""" Create SQL statements for QuerySets. The code in here encapsulates all of the SQL construction so that QuerySets themselves do not have to (and could be backed by things other than SQL databases). The abstraction barrier only works one way: this module has to know all about the internals of models in order to get the information it needs. """ import copy import difflib import functools import sys from collections import Counter, namedtuple from collections.abc import Iterator, Mapping from itertools import chain, count, product from string import ascii_uppercase from django.core.exceptions import FieldDoesNotExist, FieldError from django.db import DEFAULT_DB_ALIAS, NotSupportedError, connections from django.db.models.aggregates import Count from django.db.models.constants import LOOKUP_SEP from django.db.models.expressions import ( BaseExpression, Col, Exists, F, OuterRef, Ref, ResolvedOuterRef, Value, ) from django.db.models.fields import Field from django.db.models.fields.related_lookups import MultiColSource from django.db.models.lookups import Lookup from django.db.models.query_utils import ( Q, check_rel_lookup_compatibility, refs_expression, ) from django.db.models.sql.constants import INNER, LOUTER, ORDER_DIR, SINGLE from django.db.models.sql.datastructures import BaseTable, Empty, Join, MultiJoin from django.db.models.sql.where import AND, OR, ExtraWhere, NothingNode, WhereNode from django.utils.functional import cached_property from django.utils.regex_helper import _lazy_re_compile from django.utils.tree import Node __all__ = ["Query", "RawQuery"] # Quotation marks ('"`[]), whitespace characters, semicolons, or inline # SQL comments are forbidden in column aliases. FORBIDDEN_ALIAS_PATTERN = _lazy_re_compile(r"['`\"\]\[;\s]|--|/\*|\*/") # Inspired from # https://www.postgresql.org/docs/current/sql-syntax-lexical.html#SQL-SYNTAX-IDENTIFIERS EXPLAIN_OPTIONS_PATTERN = _lazy_re_compile(r"[\w\-]+") def get_field_names_from_opts(opts): if opts is None: return set() return set( chain.from_iterable( (f.name, f.attname) if f.concrete else (f.name,) for f in opts.get_fields() ) ) def get_children_from_q(q): for child in q.children: if isinstance(child, Node): yield from get_children_from_q(child) else: yield child JoinInfo = namedtuple( "JoinInfo", ("final_field", "targets", "opts", "joins", "path", "transform_function"), ) class RawQuery: """A single raw SQL query.""" def __init__(self, sql, using, params=()): self.params = params self.sql = sql self.using = using self.cursor = None # Mirror some properties of a normal query so that # the compiler can be used to process results. self.low_mark, self.high_mark = 0, None # Used for offset/limit self.extra_select = {} self.annotation_select = {} def chain(self, using): return self.clone(using) def clone(self, using): return RawQuery(self.sql, using, params=self.params) def get_columns(self): if self.cursor is None: self._execute_query() converter = connections[self.using].introspection.identifier_converter return [converter(column_meta[0]) for column_meta in self.cursor.description] def __iter__(self): # Always execute a new query for a new iterator. # This could be optimized with a cache at the expense of RAM. self._execute_query() if not connections[self.using].features.can_use_chunked_reads: # If the database can't use chunked reads we need to make sure we # evaluate the entire query up front. result = list(self.cursor) else: result = self.cursor return iter(result) def __repr__(self): return "<%s: %s>" % (self.__class__.__name__, self) @property def params_type(self): if self.params is None: return None return dict if isinstance(self.params, Mapping) else tuple def __str__(self): if self.params_type is None: return self.sql return self.sql % self.params_type(self.params) def _execute_query(self): connection = connections[self.using] # Adapt parameters to the database, as much as possible considering # that the target type isn't known. See #17755. params_type = self.params_type adapter = connection.ops.adapt_unknown_value if params_type is tuple: params = tuple(adapter(val) for val in self.params) elif params_type is dict: params = {key: adapter(val) for key, val in self.params.items()} elif params_type is None: params = None else: raise RuntimeError("Unexpected params type: %s" % params_type) self.cursor = connection.cursor() self.cursor.execute(self.sql, params) ExplainInfo = namedtuple("ExplainInfo", ("format", "options")) class Query(BaseExpression): """A single SQL query.""" alias_prefix = "T" empty_result_set_value = None subq_aliases = frozenset([alias_prefix]) compiler = "SQLCompiler" base_table_class = BaseTable join_class = Join default_cols = True default_ordering = True standard_ordering = True filter_is_sticky = False subquery = False # SQL-related attributes. # Select and related select clauses are expressions to use in the SELECT # clause of the query. The select is used for cases where we want to set up # the select clause to contain other than default fields (values(), # subqueries...). Note that annotations go to annotations dictionary. select = () # The group_by attribute can have one of the following forms: # - None: no group by at all in the query # - A tuple of expressions: group by (at least) those expressions. # String refs are also allowed for now. # - True: group by all select fields of the model # See compiler.get_group_by() for details. group_by = None order_by = () low_mark = 0 # Used for offset/limit. high_mark = None # Used for offset/limit. distinct = False distinct_fields = () select_for_update = False select_for_update_nowait = False select_for_update_skip_locked = False select_for_update_of = () select_for_no_key_update = False select_related = False has_select_fields = False # Arbitrary limit for select_related to prevents infinite recursion. max_depth = 5 # Holds the selects defined by a call to values() or values_list() # excluding annotation_select and extra_select. values_select = () # SQL annotation-related attributes. annotation_select_mask = None _annotation_select_cache = None # Set combination attributes. combinator = None combinator_all = False combined_queries = () # These are for extensions. The contents are more or less appended verbatim # to the appropriate clause. extra_select_mask = None _extra_select_cache = None extra_tables = () extra_order_by = () # A tuple that is a set of model field names and either True, if these are # the fields to defer, or False if these are the only fields to load. deferred_loading = (frozenset(), True) explain_info = None def __init__(self, model, alias_cols=True): self.model = model self.alias_refcount = {} # alias_map is the most important data structure regarding joins. # It's used for recording which joins exist in the query and what # types they are. The key is the alias of the joined table (possibly # the table name) and the value is a Join-like object (see # sql.datastructures.Join for more information). self.alias_map = {} # Whether to provide alias to columns during reference resolving. self.alias_cols = alias_cols # Sometimes the query contains references to aliases in outer queries (as # a result of split_exclude). Correct alias quoting needs to know these # aliases too. # Map external tables to whether they are aliased. self.external_aliases = {} self.table_map = {} # Maps table names to list of aliases. self.used_aliases = set() self.where = WhereNode() # Maps alias -> Annotation Expression. self.annotations = {} # These are for extensions. The contents are more or less appended # verbatim to the appropriate clause. self.extra = {} # Maps col_alias -> (col_sql, params). self._filtered_relations = {} @property def output_field(self): if len(self.select) == 1: select = self.select[0] return getattr(select, "target", None) or select.field elif len(self.annotation_select) == 1: return next(iter(self.annotation_select.values())).output_field @cached_property def base_table(self): for alias in self.alias_map: return alias def __str__(self): """ Return the query as a string of SQL with the parameter values substituted in (use sql_with_params() to see the unsubstituted string). Parameter values won't necessarily be quoted correctly, since that is done by the database interface at execution time. """ sql, params = self.sql_with_params() return sql % params def sql_with_params(self): """ Return the query as an SQL string and the parameters that will be substituted into the query. """ return self.get_compiler(DEFAULT_DB_ALIAS).as_sql() def __deepcopy__(self, memo): """Limit the amount of work when a Query is deepcopied.""" result = self.clone() memo[id(self)] = result return result def get_compiler(self, using=None, connection=None, elide_empty=True): if using is None and connection is None: raise ValueError("Need either using or connection") if using: connection = connections[using] return connection.ops.compiler(self.compiler)( self, connection, using, elide_empty ) def get_meta(self): """ Return the Options instance (the model._meta) from which to start processing. Normally, this is self.model._meta, but it can be changed by subclasses. """ if self.model: return self.model._meta def clone(self): """ Return a copy of the current Query. A lightweight alternative to deepcopy(). """ obj = Empty() obj.__class__ = self.__class__ # Copy references to everything. obj.__dict__ = self.__dict__.copy() # Clone attributes that can't use shallow copy. obj.alias_refcount = self.alias_refcount.copy() obj.alias_map = self.alias_map.copy() obj.external_aliases = self.external_aliases.copy() obj.table_map = self.table_map.copy() obj.where = self.where.clone() obj.annotations = self.annotations.copy() if self.annotation_select_mask is not None: obj.annotation_select_mask = self.annotation_select_mask.copy() if self.combined_queries: obj.combined_queries = tuple( [query.clone() for query in self.combined_queries] ) # _annotation_select_cache cannot be copied, as doing so breaks the # (necessary) state in which both annotations and # _annotation_select_cache point to the same underlying objects. # It will get re-populated in the cloned queryset the next time it's # used. obj._annotation_select_cache = None obj.extra = self.extra.copy() if self.extra_select_mask is not None: obj.extra_select_mask = self.extra_select_mask.copy() if self._extra_select_cache is not None: obj._extra_select_cache = self._extra_select_cache.copy() if self.select_related is not False: # Use deepcopy because select_related stores fields in nested # dicts. obj.select_related = copy.deepcopy(obj.select_related) if "subq_aliases" in self.__dict__: obj.subq_aliases = self.subq_aliases.copy() obj.used_aliases = self.used_aliases.copy() obj._filtered_relations = self._filtered_relations.copy() # Clear the cached_property, if it exists. obj.__dict__.pop("base_table", None) return obj def chain(self, klass=None): """ Return a copy of the current Query that's ready for another operation. The klass argument changes the type of the Query, e.g. UpdateQuery. """ obj = self.clone() if klass and obj.__class__ != klass: obj.__class__ = klass if not obj.filter_is_sticky: obj.used_aliases = set() obj.filter_is_sticky = False if hasattr(obj, "_setup_query"): obj._setup_query() return obj def relabeled_clone(self, change_map): clone = self.clone() clone.change_aliases(change_map) return clone def _get_col(self, target, field, alias): if not self.alias_cols: alias = None return target.get_col(alias, field) def get_aggregation(self, using, aggregate_exprs): """ Return the dictionary with the values of the existing aggregations. """ if not aggregate_exprs: return {} # Store annotation mask prior to temporarily adding aggregations for # resolving purpose to facilitate their subsequent removal. refs_subquery = False replacements = {} annotation_select_mask = self.annotation_select_mask for alias, aggregate_expr in aggregate_exprs.items(): self.check_alias(alias) aggregate = aggregate_expr.resolve_expression( self, allow_joins=True, reuse=None, summarize=True ) if not aggregate.contains_aggregate: raise TypeError("%s is not an aggregate expression" % alias) # Temporarily add aggregate to annotations to allow remaining # members of `aggregates` to resolve against each others. self.append_annotation_mask([alias]) refs_subquery |= any( getattr(self.annotations[ref], "subquery", False) for ref in aggregate.get_refs() ) aggregate = aggregate.replace_expressions(replacements) self.annotations[alias] = aggregate replacements[Ref(alias, aggregate)] = aggregate # Stash resolved aggregates now that they have been allowed to resolve # against each other. aggregates = {alias: self.annotations.pop(alias) for alias in aggregate_exprs} self.set_annotation_mask(annotation_select_mask) # Existing usage of aggregation can be determined by the presence of # selected aggregates but also by filters against aliased aggregates. _, having, qualify = self.where.split_having_qualify() has_existing_aggregation = ( any( getattr(annotation, "contains_aggregate", True) for annotation in self.annotations.values() ) or having ) # Decide if we need to use a subquery. # # Existing aggregations would cause incorrect results as # get_aggregation() must produce just one result and thus must not use # GROUP BY. # # If the query has limit or distinct, or uses set operations, then # those operations must be done in a subquery so that the query # aggregates on the limit and/or distinct results instead of applying # the distinct and limit after the aggregation. if ( isinstance(self.group_by, tuple) or self.is_sliced or has_existing_aggregation or refs_subquery or qualify or self.distinct or self.combinator ): from django.db.models.sql.subqueries import AggregateQuery inner_query = self.clone() inner_query.subquery = True outer_query = AggregateQuery(self.model, inner_query) inner_query.select_for_update = False inner_query.select_related = False inner_query.set_annotation_mask(self.annotation_select) # Queries with distinct_fields need ordering and when a limit is # applied we must take the slice from the ordered query. Otherwise # no need for ordering. inner_query.clear_ordering(force=False) if not inner_query.distinct: # If the inner query uses default select and it has some # aggregate annotations, then we must make sure the inner # query is grouped by the main model's primary key. However, # clearing the select clause can alter results if distinct is # used. if inner_query.default_cols and has_existing_aggregation: inner_query.group_by = ( self.model._meta.pk.get_col(inner_query.get_initial_alias()), ) inner_query.default_cols = False if not qualify: # Mask existing annotations that are not referenced by # aggregates to be pushed to the outer query unless # filtering against window functions is involved as it # requires complex realising. annotation_mask = set() if isinstance(self.group_by, tuple): for expr in self.group_by: annotation_mask |= expr.get_refs() for aggregate in aggregates.values(): annotation_mask |= aggregate.get_refs() inner_query.set_annotation_mask(annotation_mask) # Add aggregates to the outer AggregateQuery. This requires making # sure all columns referenced by the aggregates are selected in the # inner query. It is achieved by retrieving all column references # by the aggregates, explicitly selecting them in the inner query, # and making sure the aggregates are repointed to them. col_refs = {} for alias, aggregate in aggregates.items(): replacements = {} for col in self._gen_cols([aggregate], resolve_refs=False): if not (col_ref := col_refs.get(col)): index = len(col_refs) + 1 col_alias = f"__col{index}" col_ref = Ref(col_alias, col) col_refs[col] = col_ref inner_query.annotations[col_alias] = col inner_query.append_annotation_mask([col_alias]) replacements[col] = col_ref outer_query.annotations[alias] = aggregate.replace_expressions( replacements ) if ( inner_query.select == () and not inner_query.default_cols and not inner_query.annotation_select_mask ): # In case of Model.objects[0:3].count(), there would be no # field selected in the inner query, yet we must use a subquery. # So, make sure at least one field is selected. inner_query.select = ( self.model._meta.pk.get_col(inner_query.get_initial_alias()), ) else: outer_query = self self.select = () self.default_cols = False self.extra = {} if self.annotations: # Inline reference to existing annotations and mask them as # they are unnecessary given only the summarized aggregations # are requested. replacements = { Ref(alias, annotation): annotation for alias, annotation in self.annotations.items() } self.annotations = { alias: aggregate.replace_expressions(replacements) for alias, aggregate in aggregates.items() } else: self.annotations = aggregates self.set_annotation_mask(aggregates) empty_set_result = [ expression.empty_result_set_value for expression in outer_query.annotation_select.values() ] elide_empty = not any(result is NotImplemented for result in empty_set_result) outer_query.clear_ordering(force=True) outer_query.clear_limits() outer_query.select_for_update = False outer_query.select_related = False compiler = outer_query.get_compiler(using, elide_empty=elide_empty) result = compiler.execute_sql(SINGLE) if result is None: result = empty_set_result else: converters = compiler.get_converters(outer_query.annotation_select.values()) result = next(compiler.apply_converters((result,), converters)) return dict(zip(outer_query.annotation_select, result)) def get_count(self, using): """ Perform a COUNT() query using the current filter constraints. """ obj = self.clone() return obj.get_aggregation(using, {"__count": Count("*")})["__count"] def has_filters(self): return self.where def exists(self, limit=True): q = self.clone() if not (q.distinct and q.is_sliced): if q.group_by is True: q.add_fields( (f.attname for f in self.model._meta.concrete_fields), False ) # Disable GROUP BY aliases to avoid orphaning references to the # SELECT clause which is about to be cleared. q.set_group_by(allow_aliases=False) q.clear_select_clause() if q.combined_queries and q.combinator == "union": q.combined_queries = tuple( combined_query.exists(limit=False) for combined_query in q.combined_queries ) q.clear_ordering(force=True) if limit: q.set_limits(high=1) q.add_annotation(Value(1), "a") return q def has_results(self, using): q = self.exists(using) compiler = q.get_compiler(using=using) return compiler.has_results() def explain(self, using, format=None, **options): q = self.clone() for option_name in options: if ( not EXPLAIN_OPTIONS_PATTERN.fullmatch(option_name) or "--" in option_name ): raise ValueError(f"Invalid option name: {option_name!r}.") q.explain_info = ExplainInfo(format, options) compiler = q.get_compiler(using=using) return "\n".join(compiler.explain_query()) def combine(self, rhs, connector): """ Merge the 'rhs' query into the current one (with any 'rhs' effects being applied *after* (that is, "to the right of") anything in the current query. 'rhs' is not modified during a call to this function. The 'connector' parameter describes how to connect filters from the 'rhs' query. """ if self.model != rhs.model: raise TypeError("Cannot combine queries on two different base models.") if self.is_sliced: raise TypeError("Cannot combine queries once a slice has been taken.") if self.distinct != rhs.distinct: raise TypeError("Cannot combine a unique query with a non-unique query.") if self.distinct_fields != rhs.distinct_fields: raise TypeError("Cannot combine queries with different distinct fields.") # If lhs and rhs shares the same alias prefix, it is possible to have # conflicting alias changes like T4 -> T5, T5 -> T6, which might end up # as T4 -> T6 while combining two querysets. To prevent this, change an # alias prefix of the rhs and update current aliases accordingly, # except if the alias is the base table since it must be present in the # query on both sides. initial_alias = self.get_initial_alias() rhs.bump_prefix(self, exclude={initial_alias}) # Work out how to relabel the rhs aliases, if necessary. change_map = {} conjunction = connector == AND # Determine which existing joins can be reused. When combining the # query with AND we must recreate all joins for m2m filters. When # combining with OR we can reuse joins. The reason is that in AND # case a single row can't fulfill a condition like: # revrel__col=1 & revrel__col=2 # But, there might be two different related rows matching this # condition. In OR case a single True is enough, so single row is # enough, too. # # Note that we will be creating duplicate joins for non-m2m joins in # the AND case. The results will be correct but this creates too many # joins. This is something that could be fixed later on. reuse = set() if conjunction else set(self.alias_map) joinpromoter = JoinPromoter(connector, 2, False) joinpromoter.add_votes( j for j in self.alias_map if self.alias_map[j].join_type == INNER ) rhs_votes = set() # Now, add the joins from rhs query into the new query (skipping base # table). rhs_tables = list(rhs.alias_map)[1:] for alias in rhs_tables: join = rhs.alias_map[alias] # If the left side of the join was already relabeled, use the # updated alias. join = join.relabeled_clone(change_map) new_alias = self.join(join, reuse=reuse) if join.join_type == INNER: rhs_votes.add(new_alias) # We can't reuse the same join again in the query. If we have two # distinct joins for the same connection in rhs query, then the # combined query must have two joins, too. reuse.discard(new_alias) if alias != new_alias: change_map[alias] = new_alias if not rhs.alias_refcount[alias]: # The alias was unused in the rhs query. Unref it so that it # will be unused in the new query, too. We have to add and # unref the alias so that join promotion has information of # the join type for the unused alias. self.unref_alias(new_alias) joinpromoter.add_votes(rhs_votes) joinpromoter.update_join_types(self) # Combine subqueries aliases to ensure aliases relabelling properly # handle subqueries when combining where and select clauses. self.subq_aliases |= rhs.subq_aliases # Now relabel a copy of the rhs where-clause and add it to the current # one. w = rhs.where.clone() w.relabel_aliases(change_map) self.where.add(w, connector) # Selection columns and extra extensions are those provided by 'rhs'. if rhs.select: self.set_select([col.relabeled_clone(change_map) for col in rhs.select]) else: self.select = () if connector == OR: # It would be nice to be able to handle this, but the queries don't # really make sense (or return consistent value sets). Not worth # the extra complexity when you can write a real query instead. if self.extra and rhs.extra: raise ValueError( "When merging querysets using 'or', you cannot have " "extra(select=...) on both sides." ) self.extra.update(rhs.extra) extra_select_mask = set() if self.extra_select_mask is not None: extra_select_mask.update(self.extra_select_mask) if rhs.extra_select_mask is not None: extra_select_mask.update(rhs.extra_select_mask) if extra_select_mask: self.set_extra_mask(extra_select_mask) self.extra_tables += rhs.extra_tables # Ordering uses the 'rhs' ordering, unless it has none, in which case # the current ordering is used. self.order_by = rhs.order_by or self.order_by self.extra_order_by = rhs.extra_order_by or self.extra_order_by def _get_defer_select_mask(self, opts, mask, select_mask=None): if select_mask is None: select_mask = {} select_mask[opts.pk] = {} # All concrete fields that are not part of the defer mask must be # loaded. If a relational field is encountered it gets added to the # mask for it be considered if `select_related` and the cycle continues # by recursively caling this function. for field in opts.concrete_fields: field_mask = mask.pop(field.name, None) field_att_mask = mask.pop(field.attname, None) if field_mask is None and field_att_mask is None: select_mask.setdefault(field, {}) elif field_mask: if not field.is_relation: raise FieldError(next(iter(field_mask))) field_select_mask = select_mask.setdefault(field, {}) related_model = field.remote_field.model._meta.concrete_model self._get_defer_select_mask( related_model._meta, field_mask, field_select_mask ) # Remaining defer entries must be references to reverse relationships. # The following code is expected to raise FieldError if it encounters # a malformed defer entry. for field_name, field_mask in mask.items(): if filtered_relation := self._filtered_relations.get(field_name): relation = opts.get_field(filtered_relation.relation_name) field_select_mask = select_mask.setdefault((field_name, relation), {}) field = relation.field else: reverse_rel = opts.get_field(field_name) # While virtual fields such as many-to-many and generic foreign # keys cannot be effectively deferred we've historically # allowed them to be passed to QuerySet.defer(). Ignore such # field references until a layer of validation at mask # alteration time will be implemented eventually. if not hasattr(reverse_rel, "field"): continue field = reverse_rel.field field_select_mask = select_mask.setdefault(field, {}) related_model = field.model._meta.concrete_model self._get_defer_select_mask( related_model._meta, field_mask, field_select_mask ) return select_mask def _get_only_select_mask(self, opts, mask, select_mask=None): if select_mask is None: select_mask = {} select_mask[opts.pk] = {} # Only include fields mentioned in the mask. for field_name, field_mask in mask.items(): field = opts.get_field(field_name) # Retrieve the actual field associated with reverse relationships # as that's what is expected in the select mask. if field in opts.related_objects: field_key = field.field else: field_key = field field_select_mask = select_mask.setdefault(field_key, {}) if field_mask: if not field.is_relation: raise FieldError(next(iter(field_mask))) related_model = field.remote_field.model._meta.concrete_model self._get_only_select_mask( related_model._meta, field_mask, field_select_mask ) return select_mask def get_select_mask(self): """ Convert the self.deferred_loading data structure to an alternate data structure, describing the field that *will* be loaded. This is used to compute the columns to select from the database and also by the QuerySet class to work out which fields are being initialized on each model. Models that have all their fields included aren't mentioned in the result, only those that have field restrictions in place. """ field_names, defer = self.deferred_loading if not field_names: return {} mask = {} for field_name in field_names: part_mask = mask for part in field_name.split(LOOKUP_SEP): part_mask = part_mask.setdefault(part, {}) opts = self.get_meta() if defer: return self._get_defer_select_mask(opts, mask) return self._get_only_select_mask(opts, mask) def table_alias(self, table_name, create=False, filtered_relation=None): """ Return a table alias for the given table_name and whether this is a new alias or not. If 'create' is true, a new alias is always created. Otherwise, the most recently created alias for the table (if one exists) is reused. """ alias_list = self.table_map.get(table_name) if not create and alias_list: alias = alias_list[0] self.alias_refcount[alias] += 1 return alias, False # Create a new alias for this table. if alias_list: alias = "%s%d" % (self.alias_prefix, len(self.alias_map) + 1) alias_list.append(alias) else: # The first occurrence of a table uses the table name directly. alias = ( filtered_relation.alias if filtered_relation is not None else table_name ) self.table_map[table_name] = [alias] self.alias_refcount[alias] = 1 return alias, True def ref_alias(self, alias): """Increases the reference count for this alias.""" self.alias_refcount[alias] += 1 def unref_alias(self, alias, amount=1): """Decreases the reference count for this alias.""" self.alias_refcount[alias] -= amount def promote_joins(self, aliases): """ Promote recursively the join type of given aliases and its children to an outer join. If 'unconditional' is False, only promote the join if it is nullable or the parent join is an outer join. The children promotion is done to avoid join chains that contain a LOUTER b INNER c. So, if we have currently a INNER b INNER c and a->b is promoted, then we must also promote b->c automatically, or otherwise the promotion of a->b doesn't actually change anything in the query results. """ aliases = list(aliases) while aliases: alias = aliases.pop(0) if self.alias_map[alias].join_type is None: # This is the base table (first FROM entry) - this table # isn't really joined at all in the query, so we should not # alter its join type. continue # Only the first alias (skipped above) should have None join_type assert self.alias_map[alias].join_type is not None parent_alias = self.alias_map[alias].parent_alias parent_louter = ( parent_alias and self.alias_map[parent_alias].join_type == LOUTER ) already_louter = self.alias_map[alias].join_type == LOUTER if (self.alias_map[alias].nullable or parent_louter) and not already_louter: self.alias_map[alias] = self.alias_map[alias].promote() # Join type of 'alias' changed, so re-examine all aliases that # refer to this one. aliases.extend( join for join in self.alias_map if self.alias_map[join].parent_alias == alias and join not in aliases ) def demote_joins(self, aliases): """ Change join type from LOUTER to INNER for all joins in aliases. Similarly to promote_joins(), this method must ensure no join chains containing first an outer, then an inner join are generated. If we are demoting b->c join in chain a LOUTER b LOUTER c then we must demote a->b automatically, or otherwise the demotion of b->c doesn't actually change anything in the query results. . """ aliases = list(aliases) while aliases: alias = aliases.pop(0) if self.alias_map[alias].join_type == LOUTER: self.alias_map[alias] = self.alias_map[alias].demote() parent_alias = self.alias_map[alias].parent_alias if self.alias_map[parent_alias].join_type == INNER: aliases.append(parent_alias) def reset_refcounts(self, to_counts): """ Reset reference counts for aliases so that they match the value passed in `to_counts`. """ for alias, cur_refcount in self.alias_refcount.copy().items(): unref_amount = cur_refcount - to_counts.get(alias, 0) self.unref_alias(alias, unref_amount) def change_aliases(self, change_map): """ Change the aliases in change_map (which maps old-alias -> new-alias), relabelling any references to them in select columns and the where clause. """ # If keys and values of change_map were to intersect, an alias might be # updated twice (e.g. T4 -> T5, T5 -> T6, so also T4 -> T6) depending # on their order in change_map. assert set(change_map).isdisjoint(change_map.values()) # 1. Update references in "select" (normal columns plus aliases), # "group by" and "where". self.where.relabel_aliases(change_map) if isinstance(self.group_by, tuple): self.group_by = tuple( [col.relabeled_clone(change_map) for col in self.group_by] ) self.select = tuple([col.relabeled_clone(change_map) for col in self.select]) self.annotations = self.annotations and { key: col.relabeled_clone(change_map) for key, col in self.annotations.items() } # 2. Rename the alias in the internal table/alias datastructures. for old_alias, new_alias in change_map.items(): if old_alias not in self.alias_map: continue alias_data = self.alias_map[old_alias].relabeled_clone(change_map) self.alias_map[new_alias] = alias_data self.alias_refcount[new_alias] = self.alias_refcount[old_alias] del self.alias_refcount[old_alias] del self.alias_map[old_alias] table_aliases = self.table_map[alias_data.table_name] for pos, alias in enumerate(table_aliases): if alias == old_alias: table_aliases[pos] = new_alias break self.external_aliases = { # Table is aliased or it's being changed and thus is aliased. change_map.get(alias, alias): (aliased or alias in change_map) for alias, aliased in self.external_aliases.items() } def bump_prefix(self, other_query, exclude=None): """ Change the alias prefix to the next letter in the alphabet in a way that the other query's aliases and this query's aliases will not conflict. Even tables that previously had no alias will get an alias after this call. To prevent changing aliases use the exclude parameter. """ def prefix_gen(): """ Generate a sequence of characters in alphabetical order: -> 'A', 'B', 'C', ... When the alphabet is finished, the sequence will continue with the Cartesian product: -> 'AA', 'AB', 'AC', ... """ alphabet = ascii_uppercase prefix = chr(ord(self.alias_prefix) + 1) yield prefix for n in count(1): seq = alphabet[alphabet.index(prefix) :] if prefix else alphabet for s in product(seq, repeat=n): yield "".join(s) prefix = None if self.alias_prefix != other_query.alias_prefix: # No clashes between self and outer query should be possible. return # Explicitly avoid infinite loop. The constant divider is based on how # much depth recursive subquery references add to the stack. This value # might need to be adjusted when adding or removing function calls from # the code path in charge of performing these operations. local_recursion_limit = sys.getrecursionlimit() // 16 for pos, prefix in enumerate(prefix_gen()): if prefix not in self.subq_aliases: self.alias_prefix = prefix break if pos > local_recursion_limit: raise RecursionError( "Maximum recursion depth exceeded: too many subqueries." ) self.subq_aliases = self.subq_aliases.union([self.alias_prefix]) other_query.subq_aliases = other_query.subq_aliases.union(self.subq_aliases) if exclude is None: exclude = {} self.change_aliases( { alias: "%s%d" % (self.alias_prefix, pos) for pos, alias in enumerate(self.alias_map) if alias not in exclude } ) def get_initial_alias(self): """ Return the first alias for this query, after increasing its reference count. """ if self.alias_map: alias = self.base_table self.ref_alias(alias) elif self.model: alias = self.join(self.base_table_class(self.get_meta().db_table, None)) else: alias = None return alias def count_active_tables(self): """ Return the number of tables in this query with a non-zero reference count. After execution, the reference counts are zeroed, so tables added in compiler will not be seen by this method. """ return len([1 for count in self.alias_refcount.values() if count]) def join(self, join, reuse=None, reuse_with_filtered_relation=False): """ Return an alias for the 'join', either reusing an existing alias for that join or creating a new one. 'join' is either a base_table_class or join_class. The 'reuse' parameter can be either None which means all joins are reusable, or it can be a set containing the aliases that can be reused. The 'reuse_with_filtered_relation' parameter is used when computing FilteredRelation instances. A join is always created as LOUTER if the lhs alias is LOUTER to make sure chains like t1 LOUTER t2 INNER t3 aren't generated. All new joins are created as LOUTER if the join is nullable. """ if reuse_with_filtered_relation and reuse: reuse_aliases = [ a for a, j in self.alias_map.items() if a in reuse and j.equals(join) ] else: reuse_aliases = [ a for a, j in self.alias_map.items() if (reuse is None or a in reuse) and j == join ] if reuse_aliases: if join.table_alias in reuse_aliases: reuse_alias = join.table_alias else: # Reuse the most recent alias of the joined table # (a many-to-many relation may be joined multiple times). reuse_alias = reuse_aliases[-1] self.ref_alias(reuse_alias) return reuse_alias # No reuse is possible, so we need a new alias. alias, _ = self.table_alias( join.table_name, create=True, filtered_relation=join.filtered_relation ) if join.join_type: if self.alias_map[join.parent_alias].join_type == LOUTER or join.nullable: join_type = LOUTER else: join_type = INNER join.join_type = join_type join.table_alias = alias self.alias_map[alias] = join return alias def join_parent_model(self, opts, model, alias, seen): """ Make sure the given 'model' is joined in the query. If 'model' isn't a parent of 'opts' or if it is None this method is a no-op. The 'alias' is the root alias for starting the join, 'seen' is a dict of model -> alias of existing joins. It must also contain a mapping of None -> some alias. This will be returned in the no-op case. """ if model in seen: return seen[model] chain = opts.get_base_chain(model) if not chain: return alias curr_opts = opts for int_model in chain: if int_model in seen: curr_opts = int_model._meta alias = seen[int_model] continue # Proxy model have elements in base chain # with no parents, assign the new options # object and skip to the next base in that # case if not curr_opts.parents[int_model]: curr_opts = int_model._meta continue link_field = curr_opts.get_ancestor_link(int_model) join_info = self.setup_joins([link_field.name], curr_opts, alias) curr_opts = int_model._meta alias = seen[int_model] = join_info.joins[-1] return alias or seen[None] def check_alias(self, alias): if FORBIDDEN_ALIAS_PATTERN.search(alias): raise ValueError( "Column aliases cannot contain whitespace characters, quotation marks, " "semicolons, or SQL comments." ) def add_annotation(self, annotation, alias, select=True): """Add a single annotation expression to the Query.""" self.check_alias(alias) annotation = annotation.resolve_expression(self, allow_joins=True, reuse=None) if select: self.append_annotation_mask([alias]) else: self.set_annotation_mask(set(self.annotation_select).difference({alias})) self.annotations[alias] = annotation def resolve_expression(self, query, *args, **kwargs): clone = self.clone() # Subqueries need to use a different set of aliases than the outer query. clone.bump_prefix(query) clone.subquery = True clone.where.resolve_expression(query, *args, **kwargs) # Resolve combined queries. if clone.combinator: clone.combined_queries = tuple( [ combined_query.resolve_expression(query, *args, **kwargs) for combined_query in clone.combined_queries ] ) for key, value in clone.annotations.items(): resolved = value.resolve_expression(query, *args, **kwargs) if hasattr(resolved, "external_aliases"): resolved.external_aliases.update(clone.external_aliases) clone.annotations[key] = resolved # Outer query's aliases are considered external. for alias, table in query.alias_map.items(): clone.external_aliases[alias] = ( isinstance(table, Join) and table.join_field.related_model._meta.db_table != alias ) or ( isinstance(table, BaseTable) and table.table_name != table.table_alias ) return clone def get_external_cols(self): exprs = chain(self.annotations.values(), self.where.children) return [ col for col in self._gen_cols(exprs, include_external=True) if col.alias in self.external_aliases ] def get_group_by_cols(self, wrapper=None): # If wrapper is referenced by an alias for an explicit GROUP BY through # values() a reference to this expression and not the self must be # returned to ensure external column references are not grouped against # as well. external_cols = self.get_external_cols() if any(col.possibly_multivalued for col in external_cols): return [wrapper or self] return external_cols def as_sql(self, compiler, connection): # Some backends (e.g. Oracle) raise an error when a subquery contains # unnecessary ORDER BY clause. if ( self.subquery and not connection.features.ignores_unnecessary_order_by_in_subqueries ): self.clear_ordering(force=False) for query in self.combined_queries: query.clear_ordering(force=False) sql, params = self.get_compiler(connection=connection).as_sql() if self.subquery: sql = "(%s)" % sql return sql, params def resolve_lookup_value(self, value, can_reuse, allow_joins): if hasattr(value, "resolve_expression"): value = value.resolve_expression( self, reuse=can_reuse, allow_joins=allow_joins, ) elif isinstance(value, (list, tuple)): # The items of the iterable may be expressions and therefore need # to be resolved independently. values = ( self.resolve_lookup_value(sub_value, can_reuse, allow_joins) for sub_value in value ) type_ = type(value) if hasattr(type_, "_make"): # namedtuple return type_(*values) return type_(values) return value def solve_lookup_type(self, lookup, summarize=False): """ Solve the lookup type from the lookup (e.g.: 'foobar__id__icontains'). """ lookup_splitted = lookup.split(LOOKUP_SEP) if self.annotations: annotation, expression_lookups = refs_expression( lookup_splitted, self.annotations ) if annotation: expression = self.annotations[annotation] if summarize: expression = Ref(annotation, expression) return expression_lookups, (), expression _, field, _, lookup_parts = self.names_to_path(lookup_splitted, self.get_meta()) field_parts = lookup_splitted[0 : len(lookup_splitted) - len(lookup_parts)] if len(lookup_parts) > 1 and not field_parts: raise FieldError( 'Invalid lookup "%s" for model %s".' % (lookup, self.get_meta().model.__name__) ) return lookup_parts, field_parts, False def check_query_object_type(self, value, opts, field): """ Check whether the object passed while querying is of the correct type. If not, raise a ValueError specifying the wrong object. """ if hasattr(value, "_meta"): if not check_rel_lookup_compatibility(value._meta.model, opts, field): raise ValueError( 'Cannot query "%s": Must be "%s" instance.' % (value, opts.object_name) ) def check_related_objects(self, field, value, opts): """Check the type of object passed to query relations.""" if field.is_relation: # Check that the field and the queryset use the same model in a # query like .filter(author=Author.objects.all()). For example, the # opts would be Author's (from the author field) and value.model # would be Author.objects.all() queryset's .model (Author also). # The field is the related field on the lhs side. if ( isinstance(value, Query) and not value.has_select_fields and not check_rel_lookup_compatibility(value.model, opts, field) ): raise ValueError( 'Cannot use QuerySet for "%s": Use a QuerySet for "%s".' % (value.model._meta.object_name, opts.object_name) ) elif hasattr(value, "_meta"): self.check_query_object_type(value, opts, field) elif hasattr(value, "__iter__"): for v in value: self.check_query_object_type(v, opts, field) def check_filterable(self, expression): """Raise an error if expression cannot be used in a WHERE clause.""" if hasattr(expression, "resolve_expression") and not getattr( expression, "filterable", True ): raise NotSupportedError( expression.__class__.__name__ + " is disallowed in the filter " "clause." ) if hasattr(expression, "get_source_expressions"): for expr in expression.get_source_expressions(): self.check_filterable(expr) def build_lookup(self, lookups, lhs, rhs): """ Try to extract transforms and lookup from given lhs. The lhs value is something that works like SQLExpression. The rhs value is what the lookup is going to compare against. The lookups is a list of names to extract using get_lookup() and get_transform(). """ # __exact is the default lookup if one isn't given. *transforms, lookup_name = lookups or ["exact"] for name in transforms: lhs = self.try_transform(lhs, name) # First try get_lookup() so that the lookup takes precedence if the lhs # supports both transform and lookup for the name. lookup_class = lhs.get_lookup(lookup_name) if not lookup_class: # A lookup wasn't found. Try to interpret the name as a transform # and do an Exact lookup against it. lhs = self.try_transform(lhs, lookup_name) lookup_name = "exact" lookup_class = lhs.get_lookup(lookup_name) if not lookup_class: return lookup = lookup_class(lhs, rhs) # Interpret '__exact=None' as the sql 'is NULL'; otherwise, reject all # uses of None as a query value unless the lookup supports it. if lookup.rhs is None and not lookup.can_use_none_as_rhs: if lookup_name not in ("exact", "iexact"): raise ValueError("Cannot use None as a query value") return lhs.get_lookup("isnull")(lhs, True) # For Oracle '' is equivalent to null. The check must be done at this # stage because join promotion can't be done in the compiler. Using # DEFAULT_DB_ALIAS isn't nice but it's the best that can be done here. # A similar thing is done in is_nullable(), too. if ( lookup_name == "exact" and lookup.rhs == "" and connections[DEFAULT_DB_ALIAS].features.interprets_empty_strings_as_nulls ): return lhs.get_lookup("isnull")(lhs, True) return lookup def try_transform(self, lhs, name): """ Helper method for build_lookup(). Try to fetch and initialize a transform for name parameter from lhs. """ transform_class = lhs.get_transform(name) if transform_class: return transform_class(lhs) else: output_field = lhs.output_field.__class__ suggested_lookups = difflib.get_close_matches( name, lhs.output_field.get_lookups() ) if suggested_lookups: suggestion = ", perhaps you meant %s?" % " or ".join(suggested_lookups) else: suggestion = "." raise FieldError( "Unsupported lookup '%s' for %s or join on the field not " "permitted%s" % (name, output_field.__name__, suggestion) ) def build_filter( self, filter_expr, branch_negated=False, current_negated=False, can_reuse=None, allow_joins=True, split_subq=True, reuse_with_filtered_relation=False, check_filterable=True, summarize=False, ): """ Build a WhereNode for a single filter clause but don't add it to this Query. Query.add_q() will then add this filter to the where Node. The 'branch_negated' tells us if the current branch contains any negations. This will be used to determine if subqueries are needed. The 'current_negated' is used to determine if the current filter is negated or not and this will be used to determine if IS NULL filtering is needed. The difference between current_negated and branch_negated is that branch_negated is set on first negation, but current_negated is flipped for each negation. Note that add_filter will not do any negating itself, that is done upper in the code by add_q(). The 'can_reuse' is a set of reusable joins for multijoins. If 'reuse_with_filtered_relation' is True, then only joins in can_reuse will be reused. The method will create a filter clause that can be added to the current query. However, if the filter isn't added to the query then the caller is responsible for unreffing the joins used. """ if isinstance(filter_expr, dict): raise FieldError("Cannot parse keyword query as dict") if isinstance(filter_expr, Q): return self._add_q( filter_expr, branch_negated=branch_negated, current_negated=current_negated, used_aliases=can_reuse, allow_joins=allow_joins, split_subq=split_subq, check_filterable=check_filterable, summarize=summarize, ) if hasattr(filter_expr, "resolve_expression"): if not getattr(filter_expr, "conditional", False): raise TypeError("Cannot filter against a non-conditional expression.") condition = filter_expr.resolve_expression( self, allow_joins=allow_joins, summarize=summarize ) if not isinstance(condition, Lookup): condition = self.build_lookup(["exact"], condition, True) return WhereNode([condition], connector=AND), [] arg, value = filter_expr if not arg: raise FieldError("Cannot parse keyword query %r" % arg) lookups, parts, reffed_expression = self.solve_lookup_type(arg, summarize) if check_filterable: self.check_filterable(reffed_expression) if not allow_joins and len(parts) > 1: raise FieldError("Joined field references are not permitted in this query") pre_joins = self.alias_refcount.copy() value = self.resolve_lookup_value(value, can_reuse, allow_joins) used_joins = { k for k, v in self.alias_refcount.items() if v > pre_joins.get(k, 0) } if check_filterable: self.check_filterable(value) if reffed_expression: condition = self.build_lookup(lookups, reffed_expression, value) return WhereNode([condition], connector=AND), [] opts = self.get_meta() alias = self.get_initial_alias() allow_many = not branch_negated or not split_subq try: join_info = self.setup_joins( parts, opts, alias, can_reuse=can_reuse, allow_many=allow_many, reuse_with_filtered_relation=reuse_with_filtered_relation, ) # Prevent iterator from being consumed by check_related_objects() if isinstance(value, Iterator): value = list(value) self.check_related_objects(join_info.final_field, value, join_info.opts) # split_exclude() needs to know which joins were generated for the # lookup parts self._lookup_joins = join_info.joins except MultiJoin as e: return self.split_exclude(filter_expr, can_reuse, e.names_with_path) # Update used_joins before trimming since they are reused to determine # which joins could be later promoted to INNER. used_joins.update(join_info.joins) targets, alias, join_list = self.trim_joins( join_info.targets, join_info.joins, join_info.path ) if can_reuse is not None: can_reuse.update(join_list) if join_info.final_field.is_relation: if len(targets) == 1: col = self._get_col(targets[0], join_info.final_field, alias) else: col = MultiColSource( alias, targets, join_info.targets, join_info.final_field ) else: col = self._get_col(targets[0], join_info.final_field, alias) condition = self.build_lookup(lookups, col, value) lookup_type = condition.lookup_name clause = WhereNode([condition], connector=AND) require_outer = ( lookup_type == "isnull" and condition.rhs is True and not current_negated ) if ( current_negated and (lookup_type != "isnull" or condition.rhs is False) and condition.rhs is not None ): require_outer = True if lookup_type != "isnull": # The condition added here will be SQL like this: # NOT (col IS NOT NULL), where the first NOT is added in # upper layers of code. The reason for addition is that if col # is null, then col != someval will result in SQL "unknown" # which isn't the same as in Python. The Python None handling # is wanted, and it can be gotten by # (col IS NULL OR col != someval) # <=> # NOT (col IS NOT NULL AND col = someval). if ( self.is_nullable(targets[0]) or self.alias_map[join_list[-1]].join_type == LOUTER ): lookup_class = targets[0].get_lookup("isnull") col = self._get_col(targets[0], join_info.targets[0], alias) clause.add(lookup_class(col, False), AND) # If someval is a nullable column, someval IS NOT NULL is # added. if isinstance(value, Col) and self.is_nullable(value.target): lookup_class = value.target.get_lookup("isnull") clause.add(lookup_class(value, False), AND) return clause, used_joins if not require_outer else () def add_filter(self, filter_lhs, filter_rhs): self.add_q(Q((filter_lhs, filter_rhs))) def add_q(self, q_object): """ A preprocessor for the internal _add_q(). Responsible for doing final join promotion. """ # For join promotion this case is doing an AND for the added q_object # and existing conditions. So, any existing inner join forces the join # type to remain inner. Existing outer joins can however be demoted. # (Consider case where rel_a is LOUTER and rel_a__col=1 is added - if # rel_a doesn't produce any rows, then the whole condition must fail. # So, demotion is OK. existing_inner = { a for a in self.alias_map if self.alias_map[a].join_type == INNER } clause, _ = self._add_q(q_object, self.used_aliases) if clause: self.where.add(clause, AND) self.demote_joins(existing_inner) def build_where(self, filter_expr): return self.build_filter(filter_expr, allow_joins=False)[0] def clear_where(self): self.where = WhereNode() def _add_q( self, q_object, used_aliases, branch_negated=False, current_negated=False, allow_joins=True, split_subq=True, check_filterable=True, summarize=False, ): """Add a Q-object to the current filter.""" connector = q_object.connector current_negated ^= q_object.negated branch_negated = branch_negated or q_object.negated target_clause = WhereNode(connector=connector, negated=q_object.negated) joinpromoter = JoinPromoter( q_object.connector, len(q_object.children), current_negated ) for child in q_object.children: child_clause, needed_inner = self.build_filter( child, can_reuse=used_aliases, branch_negated=branch_negated, current_negated=current_negated, allow_joins=allow_joins, split_subq=split_subq, check_filterable=check_filterable, summarize=summarize, ) joinpromoter.add_votes(needed_inner) if child_clause: target_clause.add(child_clause, connector) needed_inner = joinpromoter.update_join_types(self) return target_clause, needed_inner def build_filtered_relation_q( self, q_object, reuse, branch_negated=False, current_negated=False ): """Add a FilteredRelation object to the current filter.""" connector = q_object.connector current_negated ^= q_object.negated branch_negated = branch_negated or q_object.negated target_clause = WhereNode(connector=connector, negated=q_object.negated) for child in q_object.children: if isinstance(child, Node): child_clause = self.build_filtered_relation_q( child, reuse=reuse, branch_negated=branch_negated, current_negated=current_negated, ) else: child_clause, _ = self.build_filter( child, can_reuse=reuse, branch_negated=branch_negated, current_negated=current_negated, allow_joins=True, split_subq=False, reuse_with_filtered_relation=True, ) target_clause.add(child_clause, connector) return target_clause def add_filtered_relation(self, filtered_relation, alias): filtered_relation.alias = alias lookups = dict(get_children_from_q(filtered_relation.condition)) relation_lookup_parts, relation_field_parts, _ = self.solve_lookup_type( filtered_relation.relation_name ) if relation_lookup_parts: raise ValueError( "FilteredRelation's relation_name cannot contain lookups " "(got %r)." % filtered_relation.relation_name ) for lookup in chain(lookups): lookup_parts, lookup_field_parts, _ = self.solve_lookup_type(lookup) shift = 2 if not lookup_parts else 1 lookup_field_path = lookup_field_parts[:-shift] for idx, lookup_field_part in enumerate(lookup_field_path): if len(relation_field_parts) > idx: if relation_field_parts[idx] != lookup_field_part: raise ValueError( "FilteredRelation's condition doesn't support " "relations outside the %r (got %r)." % (filtered_relation.relation_name, lookup) ) else: raise ValueError( "FilteredRelation's condition doesn't support nested " "relations deeper than the relation_name (got %r for " "%r)." % (lookup, filtered_relation.relation_name) ) self._filtered_relations[filtered_relation.alias] = filtered_relation def names_to_path(self, names, opts, allow_many=True, fail_on_missing=False): """ Walk the list of names and turns them into PathInfo tuples. A single name in 'names' can generate multiple PathInfos (m2m, for example). 'names' is the path of names to travel, 'opts' is the model Options we start the name resolving from, 'allow_many' is as for setup_joins(). If fail_on_missing is set to True, then a name that can't be resolved will generate a FieldError. Return a list of PathInfo tuples. In addition return the final field (the last used join field) and target (which is a field guaranteed to contain the same value as the final field). Finally, return those names that weren't found (which are likely transforms and the final lookup). """ path, names_with_path = [], [] for pos, name in enumerate(names): cur_names_with_path = (name, []) if name == "pk": name = opts.pk.name field = None filtered_relation = None try: if opts is None: raise FieldDoesNotExist field = opts.get_field(name) except FieldDoesNotExist: if name in self.annotation_select: field = self.annotation_select[name].output_field elif name in self._filtered_relations and pos == 0: filtered_relation = self._filtered_relations[name] if LOOKUP_SEP in filtered_relation.relation_name: parts = filtered_relation.relation_name.split(LOOKUP_SEP) filtered_relation_path, field, _, _ = self.names_to_path( parts, opts, allow_many, fail_on_missing, ) path.extend(filtered_relation_path[:-1]) else: field = opts.get_field(filtered_relation.relation_name) if field is not None: # Fields that contain one-to-many relations with a generic # model (like a GenericForeignKey) cannot generate reverse # relations and therefore cannot be used for reverse querying. if field.is_relation and not field.related_model: raise FieldError( "Field %r does not generate an automatic reverse " "relation and therefore cannot be used for reverse " "querying. If it is a GenericForeignKey, consider " "adding a GenericRelation." % name ) try: model = field.model._meta.concrete_model except AttributeError: # QuerySet.annotate() may introduce fields that aren't # attached to a model. model = None else: # We didn't find the current field, so move position back # one step. pos -= 1 if pos == -1 or fail_on_missing: available = sorted( [ *get_field_names_from_opts(opts), *self.annotation_select, *self._filtered_relations, ] ) raise FieldError( "Cannot resolve keyword '%s' into field. " "Choices are: %s" % (name, ", ".join(available)) ) break # Check if we need any joins for concrete inheritance cases (the # field lives in parent, but we are currently in one of its # children) if opts is not None and model is not opts.model: path_to_parent = opts.get_path_to_parent(model) if path_to_parent: path.extend(path_to_parent) cur_names_with_path[1].extend(path_to_parent) opts = path_to_parent[-1].to_opts if hasattr(field, "path_infos"): if filtered_relation: pathinfos = field.get_path_info(filtered_relation) else: pathinfos = field.path_infos if not allow_many: for inner_pos, p in enumerate(pathinfos): if p.m2m: cur_names_with_path[1].extend(pathinfos[0 : inner_pos + 1]) names_with_path.append(cur_names_with_path) raise MultiJoin(pos + 1, names_with_path) last = pathinfos[-1] path.extend(pathinfos) final_field = last.join_field opts = last.to_opts targets = last.target_fields cur_names_with_path[1].extend(pathinfos) names_with_path.append(cur_names_with_path) else: # Local non-relational field. final_field = field targets = (field,) if fail_on_missing and pos + 1 != len(names): raise FieldError( "Cannot resolve keyword %r into field. Join on '%s'" " not permitted." % (names[pos + 1], name) ) break return path, final_field, targets, names[pos + 1 :] def setup_joins( self, names, opts, alias, can_reuse=None, allow_many=True, reuse_with_filtered_relation=False, ): """ Compute the necessary table joins for the passage through the fields given in 'names'. 'opts' is the Options class for the current model (which gives the table we are starting from), 'alias' is the alias for the table to start the joining from. The 'can_reuse' defines the reverse foreign key joins we can reuse. It can be None in which case all joins are reusable or a set of aliases that can be reused. Note that non-reverse foreign keys are always reusable when using setup_joins(). The 'reuse_with_filtered_relation' can be used to force 'can_reuse' parameter and force the relation on the given connections. If 'allow_many' is False, then any reverse foreign key seen will generate a MultiJoin exception. Return the final field involved in the joins, the target field (used for any 'where' constraint), the final 'opts' value, the joins, the field path traveled to generate the joins, and a transform function that takes a field and alias and is equivalent to `field.get_col(alias)` in the simple case but wraps field transforms if they were included in names. The target field is the field containing the concrete value. Final field can be something different, for example foreign key pointing to that value. Final field is needed for example in some value conversions (convert 'obj' in fk__id=obj to pk val using the foreign key field for example). """ joins = [alias] # The transform can't be applied yet, as joins must be trimmed later. # To avoid making every caller of this method look up transforms # directly, compute transforms here and create a partial that converts # fields to the appropriate wrapped version. def final_transformer(field, alias): if not self.alias_cols: alias = None return field.get_col(alias) # Try resolving all the names as fields first. If there's an error, # treat trailing names as lookups until a field can be resolved. last_field_exception = None for pivot in range(len(names), 0, -1): try: path, final_field, targets, rest = self.names_to_path( names[:pivot], opts, allow_many, fail_on_missing=True, ) except FieldError as exc: if pivot == 1: # The first item cannot be a lookup, so it's safe # to raise the field error here. raise else: last_field_exception = exc else: # The transforms are the remaining items that couldn't be # resolved into fields. transforms = names[pivot:] break for name in transforms: def transform(field, alias, *, name, previous): try: wrapped = previous(field, alias) return self.try_transform(wrapped, name) except FieldError: # FieldError is raised if the transform doesn't exist. if isinstance(final_field, Field) and last_field_exception: raise last_field_exception else: raise final_transformer = functools.partial( transform, name=name, previous=final_transformer ) final_transformer.has_transforms = True # Then, add the path to the query's joins. Note that we can't trim # joins at this stage - we will need the information about join type # of the trimmed joins. for join in path: if join.filtered_relation: filtered_relation = join.filtered_relation.clone() table_alias = filtered_relation.alias else: filtered_relation = None table_alias = None opts = join.to_opts if join.direct: nullable = self.is_nullable(join.join_field) else: nullable = True connection = self.join_class( opts.db_table, alias, table_alias, INNER, join.join_field, nullable, filtered_relation=filtered_relation, ) reuse = can_reuse if join.m2m or reuse_with_filtered_relation else None alias = self.join( connection, reuse=reuse, reuse_with_filtered_relation=reuse_with_filtered_relation, ) joins.append(alias) if filtered_relation: filtered_relation.path = joins[:] return JoinInfo(final_field, targets, opts, joins, path, final_transformer) def trim_joins(self, targets, joins, path): """ The 'target' parameter is the final field being joined to, 'joins' is the full list of join aliases. The 'path' contain the PathInfos used to create the joins. Return the final target field and table alias and the new active joins. Always trim any direct join if the target column is already in the previous table. Can't trim reverse joins as it's unknown if there's anything on the other side of the join. """ joins = joins[:] for pos, info in enumerate(reversed(path)): if len(joins) == 1 or not info.direct: break if info.filtered_relation: break join_targets = {t.column for t in info.join_field.foreign_related_fields} cur_targets = {t.column for t in targets} if not cur_targets.issubset(join_targets): break targets_dict = { r[1].column: r[0] for r in info.join_field.related_fields if r[1].column in cur_targets } targets = tuple(targets_dict[t.column] for t in targets) self.unref_alias(joins.pop()) return targets, joins[-1], joins @classmethod def _gen_cols(cls, exprs, include_external=False, resolve_refs=True): for expr in exprs: if isinstance(expr, Col): yield expr elif include_external and callable( getattr(expr, "get_external_cols", None) ): yield from expr.get_external_cols() elif hasattr(expr, "get_source_expressions"): if not resolve_refs and isinstance(expr, Ref): continue yield from cls._gen_cols( expr.get_source_expressions(), include_external=include_external, resolve_refs=resolve_refs, ) @classmethod def _gen_col_aliases(cls, exprs): yield from (expr.alias for expr in cls._gen_cols(exprs)) def resolve_ref(self, name, allow_joins=True, reuse=None, summarize=False): annotation = self.annotations.get(name) if annotation is not None: if not allow_joins: for alias in self._gen_col_aliases([annotation]): if isinstance(self.alias_map[alias], Join): raise FieldError( "Joined field references are not permitted in this query" ) if summarize: # Summarize currently means we are doing an aggregate() query # which is executed as a wrapped subquery if any of the # aggregate() elements reference an existing annotation. In # that case we need to return a Ref to the subquery's annotation. if name not in self.annotation_select: raise FieldError( "Cannot aggregate over the '%s' alias. Use annotate() " "to promote it." % name ) return Ref(name, self.annotation_select[name]) else: return annotation else: field_list = name.split(LOOKUP_SEP) annotation = self.annotations.get(field_list[0]) if annotation is not None: for transform in field_list[1:]: annotation = self.try_transform(annotation, transform) return annotation join_info = self.setup_joins( field_list, self.get_meta(), self.get_initial_alias(), can_reuse=reuse ) targets, final_alias, join_list = self.trim_joins( join_info.targets, join_info.joins, join_info.path ) if not allow_joins and len(join_list) > 1: raise FieldError( "Joined field references are not permitted in this query" ) if len(targets) > 1: raise FieldError( "Referencing multicolumn fields with F() objects isn't supported" ) # Verify that the last lookup in name is a field or a transform: # transform_function() raises FieldError if not. transform = join_info.transform_function(targets[0], final_alias) if reuse is not None: reuse.update(join_list) return transform def split_exclude(self, filter_expr, can_reuse, names_with_path): """ When doing an exclude against any kind of N-to-many relation, we need to use a subquery. This method constructs the nested query, given the original exclude filter (filter_expr) and the portion up to the first N-to-many relation field. For example, if the origin filter is ~Q(child__name='foo'), filter_expr is ('child__name', 'foo') and can_reuse is a set of joins usable for filters in the original query. We will turn this into equivalent of: WHERE NOT EXISTS( SELECT 1 FROM child WHERE name = 'foo' AND child.parent_id = parent.id LIMIT 1 ) """ # Generate the inner query. query = self.__class__(self.model) query._filtered_relations = self._filtered_relations filter_lhs, filter_rhs = filter_expr if isinstance(filter_rhs, OuterRef): filter_rhs = OuterRef(filter_rhs) elif isinstance(filter_rhs, F): filter_rhs = OuterRef(filter_rhs.name) query.add_filter(filter_lhs, filter_rhs) query.clear_ordering(force=True) # Try to have as simple as possible subquery -> trim leading joins from # the subquery. trimmed_prefix, contains_louter = query.trim_start(names_with_path) col = query.select[0] select_field = col.target alias = col.alias if alias in can_reuse: pk = select_field.model._meta.pk # Need to add a restriction so that outer query's filters are in effect for # the subquery, too. query.bump_prefix(self) lookup_class = select_field.get_lookup("exact") # Note that the query.select[0].alias is different from alias # due to bump_prefix above. lookup = lookup_class(pk.get_col(query.select[0].alias), pk.get_col(alias)) query.where.add(lookup, AND) query.external_aliases[alias] = True lookup_class = select_field.get_lookup("exact") lookup = lookup_class(col, ResolvedOuterRef(trimmed_prefix)) query.where.add(lookup, AND) condition, needed_inner = self.build_filter(Exists(query)) if contains_louter: or_null_condition, _ = self.build_filter( ("%s__isnull" % trimmed_prefix, True), current_negated=True, branch_negated=True, can_reuse=can_reuse, ) condition.add(or_null_condition, OR) # Note that the end result will be: # (outercol NOT IN innerq AND outercol IS NOT NULL) OR outercol IS NULL. # This might look crazy but due to how IN works, this seems to be # correct. If the IS NOT NULL check is removed then outercol NOT # IN will return UNKNOWN. If the IS NULL check is removed, then if # outercol IS NULL we will not match the row. return condition, needed_inner def set_empty(self): self.where.add(NothingNode(), AND) for query in self.combined_queries: query.set_empty() def is_empty(self): return any(isinstance(c, NothingNode) for c in self.where.children) def set_limits(self, low=None, high=None): """ Adjust the limits on the rows retrieved. Use low/high to set these, as it makes it more Pythonic to read and write. When the SQL query is created, convert them to the appropriate offset and limit values. Apply any limits passed in here to the existing constraints. Add low to the current low value and clamp both to any existing high value. """ if high is not None: if self.high_mark is not None: self.high_mark = min(self.high_mark, self.low_mark + high) else: self.high_mark = self.low_mark + high if low is not None: if self.high_mark is not None: self.low_mark = min(self.high_mark, self.low_mark + low) else: self.low_mark = self.low_mark + low if self.low_mark == self.high_mark: self.set_empty() def clear_limits(self): """Clear any existing limits.""" self.low_mark, self.high_mark = 0, None @property def is_sliced(self): return self.low_mark != 0 or self.high_mark is not None def has_limit_one(self): return self.high_mark is not None and (self.high_mark - self.low_mark) == 1 def can_filter(self): """ Return True if adding filters to this instance is still possible. Typically, this means no limits or offsets have been put on the results. """ return not self.is_sliced def clear_select_clause(self): """Remove all fields from SELECT clause.""" self.select = () self.default_cols = False self.select_related = False self.set_extra_mask(()) self.set_annotation_mask(()) def clear_select_fields(self): """ Clear the list of fields to select (but not extra_select columns). Some queryset types completely replace any existing list of select columns. """ self.select = () self.values_select = () def add_select_col(self, col, name): self.select += (col,) self.values_select += (name,) def set_select(self, cols): self.default_cols = False self.select = tuple(cols) def add_distinct_fields(self, *field_names): """ Add and resolve the given fields to the query's "distinct on" clause. """ self.distinct_fields = field_names self.distinct = True def add_fields(self, field_names, allow_m2m=True): """ Add the given (model) fields to the select set. Add the field names in the order specified. """ alias = self.get_initial_alias() opts = self.get_meta() try: cols = [] for name in field_names: # Join promotion note - we must not remove any rows here, so # if there is no existing joins, use outer join. join_info = self.setup_joins( name.split(LOOKUP_SEP), opts, alias, allow_many=allow_m2m ) targets, final_alias, joins = self.trim_joins( join_info.targets, join_info.joins, join_info.path, ) for target in targets: cols.append(join_info.transform_function(target, final_alias)) if cols: self.set_select(cols) except MultiJoin: raise FieldError("Invalid field name: '%s'" % name) except FieldError: if LOOKUP_SEP in name: # For lookups spanning over relationships, show the error # from the model on which the lookup failed. raise elif name in self.annotations: raise FieldError( "Cannot select the '%s' alias. Use annotate() to promote " "it." % name ) else: names = sorted( [ *get_field_names_from_opts(opts), *self.extra, *self.annotation_select, *self._filtered_relations, ] ) raise FieldError( "Cannot resolve keyword %r into field. " "Choices are: %s" % (name, ", ".join(names)) ) def add_ordering(self, *ordering): """ Add items from the 'ordering' sequence to the query's "order by" clause. These items are either field names (not column names) -- possibly with a direction prefix ('-' or '?') -- or OrderBy expressions. If 'ordering' is empty, clear all ordering from the query. """ errors = [] for item in ordering: if isinstance(item, str): if item == "?": continue if item.startswith("-"): item = item[1:] if item in self.annotations: continue if self.extra and item in self.extra: continue # names_to_path() validates the lookup. A descriptive # FieldError will be raise if it's not. self.names_to_path(item.split(LOOKUP_SEP), self.model._meta) elif not hasattr(item, "resolve_expression"): errors.append(item) if getattr(item, "contains_aggregate", False): raise FieldError( "Using an aggregate in order_by() without also including " "it in annotate() is not allowed: %s" % item ) if errors: raise FieldError("Invalid order_by arguments: %s" % errors) if ordering: self.order_by += ordering else: self.default_ordering = False def clear_ordering(self, force=False, clear_default=True): """ Remove any ordering settings if the current query allows it without side effects, set 'force' to True to clear the ordering regardless. If 'clear_default' is True, there will be no ordering in the resulting query (not even the model's default). """ if not force and ( self.is_sliced or self.distinct_fields or self.select_for_update ): return self.order_by = () self.extra_order_by = () if clear_default: self.default_ordering = False def set_group_by(self, allow_aliases=True): """ Expand the GROUP BY clause required by the query. This will usually be the set of all non-aggregate fields in the return data. If the database backend supports grouping by the primary key, and the query would be equivalent, the optimization will be made automatically. """ if allow_aliases and self.values_select: # If grouping by aliases is allowed assign selected value aliases # by moving them to annotations. group_by_annotations = {} values_select = {} for alias, expr in zip(self.values_select, self.select): if isinstance(expr, Col): values_select[alias] = expr else: group_by_annotations[alias] = expr self.annotations = {**group_by_annotations, **self.annotations} self.append_annotation_mask(group_by_annotations) self.select = tuple(values_select.values()) self.values_select = tuple(values_select) group_by = list(self.select) for alias, annotation in self.annotation_select.items(): if not (group_by_cols := annotation.get_group_by_cols()): continue if allow_aliases and not annotation.contains_aggregate: group_by.append(Ref(alias, annotation)) else: group_by.extend(group_by_cols) self.group_by = tuple(group_by) def add_select_related(self, fields): """ Set up the select_related data structure so that we only select certain related models (as opposed to all models, when self.select_related=True). """ if isinstance(self.select_related, bool): field_dict = {} else: field_dict = self.select_related for field in fields: d = field_dict for part in field.split(LOOKUP_SEP): d = d.setdefault(part, {}) self.select_related = field_dict def add_extra(self, select, select_params, where, params, tables, order_by): """ Add data to the various extra_* attributes for user-created additions to the query. """ if select: # We need to pair any placeholder markers in the 'select' # dictionary with their parameters in 'select_params' so that # subsequent updates to the select dictionary also adjust the # parameters appropriately. select_pairs = {} if select_params: param_iter = iter(select_params) else: param_iter = iter([]) for name, entry in select.items(): self.check_alias(name) entry = str(entry) entry_params = [] pos = entry.find("%s") while pos != -1: if pos == 0 or entry[pos - 1] != "%": entry_params.append(next(param_iter)) pos = entry.find("%s", pos + 2) select_pairs[name] = (entry, entry_params) self.extra.update(select_pairs) if where or params: self.where.add(ExtraWhere(where, params), AND) if tables: self.extra_tables += tuple(tables) if order_by: self.extra_order_by = order_by def clear_deferred_loading(self): """Remove any fields from the deferred loading set.""" self.deferred_loading = (frozenset(), True) def add_deferred_loading(self, field_names): """ Add the given list of model field names to the set of fields to exclude from loading from the database when automatic column selection is done. Add the new field names to any existing field names that are deferred (or removed from any existing field names that are marked as the only ones for immediate loading). """ # Fields on related models are stored in the literal double-underscore # format, so that we can use a set datastructure. We do the foo__bar # splitting and handling when computing the SQL column names (as part of # get_columns()). existing, defer = self.deferred_loading if defer: # Add to existing deferred names. self.deferred_loading = existing.union(field_names), True else: # Remove names from the set of any existing "immediate load" names. if new_existing := existing.difference(field_names): self.deferred_loading = new_existing, False else: self.clear_deferred_loading() if new_only := set(field_names).difference(existing): self.deferred_loading = new_only, True def add_immediate_loading(self, field_names): """ Add the given list of model field names to the set of fields to retrieve when the SQL is executed ("immediate loading" fields). The field names replace any existing immediate loading field names. If there are field names already specified for deferred loading, remove those names from the new field_names before storing the new names for immediate loading. (That is, immediate loading overrides any existing immediate values, but respects existing deferrals.) """ existing, defer = self.deferred_loading field_names = set(field_names) if "pk" in field_names: field_names.remove("pk") field_names.add(self.get_meta().pk.name) if defer: # Remove any existing deferred names from the current set before # setting the new names. self.deferred_loading = field_names.difference(existing), False else: # Replace any existing "immediate load" field names. self.deferred_loading = frozenset(field_names), False def set_annotation_mask(self, names): """Set the mask of annotations that will be returned by the SELECT.""" if names is None: self.annotation_select_mask = None else: self.annotation_select_mask = set(names) self._annotation_select_cache = None def append_annotation_mask(self, names): if self.annotation_select_mask is not None: self.set_annotation_mask(self.annotation_select_mask.union(names)) def set_extra_mask(self, names): """ Set the mask of extra select items that will be returned by SELECT. Don't remove them from the Query since they might be used later. """ if names is None: self.extra_select_mask = None else: self.extra_select_mask = set(names) self._extra_select_cache = None def set_values(self, fields): self.select_related = False self.clear_deferred_loading() self.clear_select_fields() self.has_select_fields = True if fields: field_names = [] extra_names = [] annotation_names = [] if not self.extra and not self.annotations: # Shortcut - if there are no extra or annotations, then # the values() clause must be just field names. field_names = list(fields) else: self.default_cols = False for f in fields: if f in self.extra_select: extra_names.append(f) elif f in self.annotation_select: annotation_names.append(f) else: field_names.append(f) self.set_extra_mask(extra_names) self.set_annotation_mask(annotation_names) selected = frozenset(field_names + extra_names + annotation_names) else: field_names = [f.attname for f in self.model._meta.concrete_fields] selected = frozenset(field_names) # Selected annotations must be known before setting the GROUP BY # clause. if self.group_by is True: self.add_fields( (f.attname for f in self.model._meta.concrete_fields), False ) # Disable GROUP BY aliases to avoid orphaning references to the # SELECT clause which is about to be cleared. self.set_group_by(allow_aliases=False) self.clear_select_fields() elif self.group_by: # Resolve GROUP BY annotation references if they are not part of # the selected fields anymore. group_by = [] for expr in self.group_by: if isinstance(expr, Ref) and expr.refs not in selected: expr = self.annotations[expr.refs] group_by.append(expr) self.group_by = tuple(group_by) self.values_select = tuple(field_names) self.add_fields(field_names, True) @property def annotation_select(self): """ Return the dictionary of aggregate columns that are not masked and should be used in the SELECT clause. Cache this result for performance. """ if self._annotation_select_cache is not None: return self._annotation_select_cache elif not self.annotations: return {} elif self.annotation_select_mask is not None: self._annotation_select_cache = { k: v for k, v in self.annotations.items() if k in self.annotation_select_mask } return self._annotation_select_cache else: return self.annotations @property def extra_select(self): if self._extra_select_cache is not None: return self._extra_select_cache if not self.extra: return {} elif self.extra_select_mask is not None: self._extra_select_cache = { k: v for k, v in self.extra.items() if k in self.extra_select_mask } return self._extra_select_cache else: return self.extra def trim_start(self, names_with_path): """ Trim joins from the start of the join path. The candidates for trim are the PathInfos in names_with_path structure that are m2m joins. Also set the select column so the start matches the join. This method is meant to be used for generating the subquery joins & cols in split_exclude(). Return a lookup usable for doing outerq.filter(lookup=self) and a boolean indicating if the joins in the prefix contain a LEFT OUTER join. _""" all_paths = [] for _, paths in names_with_path: all_paths.extend(paths) contains_louter = False # Trim and operate only on tables that were generated for # the lookup part of the query. That is, avoid trimming # joins generated for F() expressions. lookup_tables = [ t for t in self.alias_map if t in self._lookup_joins or t == self.base_table ] for trimmed_paths, path in enumerate(all_paths): if path.m2m: break if self.alias_map[lookup_tables[trimmed_paths + 1]].join_type == LOUTER: contains_louter = True alias = lookup_tables[trimmed_paths] self.unref_alias(alias) # The path.join_field is a Rel, lets get the other side's field join_field = path.join_field.field # Build the filter prefix. paths_in_prefix = trimmed_paths trimmed_prefix = [] for name, path in names_with_path: if paths_in_prefix - len(path) < 0: break trimmed_prefix.append(name) paths_in_prefix -= len(path) trimmed_prefix.append(join_field.foreign_related_fields[0].name) trimmed_prefix = LOOKUP_SEP.join(trimmed_prefix) # Lets still see if we can trim the first join from the inner query # (that is, self). We can't do this for: # - LEFT JOINs because we would miss those rows that have nothing on # the outer side, # - INNER JOINs from filtered relations because we would miss their # filters. first_join = self.alias_map[lookup_tables[trimmed_paths + 1]] if first_join.join_type != LOUTER and not first_join.filtered_relation: select_fields = [r[0] for r in join_field.related_fields] select_alias = lookup_tables[trimmed_paths + 1] self.unref_alias(lookup_tables[trimmed_paths]) extra_restriction = join_field.get_extra_restriction( None, lookup_tables[trimmed_paths + 1] ) if extra_restriction: self.where.add(extra_restriction, AND) else: # TODO: It might be possible to trim more joins from the start of the # inner query if it happens to have a longer join chain containing the # values in select_fields. Lets punt this one for now. select_fields = [r[1] for r in join_field.related_fields] select_alias = lookup_tables[trimmed_paths] # The found starting point is likely a join_class instead of a # base_table_class reference. But the first entry in the query's FROM # clause must not be a JOIN. for table in self.alias_map: if self.alias_refcount[table] > 0: self.alias_map[table] = self.base_table_class( self.alias_map[table].table_name, table, ) break self.set_select([f.get_col(select_alias) for f in select_fields]) return trimmed_prefix, contains_louter def is_nullable(self, field): """ Check if the given field should be treated as nullable. Some backends treat '' as null and Django treats such fields as nullable for those backends. In such situations field.null can be False even if we should treat the field as nullable. """ # We need to use DEFAULT_DB_ALIAS here, as QuerySet does not have # (nor should it have) knowledge of which connection is going to be # used. The proper fix would be to defer all decisions where # is_nullable() is needed to the compiler stage, but that is not easy # to do currently. return field.null or ( field.empty_strings_allowed and connections[DEFAULT_DB_ALIAS].features.interprets_empty_strings_as_nulls ) def get_order_dir(field, default="ASC"): """ Return the field name and direction for an order specification. For example, '-foo' is returned as ('foo', 'DESC'). The 'default' param is used to indicate which way no prefix (or a '+' prefix) should sort. The '-' prefix always sorts the opposite way. """ dirn = ORDER_DIR[default] if field[0] == "-": return field[1:], dirn[1] return field, dirn[0] class JoinPromoter: """ A class to abstract away join promotion problems for complex filter conditions. """ def __init__(self, connector, num_children, negated): self.connector = connector self.negated = negated if self.negated: if connector == AND: self.effective_connector = OR else: self.effective_connector = AND else: self.effective_connector = self.connector self.num_children = num_children # Maps of table alias to how many times it is seen as required for # inner and/or outer joins. self.votes = Counter() def __repr__(self): return ( f"{self.__class__.__qualname__}(connector={self.connector!r}, " f"num_children={self.num_children!r}, negated={self.negated!r})" ) def add_votes(self, votes): """ Add single vote per item to self.votes. Parameter can be any iterable. """ self.votes.update(votes) def update_join_types(self, query): """ Change join types so that the generated query is as efficient as possible, but still correct. So, change as many joins as possible to INNER, but don't make OUTER joins INNER if that could remove results from the query. """ to_promote = set() to_demote = set() # The effective_connector is used so that NOT (a AND b) is treated # similarly to (a OR b) for join promotion. for table, votes in self.votes.items(): # We must use outer joins in OR case when the join isn't contained # in all of the joins. Otherwise the INNER JOIN itself could remove # valid results. Consider the case where a model with rel_a and # rel_b relations is queried with rel_a__col=1 | rel_b__col=2. Now, # if rel_a join doesn't produce any results is null (for example # reverse foreign key or null value in direct foreign key), and # there is a matching row in rel_b with col=2, then an INNER join # to rel_a would remove a valid match from the query. So, we need # to promote any existing INNER to LOUTER (it is possible this # promotion in turn will be demoted later on). if self.effective_connector == OR and votes < self.num_children: to_promote.add(table) # If connector is AND and there is a filter that can match only # when there is a joinable row, then use INNER. For example, in # rel_a__col=1 & rel_b__col=2, if either of the rels produce NULL # as join output, then the col=1 or col=2 can't match (as # NULL=anything is always false). # For the OR case, if all children voted for a join to be inner, # then we can use INNER for the join. For example: # (rel_a__col__icontains=Alex | rel_a__col__icontains=Russell) # then if rel_a doesn't produce any rows, the whole condition # can't match. Hence we can safely use INNER join. if self.effective_connector == AND or ( self.effective_connector == OR and votes == self.num_children ): to_demote.add(table) # Finally, what happens in cases where we have: # (rel_a__col=1|rel_b__col=2) & rel_a__col__gte=0 # Now, we first generate the OR clause, and promote joins for it # in the first if branch above. Both rel_a and rel_b are promoted # to LOUTER joins. After that we do the AND case. The OR case # voted no inner joins but the rel_a__col__gte=0 votes inner join # for rel_a. We demote it back to INNER join (in AND case a single # vote is enough). The demotion is OK, if rel_a doesn't produce # rows, then the rel_a__col__gte=0 clause can't be true, and thus # the whole clause must be false. So, it is safe to use INNER # join. # Note that in this example we could just as well have the __gte # clause and the OR clause swapped. Or we could replace the __gte # clause with an OR clause containing rel_a__col=1|rel_a__col=2, # and again we could safely demote to INNER. query.promote_joins(to_promote) query.demote_joins(to_demote) return to_demote