From 1ff1342283ba1f8db3d634d94bce45562ac3b2d9 Mon Sep 17 00:00:00 2001
From: tangenta <tangenta@126.com>
Date: Sat, 10 Dec 2022 12:05:41 +0800
Subject: [PATCH] ddl: mark the writes from delete-only and drop them on merge
 (#39796)

close pingcap/tidb#39731
---
 ddl/index_merge_tmp.go      |  5 +++--
 ddl/index_merge_tmp_test.go | 45 +++++++++++++++++++++++++++++++++++++
 table/tables/index.go       |  7 +++++-
 3 files changed, 54 insertions(+), 3 deletions(-)

diff --git a/ddl/index_merge_tmp.go b/ddl/index_merge_tmp.go
index b33111119facc..8f430ccce355f 100644
--- a/ddl/index_merge_tmp.go
+++ b/ddl/index_merge_tmp.go
@@ -186,8 +186,9 @@ func (w *mergeIndexWorker) fetchTempIndexVals(txn kv.Transaction, taskRange reor
 			unique := false
 			length := len(rawValue)
 			keyVer := rawValue[length-1]
-			if keyVer == tables.TempIndexKeyTypeMerge {
-				// The kv is written in the merging state. It has been written to the origin index, we can skip it.
+			if keyVer == tables.TempIndexKeyTypeMerge || keyVer == tables.TempIndexKeyTypeDelete {
+				// For 'm' version kvs, they are double-written.
+				// For 'd' version kvs, they are written in the delete-only state and can be dropped safely.
 				return true, nil
 			}
 			rawValue = rawValue[:length-1]
diff --git a/ddl/index_merge_tmp_test.go b/ddl/index_merge_tmp_test.go
index 389339ac15ad4..431e4791f4aeb 100644
--- a/ddl/index_merge_tmp_test.go
+++ b/ddl/index_merge_tmp_test.go
@@ -331,3 +331,48 @@ func TestCreateUniqueIndexKeyExist(t *testing.T) {
 	tk.MustExec("admin check table t")
 	tk.MustQuery("select * from t order by a, b").Check(testkit.Rows("0 9", "1 7", "2 7", "5 7", "8 8", "10 10"))
 }
+
+func TestAddIndexMergeIndexUpdateOnDeleteOnly(t *testing.T) {
+	store, dom := testkit.CreateMockStoreAndDomain(t)
+	tk := testkit.NewTestKit(t, store)
+	tk.MustExec("use test")
+	tk2 := testkit.NewTestKit(t, store)
+	tk2.MustExec("use test")
+	tk.MustExec(`CREATE TABLE t (a DATE NULL DEFAULT '1619-01-18', b BOOL NULL DEFAULT '0') CHARACTER SET 'utf8mb4' COLLATE 'utf8mb4_bin';`)
+	tk.MustExec(`INSERT INTO t SET b = '1';`)
+
+	updateSQLs := []string{
+		"UPDATE t SET a = '9432-05-10', b = '0';",
+		"UPDATE t SET a = '9432-05-10', b = '1';",
+	}
+
+	// Force onCreateIndex use the txn-merge process.
+	ingest.LitInitialized = false
+	tk.MustExec("set @@global.tidb_ddl_enable_fast_reorg = 1;")
+	tk.MustExec("set @@global.tidb_enable_mutation_checker = 1;")
+	tk.MustExec("set @@global.tidb_txn_assertion_level = 'STRICT';")
+
+	var checkErrs []error
+	originHook := dom.DDL().GetHook()
+	callback := &ddl.TestDDLCallback{
+		Do: dom,
+	}
+	onJobUpdatedBefore := func(job *model.Job) {
+		if job.SchemaState == model.StateDeleteOnly {
+			for _, sql := range updateSQLs {
+				_, err := tk2.Exec(sql)
+				if err != nil {
+					checkErrs = append(checkErrs, err)
+				}
+			}
+		}
+	}
+	callback.OnJobUpdatedExported.Store(&onJobUpdatedBefore)
+	dom.DDL().SetHook(callback)
+	tk.MustExec("alter table t add index idx(b);")
+	dom.DDL().SetHook(originHook)
+	for _, err := range checkErrs {
+		require.NoError(t, err)
+	}
+	tk.MustExec("admin check table t;")
+}
diff --git a/table/tables/index.go b/table/tables/index.go
index b3a481efba29f..1193f9e251758 100644
--- a/table/tables/index.go
+++ b/table/tables/index.go
@@ -130,7 +130,7 @@ func (c *index) Create(sctx sessionctx.Context, txn kv.Transaction, indexedValue
 	)
 	if !opt.FromBackFill {
 		key, tempKey, keyVer = GenTempIdxKeyByState(c.idxInfo, key)
-		if keyVer == TempIndexKeyTypeBackfill {
+		if keyVer == TempIndexKeyTypeBackfill || keyVer == TempIndexKeyTypeDelete {
 			key, tempKey = tempKey, nil
 			keyIsTempIdxKey = true
 		}
@@ -347,6 +347,8 @@ func (c *index) Delete(sc *stmtctx.StatementContext, txn kv.Transaction, indexed
 const (
 	// TempIndexKeyTypeNone means the key is not a temporary index key.
 	TempIndexKeyTypeNone byte = 0
+	// TempIndexKeyTypeDelete indicates this value is written in the delete-only stage.
+	TempIndexKeyTypeDelete byte = 'd'
 	// TempIndexKeyTypeBackfill indicates this value is written in the backfill stage.
 	TempIndexKeyTypeBackfill byte = 'b'
 	// TempIndexKeyTypeMerge indicates this value is written in the merge stage.
@@ -363,6 +365,9 @@ func GenTempIdxKeyByState(indexInfo *model.IndexInfo, indexKey kv.Key) (key, tem
 		case model.BackfillStateRunning:
 			// Write to the temporary index.
 			tablecodec.IndexKey2TempIndexKey(indexInfo.ID, indexKey)
+			if indexInfo.State == model.StateDeleteOnly {
+				return nil, indexKey, TempIndexKeyTypeDelete
+			}
 			return nil, indexKey, TempIndexKeyTypeBackfill
 		case model.BackfillStateReadyToMerge, model.BackfillStateMerging:
 			// Double write