From ada2e9aed8f03a177ec63c31e25fc474a281ee41 Mon Sep 17 00:00:00 2001 From: Alberto Piai Date: Sun, 29 Mar 2026 21:45:50 +0200 Subject: [PATCH v2 2/2] Try to avoid a rewrite when adding a stored generated column This builds upon basic support for ... ALTER COLUMN ... ADD GENERATED ALWAYS AS (expr) STORED If we can find a constraint which proves that the given column is already always equal to the new generated column expression, skip the expensive rewrite of the table. The check constraint must use an equality operator which is mergejoinable, and the expression must match exactly the generated column's default expression. --- src/backend/catalog/pg_constraint.c | 75 +++++++ src/backend/commands/tablecmds.c | 46 +++-- src/include/catalog/pg_constraint.h | 2 + src/test/regress/expected/alter_table.out | 232 +++++++++++++++++++++- src/test/regress/sql/alter_table.sql | 147 ++++++++++++++ 5 files changed, 481 insertions(+), 21 deletions(-) diff --git a/src/backend/catalog/pg_constraint.c b/src/backend/catalog/pg_constraint.c index b12765ae691..cfb0a0068cf 100644 --- a/src/backend/catalog/pg_constraint.c +++ b/src/backend/catalog/pg_constraint.c @@ -17,6 +17,7 @@ #include "access/genam.h" #include "access/gist.h" #include "access/htup_details.h" +#include "access/relation.h" #include "access/sysattr.h" #include "access/table.h" #include "catalog/catalog.h" @@ -29,6 +30,8 @@ #include "catalog/pg_type.h" #include "commands/defrem.h" #include "common/int.h" +#include "nodes/nodeFuncs.h" +#include "parser/parse_relation.h" #include "utils/array.h" #include "utils/builtins.h" #include "utils/fmgroids.h" @@ -694,6 +697,78 @@ findDomainNotNullConstraint(Oid typid) return retval; } +/* + * Given a relation, an attnum and a (cooked) expression, this returns true if + * it finds a CHECK constraint which proves that the given column is equal to + * the expression. + * + * The constraint must use a mergejoinable operator for the type of the column, + * a concept used by the planner as well to infer equivalence classes on the + * terms in a query (see op_mergejoinable()). + * + * The expressions are compared structurally, so they must match exactly for + * this check to succeed. + */ +bool +findStructuralCheckConstraintOnAttr(Oid relid, AttrNumber attnum, + const Node *target_expr) +{ + Relation pg_constraint; + HeapTuple conTup; + SysScanDesc scan; + ScanKeyData key; + bool found = false; + + pg_constraint = table_open(ConstraintRelationId, AccessShareLock); + ScanKeyInit(&key, + Anum_pg_constraint_conrelid, + BTEqualStrategyNumber, F_OIDEQ, + ObjectIdGetDatum(relid)); + scan = systable_beginscan(pg_constraint, ConstraintRelidTypidNameIndexId, + true, NULL, 1, &key); + + while (HeapTupleIsValid(conTup = systable_getnext(scan))) + { + Form_pg_constraint con = GETSTRUCT(conTup); + char *conbin; + Datum val; + Node *conexpr; + + if (con->contype != CONSTRAINT_CHECK) + continue; + if (!con->convalidated) + continue; + + val = SysCacheGetAttrNotNull(CONSTROID, conTup, + Anum_pg_constraint_conbin); + conbin = TextDatumGetCString(val); + conexpr = stringToNode(conbin); + + if (IsA(conexpr, OpExpr)) + { + OpExpr *op = (OpExpr *) conexpr; + + if (list_length(op->args) == 2 && IsA(linitial(op->args), Var)) + { + Var *var = linitial(op->args); + + if (var->varattno == attnum && + op_mergejoinable(op->opno, exprType((Node *) var)) && + equal(lsecond(op->args), target_expr)) + { + found = true; + break; + } + } + } + } + + systable_endscan(scan); + table_close(pg_constraint, AccessShareLock); + + return found; +} + /* * Given a pg_constraint tuple for a not-null constraint, return the column * number it is for. diff --git a/src/backend/commands/tablecmds.c b/src/backend/commands/tablecmds.c index 66622bf4837..1d98287af62 100644 --- a/src/backend/commands/tablecmds.c +++ b/src/backend/commands/tablecmds.c @@ -8911,6 +8911,7 @@ ATExecAddGeneratedAsExprStored(AlteredTableInfo *tab, NewColumnValue *newval; RawColumnDefault *rawEnt; Relation pg_attribute; + bool rewrite; Assert(def->raw_expr != NULL); Assert(def->cooked_expr == NULL); @@ -8974,29 +8975,36 @@ ATExecAddGeneratedAsExprStored(AlteredTableInfo *tab, /* Make above changes visible */ CommandCounterIncrement(); - /* - * Clear all the missing values if we're rewriting the table, since this - * renders them pointless. - */ - RelationClearMissing(rel); - - /* Make above changes visible */ - CommandCounterIncrement(); - - /* Drop any pg_statistic entry for the column */ - RemoveStatistics(RelationGetRelid(rel), attnum); - /* Build a concrete expression for the new default (generated) value */ defval = (Expr *) build_column_default(rel, attnum); defval = expression_planner(defval); - /* Schedule a rewrite */ - newval = palloc0_object(NewColumnValue); - newval->attnum = attnum; - newval->expr = defval; - newval->is_generated = true; - tab->newvals = lappend(tab->newvals, newval); - tab->rewrite |= AT_REWRITE_DEFAULT_VAL; + rewrite = !findStructuralCheckConstraintOnAttr(RelationGetRelid(rel), + attnum, + (Node *) defval); + + if (rewrite) + { + /* + * Clear all the missing values if we're rewriting the table, since + * this renders them pointless. + */ + RelationClearMissing(rel); + + /* Make above changes visible */ + CommandCounterIncrement(); + + /* Drop any pg_statistic entry for the column */ + RemoveStatistics(RelationGetRelid(rel), attnum); + + /* Schedule a rewrite */ + newval = palloc0_object(NewColumnValue); + newval->attnum = attnum; + newval->expr = defval; + newval->is_generated = true; + tab->newvals = lappend(tab->newvals, newval); + tab->rewrite |= AT_REWRITE_DEFAULT_VAL; + } InvokeObjectPostAlterHook(RelationRelationId, RelationGetRelid(rel), attnum); diff --git a/src/include/catalog/pg_constraint.h b/src/include/catalog/pg_constraint.h index 1b7fedf1750..7ac9e00c28b 100644 --- a/src/include/catalog/pg_constraint.h +++ b/src/include/catalog/pg_constraint.h @@ -266,6 +266,8 @@ extern char *ChooseConstraintName(const char *name1, const char *name2, extern HeapTuple findNotNullConstraintAttnum(Oid relid, AttrNumber attnum); extern HeapTuple findNotNullConstraint(Oid relid, const char *colname); extern HeapTuple findDomainNotNullConstraint(Oid typid); +extern bool findStructuralCheckConstraintOnAttr(Oid relid, AttrNumber attnum, + const Node *target_expr); extern AttrNumber extractNotNullColumn(HeapTuple constrTup); extern bool AdjustNotNullInheritance(Oid relid, AttrNumber attnum, const char *new_conname, bool is_local, bool is_no_inherit, bool is_notvalid); diff --git a/src/test/regress/expected/alter_table.out b/src/test/regress/expected/alter_table.out index 2567d918ec3..b3ee82f401e 100644 --- a/src/test/regress/expected/alter_table.out +++ b/src/test/regress/expected/alter_table.out @@ -4943,6 +4943,233 @@ select :idx_filenode_before != :idx_filenode_after as did_rewrite_idx; t (1 row) +drop table testgen.t3; +-- turning a regular column into a stored generated column +-- without rewriting the table (when a check constraint proves it isn't needed) +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b = a * 2); +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 2) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before = :t4_filenode_after as did_skip_rewrite; + did_skip_rewrite +------------------ + t +(1 row) + +\d+ testgen.t4 + Table "testgen.t4" + Column | Type | Collation | Nullable | Default | Storage | Stats target | Description +--------+---------+-----------+----------+------------------------------------+---------+--------------+------------- + a | integer | | | | plain | | + b | integer | | not null | generated always as (a * 2) stored | plain | | +Check constraints: + "chk_gen_clause" CHECK (b = (a * 2)) +Not-null constraints: + "t4_b_not_null" NOT NULL "b" + +drop table testgen.t4; +-- turning a regular column into a stored generated column +-- same as the previous case, but a rewrite happens since the constraint is not +-- valid +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b = a * 2) not valid; +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 2) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before != :t4_filenode_after as did_rewrite; + did_rewrite +------------- + t +(1 row) + +\d+ testgen.t4 + Table "testgen.t4" + Column | Type | Collation | Nullable | Default | Storage | Stats target | Description +--------+---------+-----------+----------+------------------------------------+---------+--------------+------------- + a | integer | | | | plain | | + b | integer | | not null | generated always as (a * 2) stored | plain | | +Check constraints: + "chk_gen_clause" CHECK (b = (a * 2)) NOT VALID +Not-null constraints: + "t4_b_not_null" NOT NULL "b" + +drop table testgen.t4; +-- turning a regular column into a stored generated column +-- same as the previous case, but a rewrite happens since the constraint +-- operator is not mergejoinable +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b >= a * 2); +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 3) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before != :t4_filenode_after as did_rewrite; + did_rewrite +------------- + t +(1 row) + +\d+ testgen.t4 + Table "testgen.t4" + Column | Type | Collation | Nullable | Default | Storage | Stats target | Description +--------+---------+-----------+----------+------------------------------------+---------+--------------+------------- + a | integer | | | | plain | | + b | integer | | not null | generated always as (a * 3) stored | plain | | +Check constraints: + "chk_gen_clause" CHECK (b >= (a * 2)) +Not-null constraints: + "t4_b_not_null" NOT NULL "b" + +drop table testgen.t4; +-- test the whole process for adding a stored generated column without +-- long-lived exclusive locks +create table testgen.t5 (a int); +select pg_relation_filenode('testgen.t5') as t5_filenode_before \gset +insert into testgen.t5 select x from generate_series(1, 5) x; +alter table testgen.t5 add column b int; +-- take care of new and updated columns +create function testgen.gen () returns trigger language plpgsql as $$ +begin + new.b = new.a * 2; return new; +end +$$; +create trigger testgen_gen + before insert or update on testgen.t5 + for each row execute function testgen.gen(); +-- add the constraint as not valid: enforced only for new and updated rows +begin; +alter table testgen.t5 + add constraint chk_gen_clause check (b = a * 2) not valid; +select locktype, mode from pg_locks + where relation = 'testgen.t5'::regclass and granted; + locktype | mode +----------+--------------------- + relation | AccessExclusiveLock +(1 row) + +commit; +insert into testgen.t5 (a) values (100), (200), (300); +-- backfill existing rows at the appropriate pace +update testgen.t5 set b = a * 2 where b is null; +-- validate: this scans the table, but without an exclusive lock +begin; +alter table testgen.t5 validate constraint chk_gen_clause; +select locktype, mode from pg_locks + where relation = 'testgen.t5'::regclass and granted; + locktype | mode +----------+-------------------------- + relation | ShareUpdateExclusiveLock +(1 row) + +commit; +-- now the schema update, which skips the rewrite because of the check +begin; +drop trigger testgen_gen on testgen.t5; +alter table testgen.t5 alter column b + add generated always as (a * 2) stored; +select locktype, mode from pg_locks +where relation = 'testgen.t5'::regclass and granted; + locktype | mode +----------+--------------------- + relation | AccessShareLock + relation | AccessExclusiveLock +(2 rows) + +commit; +select pg_relation_filenode('testgen.t5') as t5_filenode_after \gset +select :t5_filenode_before = :t5_filenode_after as did_skip_rewrite; + did_skip_rewrite +------------------ + t +(1 row) + +\d+ testgen.t5 + Table "testgen.t5" + Column | Type | Collation | Nullable | Default | Storage | Stats target | Description +--------+---------+-----------+----------+------------------------------------+---------+--------------+------------- + a | integer | | | | plain | | + b | integer | | | generated always as (a * 2) stored | plain | | +Check constraints: + "chk_gen_clause" CHECK (b = (a * 2)) + +-- test support for partitioned tables and inheritance +create table testgen.tpart (a int, b int) partition by hash (a); +create table testgen.tpart_p1 partition of testgen.tpart + for values with (modulus 2, remainder 0); +create table testgen.tpart_p2 partition of testgen.tpart + for values with (modulus 2, remainder 1); +insert into testgen.tpart (a, b) select x, x from generate_series(1, 5) x; +-- altering the parent table, recursing +begin; +alter table testgen.tpart alter column b + add generated always as (a * 2) stored; +-- expected: all the partitions have been rewritten +select a, b, a * 2 as expected, b = (a * 2) as correct + from testgen.tpart_p1 order by a; + a | b | expected | correct +---+---+----------+--------- + 1 | 2 | 2 | t + 2 | 4 | 4 | t +(2 rows) + +select a, b, a * 2 as expected, b = (a * 2) as correct + from testgen.tpart_p2 order by a; + a | b | expected | correct +---+----+----------+--------- + 3 | 6 | 6 | t + 4 | 8 | 8 | t + 5 | 10 | 10 | t +(3 rows) + +rollback; +-- altering a single partition is not allowed +begin; +-- expected: error +alter table testgen.tpart_p1 alter column b + add generated always as (a * 2) stored; +ERROR: cannot change inherited column to be a stored generated column +rollback; +-- altering only the parent table is not allowed +begin; +-- expected: error +alter table only testgen.tpart alter column b + add generated always as (a * 2) stored; +ERROR: ALTER TABLE / ADD GENERATED ALWAYS AS (expr) STORED must be applied to child tables too +rollback; +drop table testgen.tpart; +-- subpartitions +create table testgen.tpart (a int, b int, c int) + partition by hash (a); +create table testgen.tpart_p1 partition of testgen.tpart + for values with (modulus 2, remainder 0) + partition by hash (b); +create table testgen.tpart_p1_1 partition of testgen.tpart_p1 + for values with (modulus 2, remainder 0); +create table testgen.tpart_p1_2 partition of testgen.tpart_p1 + for values with (modulus 2, remainder 1); +create table testgen.tpart_p2 partition of testgen.tpart + for values with (modulus 2, remainder 1) + partition by hash (b); +create table testgen.tpart_p2_1 partition of testgen.tpart_p2 + for values with (modulus 2, remainder 0); +create table testgen.tpart_p2_2 partition of testgen.tpart_p2 + for values with (modulus 2, remainder 1); +insert into testgen.tpart (a, b) + select x, y + from generate_series(1, 5) x + cross join generate_series(1, 5) y; +-- currently, it is not possible to change the generated state of an +-- inheritance tree of depth >= 2 (same as in DROP EXPRESSION), so we expect an +-- error here. This might be fixed later. +begin; +alter table testgen.tpart alter column c + add generated always as (a + b) stored; +ERROR: ALTER TABLE / ADD GENERATED ALWAYS AS (expr) STORED must be applied to child tables too +rollback; +drop table testgen.tpart; -- tests for invalid invocations alter table doesnotexist alter column foo add generated always as (bar * 2) stored; @@ -4986,6 +5213,7 @@ alter table testgen.t3 alter column b ERROR: cannot use subquery in column generation expression drop table testgen.t3; drop schema testgen cascade; -NOTICE: drop cascades to 4 other objects -DETAIL: drop cascades to table testgen.t3 +NOTICE: drop cascades to 3 other objects +DETAIL: drop cascades to table testgen.t5 +drop cascades to function testgen.gen() drop cascades to table testgen.t1 diff --git a/src/test/regress/sql/alter_table.sql b/src/test/regress/sql/alter_table.sql index 76187083289..efb08b9ff70 100644 --- a/src/test/regress/sql/alter_table.sql +++ b/src/test/regress/sql/alter_table.sql @@ -3198,6 +3198,153 @@ select pg_relation_filenode('testgen.idx_b') as idx_filenode_after \gset select :idx_filenode_before != :idx_filenode_after as did_rewrite_idx; drop table testgen.t3; +-- turning a regular column into a stored generated column +-- without rewriting the table (when a check constraint proves it isn't needed) +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b = a * 2); +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 2) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before = :t4_filenode_after as did_skip_rewrite; +\d+ testgen.t4 +drop table testgen.t4; + +-- turning a regular column into a stored generated column +-- same as the previous case, but a rewrite happens since the constraint is not +-- valid +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b = a * 2) not valid; +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 2) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before != :t4_filenode_after as did_rewrite; +\d+ testgen.t4 +drop table testgen.t4; + +-- turning a regular column into a stored generated column +-- same as the previous case, but a rewrite happens since the constraint +-- operator is not mergejoinable +create table testgen.t4 (a int, b int not null); +insert into testgen.t4 (a, b) select x, x * 2 from generate_series(0, 5) x; +alter table testgen.t4 add constraint chk_gen_clause check (b >= a * 2); +select pg_relation_filenode('testgen.t4') as t4_filenode_before \gset +alter table testgen.t4 alter column b add generated always as (a * 3) stored; +select pg_relation_filenode('testgen.t4') as t4_filenode_after \gset +select :t4_filenode_before != :t4_filenode_after as did_rewrite; +\d+ testgen.t4 +drop table testgen.t4; + +-- test the whole process for adding a stored generated column without +-- long-lived exclusive locks +create table testgen.t5 (a int); +select pg_relation_filenode('testgen.t5') as t5_filenode_before \gset +insert into testgen.t5 select x from generate_series(1, 5) x; +alter table testgen.t5 add column b int; +-- take care of new and updated columns +create function testgen.gen () returns trigger language plpgsql as $$ +begin + new.b = new.a * 2; return new; +end +$$; +create trigger testgen_gen + before insert or update on testgen.t5 + for each row execute function testgen.gen(); +-- add the constraint as not valid: enforced only for new and updated rows +begin; +alter table testgen.t5 + add constraint chk_gen_clause check (b = a * 2) not valid; +select locktype, mode from pg_locks + where relation = 'testgen.t5'::regclass and granted; +commit; +insert into testgen.t5 (a) values (100), (200), (300); +-- backfill existing rows at the appropriate pace +update testgen.t5 set b = a * 2 where b is null; +-- validate: this scans the table, but without an exclusive lock +begin; +alter table testgen.t5 validate constraint chk_gen_clause; +select locktype, mode from pg_locks + where relation = 'testgen.t5'::regclass and granted; +commit; +-- now the schema update, which skips the rewrite because of the check +begin; +drop trigger testgen_gen on testgen.t5; +alter table testgen.t5 alter column b + add generated always as (a * 2) stored; +select locktype, mode from pg_locks +where relation = 'testgen.t5'::regclass and granted; +commit; +select pg_relation_filenode('testgen.t5') as t5_filenode_after \gset +select :t5_filenode_before = :t5_filenode_after as did_skip_rewrite; +\d+ testgen.t5 + +-- test support for partitioned tables and inheritance +create table testgen.tpart (a int, b int) partition by hash (a); +create table testgen.tpart_p1 partition of testgen.tpart + for values with (modulus 2, remainder 0); +create table testgen.tpart_p2 partition of testgen.tpart + for values with (modulus 2, remainder 1); +insert into testgen.tpart (a, b) select x, x from generate_series(1, 5) x; + +-- altering the parent table, recursing +begin; +alter table testgen.tpart alter column b + add generated always as (a * 2) stored; +-- expected: all the partitions have been rewritten +select a, b, a * 2 as expected, b = (a * 2) as correct + from testgen.tpart_p1 order by a; +select a, b, a * 2 as expected, b = (a * 2) as correct + from testgen.tpart_p2 order by a; +rollback; + +-- altering a single partition is not allowed +begin; +-- expected: error +alter table testgen.tpart_p1 alter column b + add generated always as (a * 2) stored; +rollback; + +-- altering only the parent table is not allowed +begin; +-- expected: error +alter table only testgen.tpart alter column b + add generated always as (a * 2) stored; +rollback; + +drop table testgen.tpart; + +-- subpartitions +create table testgen.tpart (a int, b int, c int) + partition by hash (a); +create table testgen.tpart_p1 partition of testgen.tpart + for values with (modulus 2, remainder 0) + partition by hash (b); +create table testgen.tpart_p1_1 partition of testgen.tpart_p1 + for values with (modulus 2, remainder 0); +create table testgen.tpart_p1_2 partition of testgen.tpart_p1 + for values with (modulus 2, remainder 1); +create table testgen.tpart_p2 partition of testgen.tpart + for values with (modulus 2, remainder 1) + partition by hash (b); +create table testgen.tpart_p2_1 partition of testgen.tpart_p2 + for values with (modulus 2, remainder 0); +create table testgen.tpart_p2_2 partition of testgen.tpart_p2 + for values with (modulus 2, remainder 1); +insert into testgen.tpart (a, b) + select x, y + from generate_series(1, 5) x + cross join generate_series(1, 5) y; +-- currently, it is not possible to change the generated state of an +-- inheritance tree of depth >= 2 (same as in DROP EXPRESSION), so we expect an +-- error here. This might be fixed later. +begin; +alter table testgen.tpart alter column c + add generated always as (a + b) stored; +rollback; + +drop table testgen.tpart; + -- tests for invalid invocations alter table doesnotexist alter column foo add generated always as (bar * 2) stored; -- 2.47.0