DDL: Skip collecting TiFlash status when TiFlash is down (#40872)

close #38484

ddl/ddl_tiflash_api.go

@@ -424,6 +424,14 @@ func (d *ddl) refreshTiFlashTicker(ctx sessionctx.Context, pollTiFlashContext *T
 			return err
 		}
 	}
+
+	failpoint.Inject("OneTiFlashStoreDown", func() {
+		for storeID, store := range pollTiFlashContext.TiFlashStores {
+			store.Store.StateName = "Down"
+			pollTiFlashContext.TiFlashStores[storeID] = store
+			break
+		}
+	})
 	pollTiFlashContext.PollCounter++
 
 	// Start to process every table.

ddl/tiflashtest/ddl_tiflash_test.go

@@ -1334,3 +1334,23 @@ func TestTiFlashAvailableAfterAddPartition(t *testing.T) {
 	require.NotNil(t, pi)
 	require.Equal(t, len(pi.Definitions), 2)
 }
+
+func TestTiFlashAvailableAfterDownOneStore(t *testing.T) {
+	s, teardown := createTiFlashContext(t)
+	defer teardown()
+	tk := testkit.NewTestKit(t, s.store)
+
+	tk.MustExec("use test")
+	tk.MustExec("drop table if exists ddltiflash")
+	tk.MustExec("create table ddltiflash(z int) PARTITION BY RANGE(z) (PARTITION p0 VALUES LESS THAN (10))")
+	require.NoError(t, failpoint.Enable("github.com/pingcap/tidb/ddl/OneTiFlashStoreDown", `return`))
+	require.NoError(t, failpoint.Enable("github.com/pingcap/tidb/domain/infosync/OneTiFlashStoreDown", `return`))
+	defer func() {
+		require.NoError(t, failpoint.Disable("github.com/pingcap/tidb/ddl/OneTiFlashStoreDown"))
+		require.NoError(t, failpoint.Disable("github.com/pingcap/tidb/domain/infosync/OneTiFlashStoreDown"))
+	}()
+
+	tk.MustExec("alter table ddltiflash set tiflash replica 1")
+	time.Sleep(ddl.PollTiFlashInterval * RoundToBeAvailable * 3)
+	CheckTableAvailable(s.dom, t, 1, []string{})
+}

domain/infosync/tiflash_manager.go

@@ -31,6 +31,7 @@ import (
 
 	"github.com/gorilla/mux"
 	"github.com/pingcap/errors"
+	"github.com/pingcap/failpoint"
 	"github.com/pingcap/tidb/ddl/placement"
 	"github.com/pingcap/tidb/store/helper"
 	"github.com/pingcap/tidb/tablecodec"
@@ -89,10 +90,19 @@ func getTiFlashPeerWithoutLagCount(tiFlashStores map[int64]helper.StoreStat, tab
 	for _, store := range tiFlashStores {
 		regionReplica := make(map[int64]int)
 		err := helper.CollectTiFlashStatus(store.Store.StatusAddress, tableID, &regionReplica)
+		failpoint.Inject("OneTiFlashStoreDown", func() {
+			if store.Store.StateName == "Down" {
+				err = errors.New("mock TiFlasah down")
+			}
+		})
 		if err != nil {
 			logutil.BgLogger().Error("Fail to get peer status from TiFlash.",
 				zap.Int64("tableID", tableID))
-			return 0, err
+			// Just skip down or offline or tomestone stores, because PD will migrate regions from these stores.
+			if store.Store.StateName == "Up" || store.Store.StateName == "Disconnected" {
+				return 0, err
+			}
+			continue
 		}
 		flashPeerCount += len(regionReplica)
 	}