Configuration Reference

This page covers the settings needed for most graphs. The Python signatures remain the source of truth for less common backend-specific and experimental options.

`GraphReduce`

Graph identity and execution

Parameter	Meaning
`name`	Human-readable graph name
`parent_node`	Node that anchors traversal and normally defines output grain
`compute_layer`	Backend selected from `ComputeLayerEnum`
`auto_features`	Generate schema-aware relation features during reduction
`debug`	Enable additional diagnostic logging
`train`	Include training-only behavior such as label computation; scoring graphs normally set this to `False`

Time windows

Parameter	Meaning
`cut_date`	Point-in-time reference for feature and label windows
`compute_period_val`	Size of the historical feature window
`compute_period_unit`	Unit for the feature-window value
`label_period_val`	Size of the future label window
`label_period_unit`	Unit for the label-window value

If label_node is set, both label-period settings are required. See Time windows and leakage safety for exact boundaries.

Labels

Parameter	Meaning
`label_node`	Relation from which the target is computed
`label_field`	Source field passed to automatic label aggregation
`label_operation`	Aggregation string or callable, such as `count`, `sum`, or `bool`

For DynamicNode, GraphReduce calls the default label implementation. A custom GraphReduceNode supplies do_labels(reduce_key) instead.

Automatic labels are left-joined to the parent. An entity with no matching label-window row therefore receives null, not an implicit zero.

Traversal

Parameter	Meaning
`auto_feature_hops_back`	Depth limit for descendant traversal from the parent
`auto_feature_hops_front`	Limit for optional joins through predecessor nodes

Keep auto_feature_hops_back at least as deep as the child relations that must contribute to the output. The default is 2.

SQL execution

Parameter	Meaning
`sql_client`	Backend client shared with SQL nodes
`lazy_execution`	Defer backend execution where supported
`dry_run`	Compile/log SQL without executing it
`checkpoint_schema`	Schema used for SQL intermediate materialization

Automatic SQL feature settings are documented separately under SQL auto-feature families.

`GraphReduceNode`

Source and schema

Parameter	Meaning
`fpath`	File path or SQL relation, depending on node type
`fmt`	Storage/source format such as `csv`, `parquet`, or `sql`
`pk`	Primary key as a string or ordered composite-key sequence
`prefix`	Required column-lineage prefix; must be unique in the graph
`date_key`	Timestamp used by feature and label filtering
`columns`	Known source columns, especially useful for SQL nodes
`compute_layer`	Node backend; normally matches the graph backend

Graph hydration overwrites shared execution and time settings on nodes with the graph-level values. Put one cutoff and one window policy on the graph unless a specialized workflow explicitly requires something else.

SQL-only additions

SQLNode accepts a backend client, optional table_name, and optional lists of sqlop values for each operation hook. A backend subclass may require more connection, catalog, or storage configuration.

Entity edges

GraphReduce.add_entity_edge() accepts:

Parameter	Meaning
`parent_node`	Destination of the relation's features
`relation_node`	Child relation being joined upward
`parent_key`	Key on the parent
`relation_key`	Matching key on the relation
`relation_type`	Relationship label; normally `parent_child`
`reduce`	Aggregate the relation before joining
`reduce_after_join`	Join first and reduce the enriched parent afterward

parent_key and relation_key must have the same number of columns. reduce and reduce_after_join cannot both be true.

Results and inspection

Path	Result
`graph.parent_node.df`	Materialized dataframe result
`graph.sql_ops`	Generated SQL statements in execution order
`graph.get_executed_op_records()`	Structured SQL operation records
`graph.freeze_execution_plan()`	Serializable record of a completed SQL plan
`graph.plot_graph(path)`	Interactive HTML visualization of graph topology

For SQL plan reuse, see Frozen SQL plans.