sql: rfc for foreign keys

Part of #3271

doc/rfc/3271-foreign-keys.md

+# Foreign Keys
+
+* **Status**: In progress
+* **Start date**: 27-06-2018
+* **Authors**: Nikita Pettik @korablev77 korablev@tarantool.org
+* **Issues**: [#3271](https://github.com/tarantool/tarantool/issues/3271)
+
+## Summary
+
+Introduce Foreign Keys (FK) constraints in server.
+
+## Background and motivation
+
+Used terminology:
+```
+CREATE TABLE t1 (id1 INT PRIMARY KEY, a INT);  
+CREATE TABLE t2 (id2 INT PRIMARY KEY REFERENECES t1(id1));
+```
+t1 - parent or referenced table; id1 - referenced columns;  
+t2 - child or referencing table; id2 - referencing columns;
+
+Originally, SQLite provides quite specific way of handling FK constraints:
+any checks are deferred as much as possible. For instance, existence of
+parent table's existence is verified each time on data manipulation in
+child space, but not checked on creation of FK constraint itself.
+The same is for UNIQUE index: according to ANSI referenced columns must
+make up UNIQUE constraint; otherwise error is raised. But in SQLite
+presence of such index is tested only when child table is involved in DML.
+On the other hand it allows us to get rid of checks on indexes/spaces
+destroying; moreover, it allows us to avoid storing additional dependencies
+in internal data dictionary holding only space name and extracting fresh
+pointers to DD by lookup in hash.
+
+To be closer to ANSI FK constraints, it was suggested to rework and
+move them to server DD.
+
+## Implementation details
+
+1. Prohibit actions with FK constraints which strictly contradict ANSI:
+ - Ban opportunity to create FK constraint on table which doesn't exist
+   (except for self-references). Since SQLite allows creating
+   FK constraints only within <CREATE TABLE> declaration (i.e. there is
+   no separate <ALTER TABLE ADD CONSTRAINT> statement), this step will
+   deprive us of ability to create circular FK dependencies.
+ - Ban opportunity to drop parent table even if it doesn't violate FK
+   constraints. Child table must always be dropped before parent.
+ - Ban opportunity to create FK constraints on VIEW.
+
+2. To return ability to create circular FK dependecies, we are going to
+   introduce <ALTER TABLE ADD CONSTRAINT> statement. It compels us to
+   store somehow created constraints. Hence, new system space to
+   persist FK constraints is added:
+```
+ - [350, 1, '_constraint', 'memtx', 0, {}, [{'name': 'name', 'type': 'string'}, {
+      'name': 'parent_id', 'type': 'unsigned'}, {'name': 'child_id', 'type': 'unsigned'},
+     {'name': 'deferred', 'type': 'boolean'}, {'name': 'match', 'type': 'string'},
+     {'name': 'on_delete', 'type': 'string'}, {'name': 'on_update', 'type': 'string'},
+     {'name': 'links', 'type': 'map'}]]
+```
+In other words, FK constraint is completely described by:
+ - Its name;
+ - Referenced space;
+ - Child space;
+ - Time of resolution (deferred until the end of transaction or not);
+ - Match clause;
+ - ON DELETE and ON UPDATE actions;
+ - Column links;
+
+3. Insertion on _constraint space leads to creation of FK constraints.
+   FK constraints in DD are represented as following structs:
+```
+struct fkey_def {
+       /** id of child space. */
+       uint32_t child_id;
+       /** id of parent space. */
+       uint32_t parent_id;
+       /** Number of fields (links) in this key. */
+       uint32_t field_count;
+       /** True if constraint checking is deferred till COMMIT. */
+       bool is_deferred;
+       /** Match condition for foreign key. SIMPLE by default. */
+       enum fkey_match match;
+       /** ON DELETE action. */
+       enum fkey_action on_delete;
+       /** ON UPDATE action. */
+       enum fkey_action on_update;
+       /** Mapping of fields in child to fields in parent. */
+       struct field_link *links;
+       /** Name of the constraint. */
+       char name[0];
+};
+
+struct fkey {
+       /** Space containing the REFERENCES clause (aka: child). */
+       struct space *child;
+       /** Space that the key points to (aka: parent). */
+       struct space *parent;
+       /** Number of fields (links) in this key. */
+       uint32_t field_count;
+       /** True if constraint checking is deferred till COMMIT. */
+       bool is_deferred;
+       /** Match condition for foreign key. SIMPLE by default. */
+       enum fkey_match match;
+       /** Triggers for actions. */
+       struct Trigger *on_delete_trigger;
+       struct Trigger *on_update_trigger;
+       /** Contraints are orginized into list. */
+       struct fkey *fkey_next;
+       /** Mapping of fields in child to fields in parent. */
+       struct field_link *links;
+       /** Name of the constraint. */
+       char name[0];
+};
+```
+
+struct fkey_def is used only during creation of struct fkey.
+Foreign keys will be stored in struct space as two linked lists:
+
+```
+@@ -182,7 +182,8 @@ struct space {
+         */
+        struct index **index;
+        /** Foreign key constraints. */
++       struct fkey *parent_fkey;
++       struct fkey *child_fkey;
+```
+
+on_replace_dd_constraint() would create definition of FK constrain
+from tuple; from def construct struct fkey, and finally add it
+to parent's list of FK constraints and to child's one.
+In case of WAL fail, simply remove it from those lists.
+On drop of FK actions are inverted: FK constraint is removed from
+parent and child lists and if WAL fails, it will be returned back.
+
+4. Alongside with insertion to _constraint, we must check that tuples
+   which are already in child table don't violate FK constraint under
+   construction. There several approaches to resolve this problem.
+   The first and common one is to handle this routine within
+   on_replace_dd_constraint trigger: create iterators for parent and
+   child spaces and for each tuple from child space find appropriate
+   tuple with given FK key in parent space. The question here is how
+   to process inserted or deleted tuples until FK constraint is committed?
+   Another solution is to temporary execute this check only if creation
+   of FK consraint is occurred via SQL: before insertion to _constraint
+   run VDBE program which will test this condition. Anyway, now it is
+   impossible to resolve FK constraints (in common sense) without
+   involving VDBE: we can't run ON DELETE and ON UPDATE triggers directly
+   from server. The last and the easiest way is to allow creation of
+   FK constraint only for empty spaces and remove this restriction when
+   we will be able to run any VDBE code from server.
+
+5. Introduce SQL statement to create and drop FK constraints (ANSI syntax):
+
+```
+ALTER TABLE <referencing table> ADD CONSTRAINT
+  <referential constraint name> FOREIGN KEY
+  <left parent> <referencing columns> <right paren> REFERENCES
+  <referenced table> [ <referenced columns> ] [ MATCH <match type> ]
+  [ <referential triggered action> ] [ <constraint check time> ]
+
+ALTER TABLE <referencing table> DROP CONSTRAINT <referential constrain name>
+```
+In terms of our SQL parser:
+
+```
+cmd ::= ALTER TABLE fullname(X) ADD CONSTRAINT nm(Z) FOREIGN KEY
+        LP eidlist(FA) RP REFERENCES nm(T) eidlist_opt(TA) refargs(R)
+        defer_subclause_opt(D).
+
+cmd ::= ALTER TABLE fullname(X) DROP CONSTRAINT nm(Z).
+```
+Example:
+```
+ALTER TABLE t1 ADD CONSTRAINT f1 FOREIGN KEY(id, a) REFERENCES t2 (id, b) MATCH FULL;
+ALTER TABLE t1 DROP CONSTRAINT f1;
+```
+These statements are going to emit VDBE code to construct appropriate
+tuple and insert it into _constraint space (and maybe run VDBE program
+to make sure that all tuples in child space satisfy new FK constraint).
+The rest of routine will be handled by on_replace_dd_constraint() trigger.
+
+6. Refactor SQL routine to operate on new struct fkey instead of
+   obsolete SQLite struct FKey.
+
+## Open questions
+
+Should we resolve fields and indexes right after insertion to
+_constraint (as it happens in other DBs) or defer it until usage of FK
+(as it occurs in SQLite)? If we chose first way, we would have to
+implicitly link index to particular index and ban ability to drop such
+index until all FK constraints are dropped. Such behaviour may be not
+so obvious for users, but it is used for instance in PostgreSQL and MySQL.
\ No newline at end of file