Differences from featuretools¶

tusk borrows several concepts from featuretools, including relationships, primitives and cutoff times. tusk differs where it can benefit from lazy tables, or where an SQL backend needs a different answer.

• Renamed container and entry points. tusk keeps concepts similar to featuretools but renames them. A collection of tables joined by primary and foreign keys is a database.

• Lazy output, wherever possible. tusk builds a query plan for the feature matrix. You can pass eager or lazy tables as input. If your backend supports it, tusk returns an uncomputed lazy table. Eager backends, like pandas, always compute directly.

• Almost any backend you like. Just keep it consistent within one database.

• Feature names are SQL identifiers. featuretools writes MEAN(orders.quantity). tusk writes MEAN__orders__quantity. A backend that produces SQL treats dots and parentheses as table qualifiers and function calls, not as parts of a column name. Feature.display_name is the only place that keeps the conventional form, for logs, docs and error messages.

• primary_key and row_creation_time rather than index and time_index. Narwhals has no index concept. row_creation_time names what the column means: when the row became knowable.

• row_update_times asks you what the column held before. featuretools' set_secondary_time_index always makes it null instead. See row update times.

• Simplified relationships. add_relationship(parent=, child=, foreign_key=). On the parent side, the join column is always the parent's primary key.

• Opt-in validation. You can use tusk to check whether your definitions match the real datasets. By default, tusk enables only the checks that need nothing but the table schema. If you are willing to spend compute time on the full datasets, run db.validate(), or set validate=True on db.add_table(...) or db.add_relationship(...).

• One global cutoff_time applies to the target table too. tusk leaves out a row that did not yet exist at the cutoff time. tusk treats a table with no row_creation_time as timeless and passes it through unfiltered.

• Group and ordered transforms only run within foreign-key groups. tusk has no groupby_trans_primitives argument. Primitives like cum_sum and percentile, passed in trans_primitives, run within each foreign-key group. A row only sees the rows that share its foreign key. featuretools needs groupby_trans_primitives for that behavior, and it also runs these primitives across the whole table from trans_primitives. See how transforms are applied.

• primary_key is optional. A table without one cannot be a relationship parent or a DFS target. On such a table, ordered transform and ordered aggregation primitives might behave non-deterministically.

Comparing primitives¶

Primitive coverage lists every featuretools primitive beside its tusk counterpart.