pixeltable 0.4.0rc3__py3-none-any.whl → 0.4.20__py3-none-any.whl

pixeltable/plan.py

@@ -3,16 +3,16 @@ from __future__ import annotations
 import dataclasses
 import enum
 from textwrap import dedent
-from typing import Any, Iterable, Literal, NamedTuple, Optional, Sequence
+from typing import Any, Iterable, Literal, Sequence, cast
 from uuid import UUID
 
+import pgvector.sqlalchemy  # type: ignore[import-untyped]
 import sqlalchemy as sql
 
 import pixeltable as pxt
 from pixeltable import catalog, exceptions as excs, exec, exprs
 from pixeltable.catalog import Column, TableVersionHandle
 from pixeltable.exec.sql_node import OrderByClause, OrderByItem, combine_order_by_clauses, print_order_by_clause
-from pixeltable.utils.sample import sample_key
 
 
 def _is_agg_fn_call(e: exprs.Expr) -> bool:
@@ -66,7 +66,7 @@ class JoinClause:
     """Corresponds to a single 'JOIN ... ON (...)' clause in a SELECT statement; excludes the joined table."""
 
     join_type: JoinType
-    join_predicate: Optional[exprs.Expr]  # None for join_type == CROSS
+    join_predicate: exprs.Expr | None  # None for join_type == CROSS
 
 
 @dataclasses.dataclass
@@ -86,25 +86,20 @@ class FromClause:
 class SampleClause:
     """Defines a sampling clause for a table."""
 
-    version: Optional[int]
-    n: Optional[int]
-    n_per_stratum: Optional[int]
-    fraction: Optional[float]
-    seed: Optional[int]
-    stratify_exprs: Optional[list[exprs.Expr]]
-
-    # This seed value is used if one is not supplied
-    DEFAULT_SEED = 0
+    version: int | None
+    n: int | None
+    n_per_stratum: int | None
+    fraction: float | None
+    seed: int | None
+    stratify_exprs: list[exprs.Expr] | None
 
     # The version of the hashing algorithm used for ordering and fractional sampling.
     CURRENT_VERSION = 1
 
     def __post_init__(self) -> None:
-        """If no version was provided, provide the default version"""
+        # If no version was provided, provide the default version
         if self.version is None:
             self.version = self.CURRENT_VERSION
-        if self.seed is None:
-            self.seed = self.DEFAULT_SEED
 
     @property
     def is_stratified(self) -> bool:
@@ -159,16 +154,6 @@ class SampleClause:
         return format(threshold_int, '08x') + 'ffffffffffffffffffffffff'
 
 
-class SamplingClauses(NamedTuple):
-    """Clauses provided when rewriting a SampleClause"""
-
-    where: exprs.Expr
-    group_by_clause: Optional[list[exprs.Expr]]
-    order_by_clause: Optional[list[tuple[exprs.Expr, bool]]]
-    limit: Optional[exprs.Expr]
-    sample_clause: Optional[SampleClause]
-
-
 class Analyzer:
     """
     Performs semantic analysis of a query and stores the analysis state.
@@ -177,17 +162,19 @@ class Analyzer:
     from_clause: FromClause
     all_exprs: list[exprs.Expr]  # union of all exprs, aside from sql_where_clause
     select_list: list[exprs.Expr]
-    group_by_clause: Optional[list[exprs.Expr]]  # None for non-aggregate queries; [] for agg query w/o grouping
+    group_by_clause: list[exprs.Expr] | None  # None for non-aggregate queries; [] for agg query w/o grouping
     grouping_exprs: list[exprs.Expr]  # [] for non-aggregate queries or agg query w/o grouping
     order_by_clause: OrderByClause
+    stratify_exprs: list[exprs.Expr]  # [] if no stratiifcation is required
+    sample_clause: SampleClause | None  # None if no sampling clause is present
 
     sql_elements: exprs.SqlElementCache
 
     # Where clause of the Select stmt of the SQL scan
-    sql_where_clause: Optional[exprs.Expr]
+    sql_where_clause: exprs.Expr | None
 
     # filter predicate applied to output rows of the SQL scan
-    filter: Optional[exprs.Expr]
+    filter: exprs.Expr | None
 
     agg_fn_calls: list[exprs.FunctionCall]  # grouping aggregation (ie, not window functions)
     window_fn_calls: list[exprs.FunctionCall]
@@ -197,9 +184,10 @@ class Analyzer:
         self,
         from_clause: FromClause,
         select_list: Sequence[exprs.Expr],
-        where_clause: Optional[exprs.Expr] = None,
-        group_by_clause: Optional[list[exprs.Expr]] = None,
-        order_by_clause: Optional[list[tuple[exprs.Expr, bool]]] = None,
+        where_clause: exprs.Expr | None = None,
+        group_by_clause: list[exprs.Expr] | None = None,
+        order_by_clause: list[tuple[exprs.Expr, bool]] | None = None,
+        sample_clause: SampleClause | None = None,
     ):
         if order_by_clause is None:
             order_by_clause = []
@@ -213,6 +201,11 @@ class Analyzer:
         self.group_by_clause = (
             [e.resolve_computed_cols() for e in group_by_clause] if group_by_clause is not None else None
         )
+        self.sample_clause = sample_clause
+        if self.sample_clause is not None and self.sample_clause.is_stratified:
+            self.stratify_exprs = [e.resolve_computed_cols() for e in sample_clause.stratify_exprs]
+        else:
+            self.stratify_exprs = []
         self.order_by_clause = [OrderByItem(e.resolve_computed_cols(), asc) for e, asc in order_by_clause]
 
         self.sql_where_clause = None
@@ -228,8 +221,11 @@ class Analyzer:
                 self.all_exprs.append(join_clause.join_predicate)
         if self.group_by_clause is not None:
             self.all_exprs.extend(self.group_by_clause)
+        self.all_exprs.extend(self.stratify_exprs)
         self.all_exprs.extend(e for e, _ in self.order_by_clause)
         if self.filter is not None:
+            if sample_clause is not None:
+                raise excs.Error(f'Filter {self.filter} not expressible in SQL')
             self.all_exprs.append(self.filter)
 
         self.agg_order_by = []
@@ -334,7 +330,7 @@ class Analyzer:
         row_builder.set_slot_idxs(self.agg_fn_calls)
         row_builder.set_slot_idxs(self.agg_order_by)
 
-    def get_window_fn_ob_clause(self) -> Optional[OrderByClause]:
+    def get_window_fn_ob_clause(self) -> OrderByClause | None:
         clause: list[OrderByClause] = []
         for fn_call in self.window_fn_calls:
             # window functions require ordering by the group_by/order_by clauses
@@ -352,7 +348,7 @@ class Analyzer:
 class Planner:
     # TODO: create an exec.CountNode and change this to create_count_plan()
     @classmethod
-    def create_count_stmt(cls, tbl: catalog.TableVersionPath, where_clause: Optional[exprs.Expr] = None) -> sql.Select:
+    def create_count_stmt(cls, tbl: catalog.TableVersionPath, where_clause: exprs.Expr | None = None) -> sql.Select:
         stmt = sql.select(sql.func.count().label('all_count'))
         refd_tbl_ids: set[UUID] = set()
         if where_clause is not None:
@@ -378,21 +374,14 @@ class Planner:
 
         cls.__check_valid_columns(tbl, stored_cols, 'inserted into')
 
-        row_builder = exprs.RowBuilder([], stored_cols, [])
+        row_builder = exprs.RowBuilder([], stored_cols, [], tbl)
 
         # create InMemoryDataNode for 'rows'
         plan: exec.ExecNode = exec.InMemoryDataNode(
             TableVersionHandle(tbl.id, tbl.effective_version), rows, row_builder, tbl.next_row_id
         )
 
-        media_input_col_info = [
-            exprs.ColumnSlotIdx(col_ref.col, col_ref.slot_idx)
-            for col_ref in row_builder.input_exprs
-            if isinstance(col_ref, exprs.ColumnRef) and col_ref.col_type.is_media_type()
-        ]
-        if len(media_input_col_info) > 0:
-            # prefetch external files for all input column refs
-            plan = exec.CachePrefetchNode(tbl.id, media_input_col_info, input=plan)
+        plan = cls._add_prefetch_node(tbl.id, row_builder.input_exprs, input_node=plan)
 
         computed_exprs = row_builder.output_exprs - row_builder.input_exprs
         if len(computed_exprs) > 0:
@@ -400,10 +389,9 @@ class Planner:
             plan = exec.ExprEvalNode(
                 row_builder, computed_exprs, plan.output_exprs, input=plan, maintain_input_order=False
             )
+        if any(c.col_type.is_json_type() or c.col_type.is_array_type() for c in stored_cols):
+            plan = exec.CellMaterializationNode(plan)
 
-        stored_col_info = row_builder.output_slot_idxs()
-        stored_img_col_info = [info for info in stored_col_info if info.col.col_type.is_image_type()]
-        plan.set_stored_img_cols(stored_img_col_info)
         plan.set_ctx(
             exec.ExecContext(
                 row_builder,
@@ -413,10 +401,12 @@ class Planner:
                 ignore_errors=ignore_errors,
             )
         )
+        plan = cls._add_save_node(plan)
+
         return plan
 
     @classmethod
-    def rowid_columns(cls, target: TableVersionHandle, num_rowid_cols: Optional[int] = None) -> list[exprs.Expr]:
+    def rowid_columns(cls, target: TableVersionHandle, num_rowid_cols: int | None = None) -> list[exprs.Expr]:
         """Return list of RowidRef for the given number of associated rowids"""
         if num_rowid_cols is None:
             num_rowid_cols = target.get().num_rowid_columns()
@@ -430,14 +420,17 @@ class Planner:
         plan = df._create_query_plan()  # ExecNode constructed by the DataFrame
 
         # Modify the plan RowBuilder to register the output columns
+        needs_cell_materialization = False
         for col_name, expr in zip(df.schema.keys(), df._select_list_exprs):
             assert col_name in tbl.cols_by_name
             col = tbl.cols_by_name[col_name]
             plan.row_builder.add_table_column(col, expr.slot_idx)
+            needs_cell_materialization = (
+                needs_cell_materialization or col.col_type.is_json_type() or col.col_type.is_array_type()
+            )
 
-        stored_col_info = plan.row_builder.output_slot_idxs()
-        stored_img_col_info = [info for info in stored_col_info if info.col.col_type.is_image_type()]
-        plan.set_stored_img_cols(stored_img_col_info)
+        if needs_cell_materialization:
+            plan = exec.CellMaterializationNode(plan)
 
         plan.set_ctx(
             exec.ExecContext(
@@ -454,16 +447,18 @@ class Planner:
         tbl: catalog.TableVersionPath,
         update_targets: dict[catalog.Column, exprs.Expr],
         recompute_targets: list[catalog.Column],
-        where_clause: Optional[exprs.Expr],
+        where_clause: exprs.Expr | None,
         cascade: bool,
     ) -> tuple[exec.ExecNode, list[str], list[catalog.Column]]:
         """Creates a plan to materialize updated rows.
+
         The plan:
         - retrieves rows that are visible at the current version of the table
         - materializes all stored columns and the update targets
         - if cascade is True, recomputes all computed columns that transitively depend on the updated columns
           and copies the values of all other stored columns
         - if cascade is False, copies all columns that aren't update targets from the original rows
+
         Returns:
             - root node of the plan
             - list of qualified column names that are getting updated
@@ -473,26 +468,33 @@ class Planner:
         assert isinstance(tbl, catalog.TableVersionPath)
         target = tbl.tbl_version.get()  # the one we need to update
         updated_cols = list(update_targets.keys())
+        recomputed_cols: set[Column]
         if len(recompute_targets) > 0:
-            recomputed_cols = set(recompute_targets)
+            assert len(update_targets) == 0
+            recomputed_cols = {*recompute_targets}
+            if cascade:
+                recomputed_cols |= target.get_dependent_columns(recomputed_cols)
         else:
             recomputed_cols = target.get_dependent_columns(updated_cols) if cascade else set()
-            # regardless of cascade, we need to update all indices on any updated column
-            idx_val_cols = target.get_idx_val_columns(updated_cols)
-            recomputed_cols.update(idx_val_cols)
-            # we only need to recompute stored columns (unstored ones are substituted away)
-            recomputed_cols = {c for c in recomputed_cols if c.is_stored}
+        # regardless of cascade, we need to update all indices on any updated/recomputed column
+        modified_base_cols = [c for c in set(updated_cols) | recomputed_cols if c.get_tbl().id == target.id]
+        idx_val_cols = target.get_idx_val_columns(modified_base_cols)
+        recomputed_cols.update(idx_val_cols)
+        # we only need to recompute stored columns (unstored ones are substituted away)
+        recomputed_cols = {c for c in recomputed_cols if c.is_stored}
 
         cls.__check_valid_columns(tbl.tbl_version.get(), recomputed_cols, 'updated in')
 
-        recomputed_base_cols = {col for col in recomputed_cols if col.tbl.id == tbl.tbl_version.id}
+        # our query plan
+        # - evaluates the update targets and recomputed columns
+        # - copies all other stored columns
+        recomputed_base_cols = {col for col in recomputed_cols if col.get_tbl().id == tbl.tbl_version.id}
         copied_cols = [
             col
             for col in target.cols_by_id.values()
             if col.is_stored and col not in updated_cols and col not in recomputed_base_cols
         ]
-        select_list: list[exprs.Expr] = [exprs.ColumnRef(col) for col in copied_cols]
-        select_list.extend(update_targets.values())
+        select_list: list[exprs.Expr] = list(update_targets.values())
 
         recomputed_exprs = [
             c.value_expr.copy().resolve_computed_cols(resolve_cols=recomputed_base_cols) for c in recomputed_base_cols
@@ -503,13 +505,25 @@ class Planner:
         select_list.extend(recomputed_exprs)
 
         # we need to retrieve the PK columns of the existing rows
-        plan = cls.create_query_plan(FromClause(tbls=[tbl]), select_list, where_clause=where_clause, ignore_errors=True)
-        all_base_cols = copied_cols + updated_cols + list(recomputed_base_cols)  # same order as select_list
+        plan = cls.create_query_plan(
+            FromClause(tbls=[tbl]),
+            select_list=select_list,
+            columns=copied_cols,
+            where_clause=where_clause,
+            ignore_errors=True,
+        )
+        evaluated_cols = updated_cols + list(recomputed_base_cols)  # same order as select_list
         # update row builder with column information
-        for i, col in enumerate(all_base_cols):
+        plan.row_builder.add_table_columns(copied_cols)
+        for i, col in enumerate(evaluated_cols):
             plan.row_builder.add_table_column(col, select_list[i].slot_idx)
+        plan.ctx.num_computed_exprs = len(recomputed_exprs)
+
+        plan = cls._add_cell_materialization_node(plan)
+        plan = cls._add_save_node(plan)
+
         recomputed_user_cols = [c for c in recomputed_cols if c.name is not None]
-        return plan, [f'{c.tbl.name}.{c.name}' for c in updated_cols + recomputed_user_cols], recomputed_user_cols
+        return plan, [f'{c.get_tbl().name}.{c.name}' for c in updated_cols + recomputed_user_cols], recomputed_user_cols
 
     @classmethod
     def __check_valid_columns(
@@ -529,6 +543,79 @@ class Planner:
                     .format(validation_error=col.value_expr.validation_error)
                 )
 
+    @classmethod
+    def _cell_md_col_refs(cls, expr_list: Iterable[exprs.Expr]) -> list[exprs.ColumnRef]:
+        """Return list of ColumnRefs that need their cellmd values for reconstruction"""
+        json_col_refs = list(
+            exprs.Expr.list_subexprs(
+                expr_list,
+                expr_class=exprs.ColumnRef,
+                filter=lambda e: cast(exprs.ColumnRef, e).col.col_type.is_json_type(),
+                traverse_matches=False,
+            )
+        )
+
+        def needs_reconstruction(e: exprs.Expr) -> bool:
+            assert isinstance(e, exprs.ColumnRef)
+            # Vector-typed array columns are used for vector indexes, and are stored in the db
+            return e.col.col_type.is_array_type() and not isinstance(e.col.sa_col_type, pgvector.sqlalchemy.Vector)
+
+        array_col_refs = list(
+            exprs.Expr.list_subexprs(
+                expr_list, expr_class=exprs.ColumnRef, filter=needs_reconstruction, traverse_matches=False
+            )
+        )
+
+        return json_col_refs + array_col_refs
+
+    @classmethod
+    def _add_cell_materialization_node(cls, input: exec.ExecNode) -> exec.ExecNode:
+        # we need a CellMaterializationNode if any of the evaluated output columns are json or array-typed
+        has_target_cols = any(
+            col.col_type.is_json_type() or col.col_type.is_array_type()
+            for col, slot_idx in input.row_builder.table_columns.items()
+            if slot_idx is not None
+        )
+        if has_target_cols:
+            return exec.CellMaterializationNode(input)
+        else:
+            return input
+
+    @classmethod
+    def _add_cell_reconstruction_node(cls, expr_list: list[exprs.Expr], input: exec.ExecNode) -> exec.ExecNode:
+        """
+        Add a CellReconstructionNode, if required by any of the exprs in expr_list.
+
+        Cell reconstruction is required for
+        1) all json-typed ColumnRefs that are not used as part of a JsonPath (the latter does its own reconstruction)
+           or as part of a ColumnPropertyRef
+        2) all array-typed ColumnRefs that are not used as part of a ColumnPropertyRef
+        """
+
+        def json_filter(e: exprs.Expr) -> bool:
+            if isinstance(e, exprs.JsonPath):
+                return not e.is_relative_path() and isinstance(e.anchor, exprs.ColumnRef)
+            if isinstance(e, exprs.ColumnPropertyRef):
+                return e.col_ref.col.col_type.is_json_type()
+            return isinstance(e, exprs.ColumnRef) and e.col.col_type.is_json_type()
+
+        def array_filter(e: exprs.Expr) -> bool:
+            if isinstance(e, exprs.ColumnPropertyRef):
+                return e.col_ref.col.col_type.is_array_type()
+            if not isinstance(e, exprs.ColumnRef):
+                return False
+            # Vector-typed array columns are used for vector indexes, and are stored in the db
+            return e.col.col_type.is_array_type() and not isinstance(e.col.sa_col_type, pgvector.sqlalchemy.Vector)
+
+        json_candidates = list(exprs.Expr.list_subexprs(expr_list, filter=json_filter, traverse_matches=False))
+        json_refs = [e for e in json_candidates if isinstance(e, exprs.ColumnRef)]
+        array_candidates = list(exprs.Expr.list_subexprs(expr_list, filter=array_filter, traverse_matches=False))
+        array_refs = [e for e in array_candidates if isinstance(e, exprs.ColumnRef)]
+        if len(json_refs) > 0 or len(array_refs) > 0:
+            return exec.CellReconstructionNode(json_refs, array_refs, input.row_builder, input=input)
+        else:
+            return input
+
     @classmethod
     def create_batch_update_plan(
         cls,
@@ -547,8 +634,8 @@ class Planner:
         """
         assert isinstance(tbl, catalog.TableVersionPath)
         target = tbl.tbl_version.get()  # the one we need to update
-        sa_key_cols: list[sql.Column] = []
-        key_vals: list[tuple] = []
+        sa_key_cols: list[sql.Column]
+        key_vals: list[tuple]
         if len(rowids) > 0:
             sa_key_cols = target.store_tbl.rowid_columns()
             key_vals = rowids
@@ -561,18 +648,18 @@ class Planner:
         updated_cols = batch[0].keys() - target.primary_key_columns()
         recomputed_cols = target.get_dependent_columns(updated_cols) if cascade else set()
         # regardless of cascade, we need to update all indices on any updated column
-        idx_val_cols = target.get_idx_val_columns(updated_cols)
+        modified_base_cols = [c for c in set(updated_cols) | recomputed_cols if c.get_tbl().id == target.id]
+        idx_val_cols = target.get_idx_val_columns(modified_base_cols)
         recomputed_cols.update(idx_val_cols)
         # we only need to recompute stored columns (unstored ones are substituted away)
         recomputed_cols = {c for c in recomputed_cols if c.is_stored}
-        recomputed_base_cols = {col for col in recomputed_cols if col.tbl.id == target.id}
+        recomputed_base_cols = {col for col in recomputed_cols if col.get_tbl().id == target.id}
         copied_cols = [
             col
             for col in target.cols_by_id.values()
             if col.is_stored and col not in updated_cols and col not in recomputed_base_cols
         ]
-        select_list: list[exprs.Expr] = [exprs.ColumnRef(col) for col in copied_cols]
-        select_list.extend(exprs.ColumnRef(col) for col in updated_cols)
+        select_list: list[exprs.Expr] = [exprs.ColumnRef(col) for col in updated_cols]
 
         recomputed_exprs = [
             c.value_expr.copy().resolve_computed_cols(resolve_cols=recomputed_base_cols) for c in recomputed_base_cols
@@ -588,25 +675,39 @@ class Planner:
         sql_exprs = list(
             exprs.Expr.list_subexprs(analyzer.all_exprs, filter=analyzer.sql_elements.contains, traverse_matches=False)
         )
-        row_builder = exprs.RowBuilder(analyzer.all_exprs, [], sql_exprs)
+        row_builder = exprs.RowBuilder(analyzer.all_exprs, [], sql_exprs, target)
         analyzer.finalize(row_builder)
-        sql_lookup_node = exec.SqlLookupNode(tbl, row_builder, sql_exprs, sa_key_cols, key_vals)
+
+        cell_md_col_refs = cls._cell_md_col_refs(sql_exprs)
+        sql_lookup_node = exec.SqlLookupNode(
+            tbl,
+            row_builder,
+            sql_exprs,
+            columns=copied_cols,
+            sa_key_cols=sa_key_cols,
+            key_vals=key_vals,
+            cell_md_col_refs=cell_md_col_refs,
+        )
         col_vals = [{col: row[col].val for col in updated_cols} for row in batch]
         row_update_node = exec.RowUpdateNode(tbl, key_vals, len(rowids) > 0, col_vals, row_builder, sql_lookup_node)
         plan: exec.ExecNode = row_update_node
         if not cls._is_contained_in(analyzer.select_list, sql_exprs):
             # we need an ExprEvalNode to evaluate the remaining output exprs
             plan = exec.ExprEvalNode(row_builder, analyzer.select_list, sql_exprs, input=plan)
+
         # update row builder with column information
-        all_base_cols = copied_cols + list(updated_cols) + list(recomputed_base_cols)  # same order as select_list
+        evaluated_cols = list(updated_cols) + list(recomputed_base_cols)  # same order as select_list
         row_builder.set_slot_idxs(select_list, remove_duplicates=False)
-        for i, col in enumerate(all_base_cols):
+        plan.row_builder.add_table_columns(copied_cols)
+        for i, col in enumerate(evaluated_cols):
             plan.row_builder.add_table_column(col, select_list[i].slot_idx)
-
-        ctx = exec.ExecContext(row_builder)
-        # we're returning everything to the user, so we might as well do it in a single batch
+        ctx = exec.ExecContext(row_builder, num_computed_exprs=len(recomputed_exprs))
+        # TODO: correct batch size?
         ctx.batch_size = 0
         plan.set_ctx(ctx)
+
+        plan = cls._add_cell_materialization_node(plan)
+        plan = cls._add_save_node(plan)
         recomputed_user_cols = [c for c in recomputed_cols if c.name is not None]
         return (
             plan,
@@ -656,13 +757,13 @@ class Planner:
             ignore_errors=True,
             exact_version_only=view.get_bases(),
         )
-        for i, col in enumerate(copied_cols + list(recomputed_cols)):  # same order as select_list
+        plan.ctx.num_computed_exprs = len(recomputed_exprs)
+        materialized_cols = copied_cols + list(recomputed_cols)  # same order as select_list
+        for i, col in enumerate(materialized_cols):
             plan.row_builder.add_table_column(col, select_list[i].slot_idx)
-        # TODO: avoid duplication with view_load_plan() logic (where does this belong?)
-        stored_img_col_info = [
-            info for info in plan.row_builder.output_slot_idxs() if info.col.col_type.is_image_type()
-        ]
-        plan.set_stored_img_cols(stored_img_col_info)
+        plan = cls._add_cell_materialization_node(plan)
+        plan = cls._add_save_node(plan)
+
         return plan
 
     @classmethod
@@ -691,25 +792,13 @@ class Planner:
         # 2. for component views: iterator args
         iterator_args = [target.iterator_args] if target.iterator_args is not None else []
 
-        # If this contains a sample specification, modify / create where, group_by, order_by, and limit clauses
         from_clause = FromClause(tbls=[view.base])
-        where, group_by_clause, order_by_clause, limit, sample_clause = cls.create_sample_clauses(
-            from_clause, target.sample_clause, target.predicate, None, [], None
-        )
-
-        # if we're propagating an insert, we only want to see those base rows that were created for the current version
         base_analyzer = Analyzer(
-            from_clause,
-            iterator_args,
-            where_clause=where,
-            group_by_clause=group_by_clause,
-            order_by_clause=order_by_clause,
+            from_clause, iterator_args, where_clause=target.predicate, sample_clause=target.sample_clause
         )
-        row_builder = exprs.RowBuilder(base_analyzer.all_exprs, stored_cols, [])
-
-        if target.sample_clause is not None and base_analyzer.filter is not None:
-            raise excs.Error(f'Filter {base_analyzer.filter} not expressible in SQL')
+        row_builder = exprs.RowBuilder(base_analyzer.all_exprs, stored_cols, [], target)
 
+        # if we're propagating an insert, we only want to see those base rows that were created for the current version
         # execution plan:
         # 1. materialize exprs computed from the base that are needed for stored view columns
         # 2. if it's an iterator view, expand the base rows into component rows
@@ -723,19 +812,13 @@ class Planner:
 
         # Create a new analyzer reflecting exactly what is required from the base table
         base_analyzer = Analyzer(
-            from_clause,
-            base_output_exprs,
-            where_clause=where,
-            group_by_clause=group_by_clause,
-            order_by_clause=order_by_clause,
+            from_clause, base_output_exprs, where_clause=target.predicate, sample_clause=target.sample_clause
         )
         base_eval_ctx = row_builder.create_eval_ctx(base_analyzer.all_exprs)
         plan = cls._create_query_plan(
             row_builder=row_builder,
             analyzer=base_analyzer,
             eval_ctx=base_eval_ctx,
-            limit=limit,
-            sample_clause=sample_clause,
             with_pk=True,
             exact_version_only=view.get_bases() if propagates_insert else [],
         )
@@ -747,10 +830,12 @@ class Planner:
                 row_builder, output_exprs=view_output_exprs, input_exprs=base_output_exprs, input=plan
             )
 
-        stored_img_col_info = [info for info in row_builder.output_slot_idxs() if info.col.col_type.is_image_type()]
-        plan.set_stored_img_cols(stored_img_col_info)
         exec_ctx.ignore_errors = True
         plan.set_ctx(exec_ctx)
+        if any(c.col_type.is_json_type() or c.col_type.is_array_type() for c in stored_cols):
+            plan = exec.CellMaterializationNode(plan)
+        plan = cls._add_save_node(plan)
+
         return plan, len(row_builder.default_eval_ctx.target_exprs)
 
     @classmethod
@@ -761,7 +846,7 @@ class Planner:
                 raise excs.Error(f'Join predicate {join_clause.join_predicate} not expressible in SQL')
 
     @classmethod
-    def _create_combined_ordering(cls, analyzer: Analyzer, verify_agg: bool) -> Optional[OrderByClause]:
+    def _create_combined_ordering(cls, analyzer: Analyzer, verify_agg: bool) -> OrderByClause | None:
         """Verify that the various ordering requirements don't conflict and return a combined ordering"""
         ob_clauses: list[OrderByClause] = [analyzer.order_by_clause.copy()]
 
@@ -795,22 +880,29 @@ class Planner:
             combined_ordering = combined
         return combined_ordering
 
+    @classmethod
+    def _add_save_node(cls, input_node: exec.ExecNode) -> exec.ExecNode:
+        """Add an ObjectStoreSaveNode, if needed."""
+        media_col_info = input_node.row_builder.media_output_col_info
+        if len(media_col_info) == 0:
+            return input_node
+        else:
+            return exec.ObjectStoreSaveNode(media_col_info, input_node)
+
     @classmethod
     def _is_contained_in(cls, l1: Iterable[exprs.Expr], l2: Iterable[exprs.Expr]) -> bool:
         """Returns True if l1 is contained in l2"""
         return {e.id for e in l1} <= {e.id for e in l2}
 
     @classmethod
-    def _insert_prefetch_node(
-        cls, tbl_id: UUID, row_builder: exprs.RowBuilder, input_node: exec.ExecNode
+    def _add_prefetch_node(
+        cls, tbl_id: UUID, expressions: Iterable[exprs.Expr], input_node: exec.ExecNode
     ) -> exec.ExecNode:
-        """Returns a CachePrefetchNode into the plan if needed, otherwise returns input"""
+        """Add a CachePrefetch node, if needed."""
         # we prefetch external files for all media ColumnRefs, even those that aren't part of the dependencies
         # of output_exprs: if unstored iterator columns are present, we might need to materialize ColumnRefs that
         # aren't explicitly captured as dependencies
-        media_col_refs = [
-            e for e in list(row_builder.unique_exprs) if isinstance(e, exprs.ColumnRef) and e.col_type.is_media_type()
-        ]
+        media_col_refs = [e for e in expressions if isinstance(e, exprs.ColumnRef) and e.col_type.is_media_type()]
         if len(media_col_refs) == 0:
             return input_node
         # we need to prefetch external files for media column types
@@ -818,101 +910,52 @@ class Planner:
         prefetch_node = exec.CachePrefetchNode(tbl_id, file_col_info, input_node)
         return prefetch_node
 
-    @classmethod
-    def create_sample_clauses(
-        cls,
-        from_clause: FromClause,
-        sample_clause: SampleClause,
-        where_clause: Optional[exprs.Expr],
-        group_by_clause: Optional[list[exprs.Expr]],
-        order_by_clause: Optional[list[tuple[exprs.Expr, bool]]],
-        limit: Optional[exprs.Expr],
-    ) -> SamplingClauses:
-        """tuple[
-            exprs.Expr,
-            Optional[list[exprs.Expr]],
-            Optional[list[tuple[exprs.Expr, bool]]],
-            Optional[exprs.Expr],
-            Optional[SampleClause],
-        ]:"""
-        """Construct clauses required for sampling under various conditions.
-        If there is no sampling, then return the original clauses.
-        If the sample is stratified, then return only the group by clause. The rest of the
-        mechanism for stratified sampling is provided by the SampleSqlNode.
-        If the sample is non-stratified, then rewrite the query to accommodate the supplied where clause,
-        and provide the other clauses required for sampling
-        """
-
-        # If no sample clause, return the original clauses
-        if sample_clause is None:
-            return SamplingClauses(where_clause, group_by_clause, order_by_clause, limit, None)
-
-        # If the sample clause is stratified, create a group by clause
-        if sample_clause.is_stratified:
-            group_by = sample_clause.stratify_exprs
-            # Note that limit is not possible here
-            return SamplingClauses(where_clause, group_by, order_by_clause, None, sample_clause)
-
-        else:
-            # If non-stratified sampling, construct a where clause, order_by, and limit clauses
-            # Construct an expression for sorting rows and limiting row counts
-            s_key = sample_key(
-                exprs.Literal(sample_clause.seed), *cls.rowid_columns(from_clause._first_tbl.tbl_version)
-            )
-
-            # Construct a suitable where clause
-            where = where_clause
-            if sample_clause.fraction is not None:
-                fraction_md5_hex = exprs.Expr.from_object(
-                    sample_clause.fraction_to_md5_hex(float(sample_clause.fraction))
-                )
-                f_where = s_key < fraction_md5_hex
-                where = where & f_where if where is not None else f_where
-
-            order_by: list[tuple[exprs.Expr, bool]] = [(s_key, True)]
-            limit = exprs.Literal(sample_clause.n)
-            # Note that group_by is not possible here
-            return SamplingClauses(where, None, order_by, limit, None)
-
     @classmethod
     def create_query_plan(
         cls,
         from_clause: FromClause,
-        select_list: Optional[list[exprs.Expr]] = None,
-        where_clause: Optional[exprs.Expr] = None,
-        group_by_clause: Optional[list[exprs.Expr]] = None,
-        order_by_clause: Optional[list[tuple[exprs.Expr, bool]]] = None,
-        limit: Optional[exprs.Expr] = None,
-        sample_clause: Optional[SampleClause] = None,
+        select_list: list[exprs.Expr] | None = None,
+        columns: list[catalog.Column] | None = None,
+        where_clause: exprs.Expr | None = None,
+        group_by_clause: list[exprs.Expr] | None = None,
+        order_by_clause: list[tuple[exprs.Expr, bool]] | None = None,
+        limit: exprs.Expr | None = None,
+        sample_clause: SampleClause | None = None,
         ignore_errors: bool = False,
-        exact_version_only: Optional[list[catalog.TableVersionHandle]] = None,
+        exact_version_only: list[catalog.TableVersionHandle] | None = None,
     ) -> exec.ExecNode:
-        """Return plan for executing a query.
+        """
+        Return plan for executing a query.
+
+        The plan:
+        - materializes the values of select_list exprs into their respective slots
+        - materializes cell values of 'columns' (and their cellmd, if applicable) into DataRow.cell_vals/cell_md
+
         Updates 'select_list' in place to make it executable.
         TODO: make exact_version_only a flag and use the versions from tbl
         """
         if select_list is None:
             select_list = []
+        if columns is None:
+            columns = []
         if order_by_clause is None:
             order_by_clause = []
         if exact_version_only is None:
             exact_version_only = []
 
-        # Modify clauses to include sample clause
-        where, group_by_clause, order_by_clause, limit, sample = cls.create_sample_clauses(
-            from_clause, sample_clause, where_clause, group_by_clause, order_by_clause, limit
-        )
-
         analyzer = Analyzer(
             from_clause,
             select_list,
-            where_clause=where,
+            where_clause=where_clause,
             group_by_clause=group_by_clause,
             order_by_clause=order_by_clause,
+            sample_clause=sample_clause,
         )
-        row_builder = exprs.RowBuilder(analyzer.all_exprs, [], [])
-        if sample_clause is not None and analyzer.filter is not None:
-            raise excs.Error(f'Filter {analyzer.filter} not expressible in SQL')
+        # If the from_clause has a single table, we can use it as the context table for the RowBuilder.
+        # Otherwise there is no context table, but that's ok, because the context table is only needed for
+        # table mutations, which can't happen during a join.
+        context_tbl = from_clause.tbls[0].tbl_version.get() if len(from_clause.tbls) == 1 else None
+        row_builder = exprs.RowBuilder(analyzer.all_exprs, [], [], context_tbl)
 
         analyzer.finalize(row_builder)
         # select_list: we need to materialize everything that's been collected
@@ -922,8 +965,8 @@ class Planner:
             row_builder=row_builder,
             analyzer=analyzer,
             eval_ctx=eval_ctx,
+            columns=columns,
             limit=limit,
-            sample_clause=sample,
             with_pk=True,
             exact_version_only=exact_version_only,
         )
@@ -938,10 +981,10 @@ class Planner:
         row_builder: exprs.RowBuilder,
         analyzer: Analyzer,
         eval_ctx: exprs.RowBuilder.EvalCtx,
-        limit: Optional[exprs.Expr] = None,
-        sample_clause: Optional[SampleClause] = None,
+        columns: list[catalog.Column] | None = None,
+        limit: exprs.Expr | None = None,
         with_pk: bool = False,
-        exact_version_only: Optional[list[catalog.TableVersionHandle]] = None,
+        exact_version_only: list[catalog.TableVersionHandle] | None = None,
     ) -> exec.ExecNode:
         """
         Create plan to materialize eval_ctx.
@@ -951,6 +994,8 @@ class Planner:
                 in the context of that table version (eg, if 'tbl' is a view, 'plan_target' might be the base)
         TODO: make exact_version_only a flag and use the versions from tbl
         """
+        if columns is None:
+            columns = []
         if exact_version_only is None:
             exact_version_only = []
         sql_elements = analyzer.sql_elements
@@ -958,6 +1003,7 @@ class Planner:
             analyzer.window_fn_calls
         )
         ctx = exec.ExecContext(row_builder)
+
         combined_ordering = cls._create_combined_ordering(analyzer, verify_agg=is_python_agg)
         cls._verify_join_clauses(analyzer)
 
@@ -966,6 +1012,7 @@ class Planner:
         # - join clause subexprs
         # - subexprs of Where clause conjuncts that can't be run in SQL
         # - all grouping exprs
+        # - all stratify exprs
         candidates = list(
             exprs.Expr.list_subexprs(
                 analyzer.select_list,
@@ -980,10 +1027,12 @@ class Planner:
             candidates.extend(
                 exprs.Expr.subexprs(analyzer.filter, filter=sql_elements.contains, traverse_matches=False)
             )
-        if analyzer.group_by_clause is not None:
-            candidates.extend(
-                exprs.Expr.list_subexprs(analyzer.group_by_clause, filter=sql_elements.contains, traverse_matches=False)
-            )
+        candidates.extend(
+            exprs.Expr.list_subexprs(analyzer.grouping_exprs, filter=sql_elements.contains, traverse_matches=False)
+        )
+        candidates.extend(
+            exprs.Expr.list_subexprs(analyzer.stratify_exprs, filter=sql_elements.contains, traverse_matches=False)
+        )
         # not isinstance(...): we don't want to materialize Literals via a Select
         sql_exprs = exprs.ExprSet(e for e in candidates if not isinstance(e, exprs.Literal))
 
@@ -1005,8 +1054,15 @@ class Planner:
                     traverse_matches=False,
                 )
             )
+
             plan = exec.SqlScanNode(
-                tbl, row_builder, select_list=tbl_scan_exprs, set_pk=with_pk, exact_version_only=exact_version_only
+                tbl,
+                row_builder,
+                select_list=tbl_scan_exprs,
+                columns=[c for c in columns if c.get_tbl().id == tbl.tbl_id],
+                set_pk=with_pk,
+                cell_md_col_refs=cls._cell_md_col_refs(tbl_scan_exprs),
+                exact_version_only=exact_version_only,
             )
             tbl_scan_plans.append(plan)
 
@@ -1028,7 +1084,17 @@ class Planner:
             # we need to order the input for window functions
             plan.set_order_by(analyzer.get_window_fn_ob_clause())
 
-        plan = cls._insert_prefetch_node(tbl.tbl_version.id, row_builder, plan)
+        if analyzer.sample_clause is not None:
+            plan = exec.SqlSampleNode(
+                row_builder,
+                input=plan,
+                select_list=tbl_scan_exprs,
+                sample_clause=analyzer.sample_clause,
+                stratify_exprs=analyzer.stratify_exprs,
+            )
+
+        plan = cls._add_prefetch_node(tbl.tbl_version.id, row_builder.unique_exprs, plan)
+        plan = cls._add_cell_reconstruction_node(analyzer.all_exprs, plan)
 
         if analyzer.group_by_clause is not None:
             # we're doing grouping aggregation; the input of the AggregateNode are the grouping exprs plus the
@@ -1050,26 +1116,12 @@ class Planner:
                 sql_elements.contains_all(analyzer.select_list)
                 and sql_elements.contains_all(analyzer.grouping_exprs)
                 and isinstance(plan, exec.SqlNode)
-                and plan.to_cte(keep_pk=(sample_clause is not None)) is not None
+                and plan.to_cte() is not None
             ):
-                if sample_clause is not None:
-                    plan = exec.SqlSampleNode(
-                        row_builder,
-                        input=plan,
-                        select_list=analyzer.select_list,
-                        stratify_exprs=analyzer.group_by_clause,
-                        sample_clause=sample_clause,
-                    )
-                else:
-                    plan = exec.SqlAggregationNode(
-                        row_builder,
-                        input=plan,
-                        select_list=analyzer.select_list,
-                        group_by_items=analyzer.group_by_clause,
-                    )
+                plan = exec.SqlAggregationNode(
+                    row_builder, input=plan, select_list=analyzer.select_list, group_by_items=analyzer.group_by_clause
+                )
             else:
-                if sample_clause is not None:
-                    raise excs.Error('Sample clause not supported with Python aggregation')
                 input_sql_node = plan.get_node(exec.SqlNode)
                 assert combined_ordering is not None
                 input_sql_node.set_order_by(combined_ordering)
@@ -1086,6 +1138,7 @@ class Planner:
                 if not agg_output.issuperset(exprs.ExprSet(eval_ctx.target_exprs)):
                     # we need an ExprEvalNode to evaluate the remaining output exprs
                     plan = exec.ExprEvalNode(row_builder, eval_ctx.target_exprs, agg_output, input=plan)
+                plan = cls._add_save_node(plan)
         else:
             if not exprs.ExprSet(sql_exprs).issuperset(exprs.ExprSet(eval_ctx.target_exprs)):
                 # we need an ExprEvalNode to evaluate the remaining output exprs
@@ -1119,26 +1172,24 @@ class Planner:
         return Analyzer(FromClause(tbls=[tbl]), [], where_clause=where_clause)
 
     @classmethod
-    def create_add_column_plan(
-        cls, tbl: catalog.TableVersionPath, col: catalog.Column
-    ) -> tuple[exec.ExecNode, Optional[int]]:
+    def create_add_column_plan(cls, tbl: catalog.TableVersionPath, col: catalog.Column) -> exec.ExecNode:
         """Creates a plan for InsertableTable.add_column()
         Returns:
             plan: the plan to execute
             value_expr slot idx for the plan output (for computed cols)
         """
         assert isinstance(tbl, catalog.TableVersionPath)
-        row_builder = exprs.RowBuilder(output_exprs=[], columns=[col], input_exprs=[])
+        row_builder = exprs.RowBuilder(output_exprs=[], columns=[col], input_exprs=[], tbl=tbl.tbl_version.get())
         analyzer = Analyzer(FromClause(tbls=[tbl]), row_builder.default_eval_ctx.target_exprs)
         plan = cls._create_query_plan(
             row_builder=row_builder, analyzer=analyzer, eval_ctx=row_builder.default_eval_ctx, with_pk=True
         )
+
         plan.ctx.batch_size = 16
         plan.ctx.show_pbar = True
         plan.ctx.ignore_errors = True
+        computed_exprs = row_builder.output_exprs - row_builder.input_exprs
+        plan.ctx.num_computed_exprs = len(computed_exprs)  # we are adding a computed column, so we need to evaluate it
+        plan = cls._add_save_node(plan)
 
-        # we want to flush images
-        if col.is_computed and col.is_stored and col.col_type.is_image_type():
-            plan.set_stored_img_cols(row_builder.output_slot_idxs())
-        value_expr_slot_idx = row_builder.output_slot_idxs()[0].slot_idx if col.is_computed else None
-        return plan, value_expr_slot_idx
+        return plan