tables are sampled beyond what the planner can automatically produce this is the place to do it. If needed, you can even write direct SQL here.

The sample sub-command will sample rows from the source database into your target output (either a database or as json files) using a plan generated using the plan sub-command. By default this tool will not copy schema from the source database and expects tables to already exist. If you would like it to attempt to create tables in the output database pass the --create flag. Additionally you must pass --truncate if you wish to clear any existing data in the output tables that may otherwise interfere with the sampling process.

subsetter --config my-config.yaml sample --plan my-plan.yaml --create --truncate

The sampling process proceeds in four phases:

1. If --create is specified it will attempt to create any missing tables. Existing tables will not be touched even if the schema does not match what is expected.
2. If --truncate is specified any tables about to be sampled will be first truncated. subsetter expects there to be no existing data in the destination database unless configured to run in merge mode.
3. Any sampled tables that are referenced by other tables will first be materialized into temporary tables on the source database.
4. Data is copied for each table from the source to destination.

There's also a subset subcommand to perform the plan and sample actions together. This will automatically feed the generated plan into the sampler, in addition to ensuring the same source database configuration is used for each.

subsetter -c my-config.yaml subset --create --truncate

By default any sampled row is copied directly from the source database to the destination database. However, there are several transformation steps that can be configured at the sampling stage that can change this behavior.

Filters allow you to transform the columns in each sampled row using either a set of built-in filters or through custom plugins. Built in filters allow you to easily replace common sources of personally identifiable information with fake data using the faker library. Filters for name, email, phone number, address, and location, and more come built in. See subsetter.example.yaml for full details on what filters exist and how to create a custom filter plugin.

Often tables make use of auto-incrementing integer identifiers to function as their primary key. Sometimes we may want the identifiers in our sampled data to be compact -- instead of retaining the value in the source database we may want our N sampled rows to have identifiers ranging from 1 to N. This is useful for sample data where we want to keep the identifiers easy to reference.

Any other table that has a foreign key that references one of these compacted columns will automatically also have the column involved in that foreign key adjusted to maintain semantic consistency.

Note that enabling compaction can have a noticable impact on performance. Compaction both requires more tables to be materialized on the source database and requries more joins when streaming data into the destination database.

By default the sampler expects no data to exist in the destination database. To get around this constraint we can turn on "merge" mode. To use merge mode all sampled tables must be either marked as "passthrough" or have a single-column, non-negative, integral primary key.

When enabled, the sampler will calculate the largest existing primary key identifier for each non-passthrough table and automatically shift the primary key of each sampled row to be larger using the equation:

new_id = source_id + max(0, existing_ids...) + 1

Passthrough tables instead will be sampled as normal except they will use the 'skip' conflict strategy which will have the effect of only inserting rows in a passthrough table if no row with the matching primary key exists in the destination database.

If merging multiple times it may be necessary to turn on identifier compaction to avoid the largest identifier in each table from growing too quickly due to large gaps.

Sampling usually means condensing a large dataset into a semantically consistent small dataset. However, there are times that what you really want to do is create a semantically consistent large dataset from your existing data (e.g. for performance testing). The sampler has support for this by setting the multiplicity factor.

Multiplicity works by creating multiple copies of your sampled dataset in your output database. To ensure these datasets do not collide it remaps all foreign keys into a new key-space. Note that this process assumes your foreign keys are opaque integers identifiers.

When using multiple targets each target table will be sampled entirely independently unless another target table directly or indirectly depends on some rows from it through a series of foreign keys. In the later case the subsetter will sample a union of the rows from the independently sampling of the table and those rows that other targets depend on.

The subsetter uses the foreign keys present in the database schema to understand relationships between data and generate a sampling plan. Foreign key relationships can be followed in both directions if need be. For example, suppose there was a users and an orders table where orders had a foreign key to the users table.

If users was sampled first the subsetter would sample orders from users by sampling all rows from orders such that their corresponding user row existed. This represents the maximal set of rows that can be included without violating foreign key constraints.

Otherwise if orders was sampled first the subsetter would sample users from orders by sampling all rows from users such that they had at least one order. This represents the minimal set of rows that can be included without violating foreign key constraints.

In general the subsetter will always sample tables in an order such that all foreign key relationships to previously sampled tables are going in the same direction. If they are followed in the forwards direction (as in our first case) the subsetter will select the intersection of all rows that obey each foreign key relationship. Otherwise if they are followed in the backwards direction (as in our second case) the subsetter will select the union of all rows that obey each foreign key relationship. This strategy ensures no foreign key relationships are violated in the sampled data.