churnkit 1.2.8a3__tar.gz → 1.2.8a5__tar.gz

{churnkit-1.2.8a3 → churnkit-1.2.8a5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: churnkit
-Version: 1.2.8a3
+Version: 1.2.8a5
 Summary: Structured ML framework for customer churn prediction -- from exploration notebooks to production pipelines, locally or on Databricks.
 Project-URL: Homepage, https://github.com/aladjov/CR
 Project-URL: Documentation, https://github.com/aladjov/CR/wiki

{churnkit-1.2.8a3 → churnkit-1.2.8a5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "churnkit"
-version = "1.02.8a3"
+version = "1.02.8a5"
 description = "Structured ML framework for customer churn prediction -- from exploration notebooks to production pipelines, locally or on Databricks."
 readme = "README.md"
 license = {text = "Apache-2.0"}

{churnkit-1.2.8a3 → churnkit-1.2.8a5}/src/customer_retention/__init__.py

@@ -17,7 +17,7 @@ Main module categories:
                 llm_context, iteration)
 """
 
-__version__ = "1.02.8a3"
+__version__ = "1.02.8a5"
 
 # Environment utilities (always available)
 from .core.compat import (

{churnkit-1.2.8a3 → churnkit-1.2.8a5}/src/customer_retention/generators/pipeline_generator/findings_parser.py

@@ -1403,11 +1403,18 @@ class FindingsParser:
                 if name in config.landing:
                     siblings.append(name)
                     seen.add(name)
+            # Translate Python-list `in [a, b]` into Spark-SQL `in (a, b)`
+            # — exploration's filter executor goes through
+            # `_spark_safe_query_expr`, but production's generated landing
+            # script feeds the predicate straight into Spark SQL via
+            # `df.filter(...)` and rejects bracket lists with PARSE_SYNTAX_ERROR.
+            from customer_retention.core.compat import _spark_safe_query_expr
+            sql_predicate = _spark_safe_query_expr(str(predicate))
             target.filters.append(TransformationStep(
                 type=PipelineTransformationType.LANDING_FILTER,
                 column=dataset,
                 parameters={
-                    "predicate": str(predicate),
+                    "predicate": sql_predicate,
                     "sibling_views": list(siblings),
                 },
                 rationale="NB00 SAMPLE_FILTER_COLUMNS (cohort scope)",
@@ -2527,17 +2534,6 @@ class FindingsParser:
             for src in dd.source_columns:
                 for suffix in ("_delta_hours", "_hour", "_dow", "_is_weekend"):
                     columns.add(f"{src}{suffix}")
-            # The runtime aggregator computes datetime-derived features per row
-            # (`compute_temporal_features`) and then runs them through the same
-            # window aggregation as numeric `value_columns`. Mirror that fan-out
-            # so silver_derived recs referencing `<src>_<suffix>_<func>_<window>`
-            # (e.g. `PAYMENT_DATE_is_weekend_max_all_time`) survive parity.
-            if agg and agg.windows and agg.agg_funcs:
-                for src in dd.source_columns:
-                    for suffix in ("_delta_hours", "_hour", "_dow", "_is_weekend"):
-                        for window in agg.windows:
-                            for func in agg.agg_funcs:
-                                columns.add(f"{src}{suffix}_{func}_{window}")
         for text_cfg in event_cfg.text_features:
             if text_cfg.component_columns:
                 columns |= set(text_cfg.component_columns)
@@ -3300,6 +3296,8 @@ class FindingsParser:
             categorical_value_counts=categorical_value_counts,
         )
 
+    _DATETIME_DERIVED_SUFFIXES = ("_delta_hours", "_hour", "_dow", "_is_weekend")
+
     @staticmethod
     def _apply_sparse_aggregate_prune(
         *, ts, value_columns: List[str], blocked: Dict[str, List[str]],
@@ -3313,12 +3311,19 @@ class FindingsParser:
         median-imputed downstream, polluting selection with near-constant noise.
         ``count``/``sum`` remain emitted; categorical and binary aggregates are
         not affected. Fails open (no pruning) when metadata is missing.
+
+        Datetime-derived columns (``_delta_hours`` / ``_hour`` / ``_dow`` /
+        ``_is_weekend``) are exempt — they're per-row signals computed at
+        ingestion, where ``max``/``mean`` are well-defined even at one event
+        per entity (e.g. ``max(is_weekend)`` = "ever fired on a weekend").
         """
         if ts is None or ts.avg_events_per_entity is None:
             return
         if float(ts.avg_events_per_entity) >= threshold:
             return
         for col in value_columns:
+            if col.endswith(FindingsParser._DATETIME_DERIVED_SUFFIXES):
+                continue
             existing = set(blocked.get(col, []))
             existing.update(["mean", "max"])
             blocked[col] = sorted(existing)