SNOW-2872192: Support targeted delete-insert in save_as_table

CHANGELOG.md

@@ -6,6 +6,8 @@
 
 #### New Features
 
+- Added support for targeted delete-insert via the `overwrite_condition` parameter in `DataFrameWriter.save_as_table`
+
 #### Bug Fixes
 
 #### Improvements

src/snowflake/snowpark/_internal/analyzer/analyzer.py

@@ -1228,6 +1228,12 @@ def do_resolve_with_resolved_children(
                 child_attributes=resolved_child.attributes,
                 iceberg_config=iceberg_config,
                 table_exists=logical_plan.table_exists,
+                overwrite_condition=self.analyze(
+                    logical_plan.overwrite_condition,
+                    df_aliased_col_name_to_real_col_name,
+                )
+                if logical_plan.overwrite_condition
+                else None,
             )
 
         if isinstance(logical_plan, Limit):

src/snowflake/snowpark/_internal/analyzer/snowflake_plan.py

@@ -1267,6 +1267,7 @@ def save_as_table(
         child_attributes: Optional[List[Attribute]],
         iceberg_config: Optional[dict] = None,
         table_exists: Optional[bool] = None,
+        overwrite_condition: Optional[str] = None,
     ) -> SnowflakePlan:
         """Returns a SnowflakePlan to materialize the child plan into a table.
 
@@ -1417,6 +1418,47 @@ def get_create_and_insert_plan(child: SnowflakePlan, replace, error):
                 referenced_ctes=child.referenced_ctes,
             )
 
+        def get_overwrite_delete_insert_plan(child: SnowflakePlan):
+            """Build a plan for targeted delete + insert with transaction.
+
+            Deletes rows matching the overwrite_condition condition, then inserts
+            all rows from the source DataFrame. Wrapped in a transaction for atomicity.
+            """
+            child = self.add_result_scan_if_not_select(child)
+
+            return SnowflakePlan(
+                [
+                    *child.queries[:-1],
+                    Query("BEGIN TRANSACTION"),
+                    Query(
+                        delete_statement(
+                            table_name=full_table_name,
+                            condition=overwrite_condition,
+                            source_data=None,
+                        ),
+                        params=child.queries[-1].params,
+                        is_ddl_on_temp_object=is_temp_table_type,
+                    ),
+                    Query(
+                        insert_into_statement(
+                            table_name=full_table_name,
+                            child=child.queries[-1].sql,
+                            column_names=column_names,
+                        ),
+                        params=child.queries[-1].params,
+                        is_ddl_on_temp_object=is_temp_table_type,
+                    ),
+                    Query("COMMIT"),
+                ],
+                schema_query=None,
+                post_actions=child.post_actions,
+                expr_to_alias={},
+                source_plan=source_plan,
+                api_calls=child.api_calls,
+                session=self.session,
+                referenced_ctes=child.referenced_ctes,
+            )
+
         if mode == SaveMode.APPEND:
             assert table_exists is not None
             if table_exists:
@@ -1446,7 +1488,12 @@ def get_create_and_insert_plan(child: SnowflakePlan, replace, error):
                 return get_create_table_as_select_plan(child, replace=True, error=True)
 
         elif mode == SaveMode.OVERWRITE:
-            return get_create_table_as_select_plan(child, replace=True, error=True)
+            if overwrite_condition is not None and table_exists:
+                # Selective overwrite: delete matching rows, then insert
+                return get_overwrite_delete_insert_plan(child)
+            else:
+                # Default overwrite: drop and recreate table
+                return get_create_table_as_select_plan(child, replace=True, error=True)
 
         elif mode == SaveMode.IGNORE:
             return get_create_table_as_select_plan(child, replace=False, error=False)

src/snowflake/snowpark/_internal/analyzer/snowflake_plan_node.py

@@ -243,6 +243,7 @@ def __init__(
         copy_grants: bool = False,
         iceberg_config: Optional[dict] = None,
         table_exists: Optional[bool] = None,
+        overwrite_condition: Optional[Expression] = None,
     ) -> None:
         super().__init__()
 
@@ -267,6 +268,7 @@ def __init__(
         # whether the table already exists in the database
         # determines the compiled SQL for APPEND and TRUNCATE mode
         self.table_exists = table_exists
+        self.overwrite_condition = overwrite_condition
 
     @property
     def individual_node_complexity(self) -> Dict[PlanNodeCategory, int]:

src/snowflake/snowpark/_internal/proto/ast.proto

@@ -1,5 +1,5 @@
 // N.B. This file is generated by `//ir-dsl-c`. DO NOT EDIT!
-// Generated from `{[email protected]:snowflakedb/snowflake.git}/Snowpark/ast`.
+// Generated from `{[email protected]:snowflake-eng/snowflake.git}/Snowpark/ast`.
 
 syntax = "proto3";
 
@@ -987,7 +987,7 @@ message DataframeCollect {
   repeated Tuple_String_String statement_params = 7;
 }
 
-// dataframe-io.ir:165
+// dataframe-io.ir:167
 message DataframeCopyIntoTable {
   repeated Tuple_String_Expr copy_options = 1;
   Expr df = 2;
@@ -1011,7 +1011,7 @@ message DataframeCount {
   repeated Tuple_String_String statement_params = 4;
 }
 
-// dataframe-io.ir:148
+// dataframe-io.ir:150
 message DataframeCreateOrReplaceDynamicTable {
   repeated Expr clustering_keys = 1;
   google.protobuf.StringValue comment = 2;
@@ -1030,7 +1030,7 @@ message DataframeCreateOrReplaceDynamicTable {
   string warehouse = 15;
 }
 
-// dataframe-io.ir:139
+// dataframe-io.ir:141
 message DataframeCreateOrReplaceView {
   google.protobuf.StringValue comment = 1;
   bool copy_grants = 2;
@@ -2685,7 +2685,7 @@ message WindowSpecRowsBetween {
   WindowSpecExpr wnd = 4;
 }
 
-// dataframe-io.ir:116
+// dataframe-io.ir:118
 message WriteCopyIntoLocation {
   bool block = 1;
   repeated Tuple_String_Expr copy_options = 2;
@@ -2704,7 +2704,7 @@ message WriteCopyIntoLocation {
   Expr writer = 15;
 }
 
-// dataframe-io.ir:123
+// dataframe-io.ir:125
 message WriteCsv {
   bool block = 1;
   repeated Tuple_String_Expr copy_options = 2;
@@ -2731,15 +2731,15 @@ message WriteFile {
   }
 }
 
-// dataframe-io.ir:129
+// dataframe-io.ir:131
 message WriteInsertInto {
   bool overwrite = 1;
   SrcPosition src = 2;
   NameRef table_name = 3;
   Expr writer = 4;
 }
 
-// dataframe-io.ir:125
+// dataframe-io.ir:127
 message WriteJson {
   bool block = 1;
   repeated Tuple_String_Expr copy_options = 2;
@@ -2773,7 +2773,7 @@ message WritePandas {
   string table_type = 13;
 }
 
-// dataframe-io.ir:127
+// dataframe-io.ir:129
 message WriteParquet {
   bool block = 1;
   repeated Tuple_String_Expr copy_options = 2;
@@ -2790,7 +2790,7 @@ message WriteParquet {
   Expr writer = 13;
 }
 
-// dataframe-io.ir:121
+// dataframe-io.ir:123
 message WriteSave {
   bool block = 1;
   repeated Tuple_String_Expr copy_options = 2;
@@ -2821,9 +2821,11 @@ message WriteTable {
   repeated Tuple_String_Expr iceberg_config = 10;
   google.protobuf.Int64Value max_data_extension_time = 11;
   SaveMode mode = 12;
-  SrcPosition src = 13;
-  repeated Tuple_String_String statement_params = 14;
-  NameRef table_name = 15;
-  string table_type = 16;
-  Expr writer = 17;
+  Expr overwrite_condition = 13;
+  SrcPosition src = 14;
+  repeated Tuple_String_String statement_params = 15;
+  google.protobuf.BoolValue table_exists = 16;
+  NameRef table_name = 17;
+  string table_type = 18;
+  Expr writer = 19;
 }

src/snowflake/snowpark/dataframe_writer.py

@@ -48,7 +48,7 @@
     warning,
 )
 from snowflake.snowpark.async_job import AsyncJob, _AsyncResultType
-from snowflake.snowpark.column import Column, _to_col_if_str
+from snowflake.snowpark.column import Column, _to_col_if_str, _to_col_if_sql_expr
 from snowflake.snowpark.exceptions import SnowparkClientException
 from snowflake.snowpark.functions import sql_expr
 from snowflake.snowpark.mock._connection import MockServerConnection
@@ -256,6 +256,7 @@ def save_as_table(
             Dict[str, Union[str, Iterable[ColumnOrSqlExpr]]]
         ] = None,
         table_exists: Optional[bool] = None,
+        overwrite_condition: Optional[ColumnOrSqlExpr] = None,
         _emit_ast: bool = True,
         **kwargs: Optional[Dict[str, Any]],
     ) -> Optional[AsyncJob]:
@@ -330,7 +331,10 @@ def save_as_table(
                 * iceberg_version: Overrides the version of iceberg to use. Defaults to 2 when unset.
             table_exists: Optional parameter to specify if the table is known to exist or not.
                 Set to ``True`` if table exists, ``False`` if it doesn't, or ``None`` (default) for automatic detection.
-                Primarily useful for "append" and "truncate" modes to avoid running query for automatic detection.
+                Primarily useful for "append", "truncate", and "overwrite" with overwrite_condition modes to avoid running query for automatic detection.
+            overwrite_condition: Specifies the overwrite condition to perform atomic targeted delete-insert.
+                Can only be used when ``mode`` is "overwrite" and the table exists. Rows matching the
+                condition are deleted from the target table, then all rows from the DataFrame are inserted.
 
 
         Example 1::
@@ -364,6 +368,21 @@ def save_as_table(
             ...     "partition_by": ["a", bucket(3, col("b"))],
             ... }
             >>> df.write.mode("overwrite").save_as_table("my_table", iceberg_config=iceberg_config) # doctest: +SKIP
+
+        Example 3::
+
+            Using overwrite_condition for targeted delete and insert:
+
+            >>> from snowflake.snowpark.functions import col
+            >>> df = session.create_dataframe([[1, "a"], [2, "b"], [3, "c"]], schema=["id", "val"])
+            >>> df.write.mode("overwrite").save_as_table("my_table", table_type="temporary")
+            >>> session.table("my_table").order_by("id").collect()
+            [Row(ID=1, VAL='a'), Row(ID=2, VAL='b'), Row(ID=3, VAL='c')]
+
+            >>> new_df = session.create_dataframe([[2, "updated2"], [5, "updated5"]], schema=["id", "val"])
+            >>> new_df.write.mode("overwrite").save_as_table("my_table", overwrite_condition="id = 1 or val = 'b'")
+            >>> session.table("my_table").order_by("id").collect()
+            [Row(ID=2, VAL='updated2'), Row(ID=3, VAL='c'), Row(ID=5, VAL='updated5')]
         """
 
         statement_params = track_data_source_statement_params(
@@ -392,6 +411,8 @@ def save_as_table(
             # change_tracking: Optional[bool] = None,
             # copy_grants: bool = False,
             # iceberg_config: Optional[dict] = None,
+            # table_exists: Optional[bool] = None,
+            # overwrite_condition: Optional[ColumnOrSqlExpr] = None,
 
             build_table_name(expr.table_name, table_name)
 
@@ -433,6 +454,12 @@ def save_as_table(
                     t = expr.iceberg_config.add()
                     t._1 = k
                     build_expr_from_python_val(t._2, v)
+            if table_exists is not None:
+                expr.table_exists.value = table_exists
+            if overwrite_condition is not None:
+                build_expr_from_snowpark_column_or_sql_str(
+                    expr.overwrite_condition, overwrite_condition
+                )
 
             self._dataframe._session._ast_batch.eval(stmt)
 
@@ -486,18 +513,33 @@ def save_as_table(
                     f"Unsupported table type. Expected table types: {SUPPORTED_TABLE_TYPES}"
                 )
 
+            # overwrite_condition must be used with OVERWRITE mode only
+            if overwrite_condition is not None and save_mode != SaveMode.OVERWRITE:
+                raise ValueError(
+                    f"'overwrite_condition' is only supported with mode='overwrite'. "
+                    f"Got mode='{save_mode.value}'."
+                )
+
+            overwrite_condition_expr = (
+                _to_col_if_sql_expr(
+                    overwrite_condition, "DataFrameWriter.save_as_table"
+                )._expression
+                if overwrite_condition is not None
+                else None
+            )
+
             session = self._dataframe._session
+            needs_table_exists_check = save_mode in [
+                SaveMode.APPEND,
+                SaveMode.TRUNCATE,
+            ] or (save_mode == SaveMode.OVERWRITE and overwrite_condition is not None)
             if (
                 table_exists is None
                 and not isinstance(session._conn, MockServerConnection)
-                and save_mode
-                in [
-                    SaveMode.APPEND,
-                    SaveMode.TRUNCATE,
-                ]
+                and needs_table_exists_check
             ):
                 # whether the table already exists in the database
-                # determines the compiled SQL for APPEND and TRUNCATE mode
+                # determines the compiled SQL for APPEND, TRUNCATE, and OVERWRITE with overwrite_condition
                 # if the table does not exist, we need to create it first;
                 # if the table exists, we can skip the creation step and insert data directly
                 table_exists = session._table_exists(table_name)
@@ -518,6 +560,7 @@ def save_as_table(
                 copy_grants,
                 iceberg_config,
                 table_exists,
+                overwrite_condition_expr,
             )
             snowflake_plan = session._analyzer.resolve(create_table_logic_plan)
             result = session._conn.execute(