sqlglot.optimizer.pushdown_projections
1from collections import defaultdict 2 3from sqlglot import alias, exp 4from sqlglot.optimizer.qualify_columns import Resolver 5from sqlglot.optimizer.scope import Scope, traverse_scope 6from sqlglot.schema import ensure_schema 7 8# Sentinel value that means an outer query selecting ALL columns 9SELECT_ALL = object() 10 11 12# Selection to use if selection list is empty 13def default_selection(is_agg: bool) -> exp.Alias: 14 return alias(exp.Max(this=exp.Literal.number(1)) if is_agg else "1", "_") 15 16 17def pushdown_projections(expression, schema=None, remove_unused_selections=True): 18 """ 19 Rewrite sqlglot AST to remove unused columns projections. 20 21 Example: 22 >>> import sqlglot 23 >>> sql = "SELECT y.a AS a FROM (SELECT x.a AS a, x.b AS b FROM x) AS y" 24 >>> expression = sqlglot.parse_one(sql) 25 >>> pushdown_projections(expression).sql() 26 'SELECT y.a AS a FROM (SELECT x.a AS a FROM x) AS y' 27 28 Args: 29 expression (sqlglot.Expression): expression to optimize 30 remove_unused_selections (bool): remove selects that are unused 31 Returns: 32 sqlglot.Expression: optimized expression 33 """ 34 # Map of Scope to all columns being selected by outer queries. 35 schema = ensure_schema(schema) 36 source_column_alias_count = {} 37 referenced_columns = defaultdict(set) 38 39 # We build the scope tree (which is traversed in DFS postorder), then iterate 40 # over the result in reverse order. This should ensure that the set of selected 41 # columns for a particular scope are completely build by the time we get to it. 42 for scope in reversed(traverse_scope(expression)): 43 parent_selections = referenced_columns.get(scope, {SELECT_ALL}) 44 alias_count = source_column_alias_count.get(scope, 0) 45 46 # We can't remove columns SELECT DISTINCT nor UNION DISTINCT. 47 if scope.expression.args.get("distinct"): 48 parent_selections = {SELECT_ALL} 49 50 if isinstance(scope.expression, exp.Union): 51 left, right = scope.union_scopes 52 referenced_columns[left] = parent_selections 53 54 if any(select.is_star for select in right.expression.selects): 55 referenced_columns[right] = parent_selections 56 elif not any(select.is_star for select in left.expression.selects): 57 referenced_columns[right] = [ 58 right.expression.selects[i].alias_or_name 59 for i, select in enumerate(left.expression.selects) 60 if SELECT_ALL in parent_selections or select.alias_or_name in parent_selections 61 ] 62 63 if isinstance(scope.expression, exp.Select): 64 if remove_unused_selections: 65 _remove_unused_selections(scope, parent_selections, schema, alias_count) 66 67 if scope.expression.is_star: 68 continue 69 70 # Group columns by source name 71 selects = defaultdict(set) 72 for col in scope.columns: 73 table_name = col.table 74 col_name = col.name 75 selects[table_name].add(col_name) 76 77 # Push the selected columns down to the next scope 78 for name, (node, source) in scope.selected_sources.items(): 79 if isinstance(source, Scope): 80 columns = {SELECT_ALL} if scope.pivots else selects.get(name) or set() 81 referenced_columns[source].update(columns) 82 83 column_aliases = node.alias_column_names 84 if column_aliases: 85 source_column_alias_count[source] = len(column_aliases) 86 87 return expression 88 89 90def _remove_unused_selections(scope, parent_selections, schema, alias_count): 91 order = scope.expression.args.get("order") 92 93 if order: 94 # Assume columns without a qualified table are references to output columns 95 order_refs = {c.name for c in order.find_all(exp.Column) if not c.table} 96 else: 97 order_refs = set() 98 99 new_selections = [] 100 removed = False 101 star = False 102 is_agg = False 103 104 select_all = SELECT_ALL in parent_selections 105 106 for selection in scope.expression.selects: 107 name = selection.alias_or_name 108 109 if select_all or name in parent_selections or name in order_refs or alias_count > 0: 110 new_selections.append(selection) 111 alias_count -= 1 112 else: 113 if selection.is_star: 114 star = True 115 removed = True 116 117 if not is_agg and selection.find(exp.AggFunc): 118 is_agg = True 119 120 if star: 121 resolver = Resolver(scope, schema) 122 names = {s.alias_or_name for s in new_selections} 123 124 for name in sorted(parent_selections): 125 if name not in names: 126 new_selections.append( 127 alias(exp.column(name, table=resolver.get_table(name)), name, copy=False) 128 ) 129 130 # If there are no remaining selections, just select a single constant 131 if not new_selections: 132 new_selections.append(default_selection(is_agg)) 133 134 scope.expression.select(*new_selections, append=False, copy=False) 135 136 if removed: 137 scope.clear_cache()
SELECT_ALL =
<object object>
def
pushdown_projections(expression, schema=None, remove_unused_selections=True):
18def pushdown_projections(expression, schema=None, remove_unused_selections=True): 19 """ 20 Rewrite sqlglot AST to remove unused columns projections. 21 22 Example: 23 >>> import sqlglot 24 >>> sql = "SELECT y.a AS a FROM (SELECT x.a AS a, x.b AS b FROM x) AS y" 25 >>> expression = sqlglot.parse_one(sql) 26 >>> pushdown_projections(expression).sql() 27 'SELECT y.a AS a FROM (SELECT x.a AS a FROM x) AS y' 28 29 Args: 30 expression (sqlglot.Expression): expression to optimize 31 remove_unused_selections (bool): remove selects that are unused 32 Returns: 33 sqlglot.Expression: optimized expression 34 """ 35 # Map of Scope to all columns being selected by outer queries. 36 schema = ensure_schema(schema) 37 source_column_alias_count = {} 38 referenced_columns = defaultdict(set) 39 40 # We build the scope tree (which is traversed in DFS postorder), then iterate 41 # over the result in reverse order. This should ensure that the set of selected 42 # columns for a particular scope are completely build by the time we get to it. 43 for scope in reversed(traverse_scope(expression)): 44 parent_selections = referenced_columns.get(scope, {SELECT_ALL}) 45 alias_count = source_column_alias_count.get(scope, 0) 46 47 # We can't remove columns SELECT DISTINCT nor UNION DISTINCT. 48 if scope.expression.args.get("distinct"): 49 parent_selections = {SELECT_ALL} 50 51 if isinstance(scope.expression, exp.Union): 52 left, right = scope.union_scopes 53 referenced_columns[left] = parent_selections 54 55 if any(select.is_star for select in right.expression.selects): 56 referenced_columns[right] = parent_selections 57 elif not any(select.is_star for select in left.expression.selects): 58 referenced_columns[right] = [ 59 right.expression.selects[i].alias_or_name 60 for i, select in enumerate(left.expression.selects) 61 if SELECT_ALL in parent_selections or select.alias_or_name in parent_selections 62 ] 63 64 if isinstance(scope.expression, exp.Select): 65 if remove_unused_selections: 66 _remove_unused_selections(scope, parent_selections, schema, alias_count) 67 68 if scope.expression.is_star: 69 continue 70 71 # Group columns by source name 72 selects = defaultdict(set) 73 for col in scope.columns: 74 table_name = col.table 75 col_name = col.name 76 selects[table_name].add(col_name) 77 78 # Push the selected columns down to the next scope 79 for name, (node, source) in scope.selected_sources.items(): 80 if isinstance(source, Scope): 81 columns = {SELECT_ALL} if scope.pivots else selects.get(name) or set() 82 referenced_columns[source].update(columns) 83 84 column_aliases = node.alias_column_names 85 if column_aliases: 86 source_column_alias_count[source] = len(column_aliases) 87 88 return expression
Rewrite sqlglot AST to remove unused columns projections.
Example:
>>> import sqlglot >>> sql = "SELECT y.a AS a FROM (SELECT x.a AS a, x.b AS b FROM x) AS y" >>> expression = sqlglot.parse_one(sql) >>> pushdown_projections(expression).sql() 'SELECT y.a AS a FROM (SELECT x.a AS a FROM x) AS y'
Arguments:
- expression (sqlglot.Expression): expression to optimize
- remove_unused_selections (bool): remove selects that are unused
Returns:
sqlglot.Expression: optimized expression