node_diff

Compare Node Outputs Across Builds

Compares the artifact produced by a node across two historical builds of the same pipeline, or compares two different nodes. Returns a structured VDiff dictionary with a consistent envelope.

Dispatches to a type-appropriate comparison: - DataFrame → row-/column-level diff with optional key-based alignment - Model (PMML) → coefficient deltas and fit-stat comparison - Scalar → before/after with numeric delta - Python-native objects → artifact deserialization through the bundled tlang Python package, then stable JSON rendering plus git-like unified diffs - Julia-native objects → artifact deserialization through the bundled tlang Julia package, then DeepDiffs-based structural comparison - R-native objects → artifact deserialization through the bundled tlang R package, then diffobj-based structural comparison - Generic → structural comparison over string representations

Runtime-native object diffs are preserved only for artifacts using the standard default or tobj serializers. If you assign a custom serializer name (for example "rds" or "pkl"), node_diff() falls back to the normal artifact loading path; for those cases, call the companion R/Python/Julia helper packages directly with an explicit deserializer.

For Julia-native artifacts, node_diff() launches a fresh Julia helper process for each comparison. This keeps the integration simple but adds startup cost for repeated or very large diffs.

Signature

node_diff(
  node_a    :: ComputedNode,
  node_b    :: ComputedNode,
  log_a     :: String = "latest",
  log_b     :: String = "latest",
  key       :: List[Symbol] = [],
  context   :: Int = 3
) :: VDiff

Parameters

• node_a (ComputedNode): The “before” node, e.g. p.clean_data.
• node_b (ComputedNode): The “after” node, e.g. p.clean_data.
• log_a (String): Build log selector for node_a. Accepts "latest", a timestamp prefix ("20260510_120000"), or a regex matched against filenames in _pipeline/. Default: "latest".
• log_b (String): Build log selector for node_b. Same format as log_a. Default: "latest".
• key (List[Symbol]): For DataFrames: the natural key column(s) used to align rows before diffing. If empty, rows are aligned by position. Default: [].
• context (Int): Number of unchanged rows shown above and below each changed hunk. Default: 3.

Returns

Dict: A VDiff envelope dictionary with the following fields:

Field	Type	Description
kind	String	"dataframe_diff", "model_diff", "scalar_diff", "python_object_diff", "julia_object_diff", "r_object_diff", or "generic_diff"
node_a	String	Name of the first node
node_b	String	Name of the second node
log_a	String	Resolved log filename for node_a
log_b	String	Resolved log filename for node_b
value_type	String	T type name of the diffed values
identical	Bool	true if no differences were found
summary	Dict	Type-specific summary counts
detail	Dict	Type-specific detail
hunks	List[Dict]	Diff hunks or rendered diff regions when available

Examples

-- Compare the same node across two historical builds
d = node_diff(p.clean_data, p.clean_data,
      log_a = "20260510_120000",
      log_b = "20260515_090000")

-- Compare two different nodes in the current build
d = node_diff(p.clean_data, p.validated_data)

-- Same node, latest vs a named earlier run, keyed on an id column
d = node_diff(p.customers, p.customers,
      log_a = "20260501",
      log_b = "latest",
      key = [$customer_id])

-- Model comparison
d = node_diff(p.model_v1, p.model_v2)

-- Python-native artifact comparison (for example NumPy ndarrays)
d = node_diff(p.weights, p.weights, log_a = 1, log_b = 2)

-- Julia-native artifact comparison (for example serialized structs or arrays)
d = node_diff(p.julia_model, p.julia_model, log_a = 1, log_b = 2)

-- R-native artifact comparison (for example saved model objects)
d = node_diff(p.r_model, p.r_model, log_a = 1, log_b = 2)

See Also

• build_log — retrieve build log for a pipeline
• build_log_history — list historical builds
• explain — structural explanation of any value, including VDiff