sql: ALTER PRIMARY KEY rewrite secondary index if newPK subsets oldPK

Previously, during a `ALTER PRIMARY KEY`, if the new PK columns is a
subset of the old PK columns, we won't rewrite existing unique,
secondary index.

This was inadequate because the user might think this column is not used
anywhere and will want to drop it, which will unexpectedly drop the
dependent unique index.

Release note (bug fix): This PR fixed a bug where, in a
`ALTER PRIMARY KEY`, if the new PK columns is a subset of the old PK
columns, we will not rewrite existing secondary index, and hence those
secondary indexes continue to have some of the old PK columns in their
`suffixColumns`. But the user might, reasonably, think those columns
are not used anymore and proceed to drop them. The bug then caused
those dependent secondary indexes to be dropped, unexpectedly for
the user.

Release justification: fixed a bug where unique indexes will be
silently dropped after `ALTER PRIMARY KEY` followed by `DROP COLUMN`.

pkg/sql/alter_primary_key.go

@@ -438,6 +438,11 @@ func (p *planner) AlterPrimaryKey(
 				return true, nil
 			}
 		}
+		if !idx.Primary() && catalog.MakeTableColSet(newPrimaryIndexDesc.KeyColumnIDs...).SubsetOf(
+			catalog.MakeTableColSet(tableDesc.PrimaryIndex.KeyColumnIDs...)) {
+			// Always rewrite a secondary index if the new PK columns is a (strict) subset of the old PK columns.
+			return true, nil
+		}
 		if idx.IsUnique() {
 			for i := 0; i < idx.NumKeyColumns(); i++ {
 				colID := idx.GetKeyColumnID(i)

pkg/sql/logictest/testdata/logic_test/alter_primary_key

@@ -1056,9 +1056,9 @@ primary        id     DESC
 primary        id2    N/A
 primary        name   N/A
 primary        rowid  N/A
+t1_id_key      id     ASC
 t1_id_id2_key  id     ASC
 t1_id_id2_key  id2    ASC
-t1_id_key      id     ASC
 
 
 statement ok
@@ -1071,9 +1071,9 @@ primary        id     DESC
 primary        id2    N/A
 primary        name   N/A
 primary        rowid  N/A
+t1_id_key      id     ASC
 t1_id_id2_key  id     ASC
 t1_id_id2_key  id2    ASC
-t1_id_key      id     ASC
 
 statement ok
 alter table t1 alter primary key using columns(id desc);
@@ -1085,9 +1085,9 @@ primary        id     DESC
 primary        id2    N/A
 primary        name   N/A
 primary        rowid  N/A
+t1_id_key      id     ASC
 t1_id_id2_key  id     ASC
 t1_id_id2_key  id2    ASC
-t1_id_key      id     ASC
 
 statement ok
 alter table t1 alter primary key using columns(id) USING HASH WITH BUCKET_COUNT = 10
@@ -1102,11 +1102,11 @@ primary        name                       N/A
 primary        rowid                      N/A
 t1_id_key1     id                         DESC
 t1_id_key1     crdb_internal_id_shard_10  ASC
+t1_id_key      id                         ASC
+t1_id_key      crdb_internal_id_shard_10  ASC
 t1_id_id2_key  id                         ASC
 t1_id_id2_key  id2                        ASC
 t1_id_id2_key  crdb_internal_id_shard_10  ASC
-t1_id_key      id                         ASC
-t1_id_key      crdb_internal_id_shard_10  ASC
 
 statement ok
 CREATE TABLE table_with_virtual_cols (
@@ -1192,3 +1192,76 @@ query II
 SELECT * FROM t71553@t71553_b_key
 ----
 1  1
+
+# The following regression test makes sure that when the new PK columns
+# is a (strict) subset of the old PK columns, all existing secondary indexes
+# were rewritten, and hence dropping a column from the old PK columns does not
+# unexpectedly drop an existing secondary index.
+subtest regression_#84040
+
+statement ok
+DROP TABLE IF EXISTS t
+
+statement ok
+CREATE TABLE t (
+  a INT NOT NULL,
+  b INT NOT NULL,
+  c INT NOT NULL,
+  PRIMARY KEY (a, b),
+  UNIQUE INDEX uniq_idx (c)
+);
+
+statement ok
+ALTER TABLE t ALTER PRIMARY KEY USING COLUMNS (a)
+
+statement ok
+ALTER TABLE t DROP COLUMN b
+
+query TTT
+select index_name,column_name,direction from [show indexes from t];
+----
+primary   a  ASC
+primary   c  N/A
+uniq_idx  c  ASC
+uniq_idx  a  ASC
+
+# Alter primary key on a hash-sharded primary key to be non-hash-sharded.
+# This had unexpectedly caused all unique indexes to be dropped silently on v21.2.
+# See the support issue https://github.com/cockroachlabs/support/issues/1687
+statement ok
+DROP TABLE IF EXISTS t;
+
+statement ok
+set experimental_enable_hash_sharded_indexes = true;
+
+statement ok
+CREATE TABLE t (
+	i INT PRIMARY KEY USING HASH WITH bucket_count=7 DEFAULT unique_rowid(),
+	j INT NOT NULL UNIQUE
+)
+
+# Assert that the primary key is hash-sharded and the unique index is created.
+query TTT
+SELECT index_name,column_name,direction FROM [SHOW INDEXES FROM t]
+----
+primary  crdb_internal_i_shard_7  ASC
+primary  i                        ASC
+primary  j                        N/A
+t_j_key  j                        ASC
+t_j_key  crdb_internal_i_shard_7  ASC
+t_j_key  i                        ASC
+
+# Alter the primary key to be no longer hash-sharded.
+statement ok
+ALTER TABLE t ALTER PRIMARY KEY USING COLUMNS (i)
+
+# Now assert that the primary key has been modified to non-hash-sharded,
+# and the unique index still exists.
+query TTT
+SELECT index_name,column_name,direction FROM [SHOW INDEXES FROM t]
+----
+primary  i                        ASC
+primary  crdb_internal_i_shard_7  N/A
+primary  j                        N/A
+t_j_key  j                        ASC
+t_j_key  i                        ASC