From 3bafa7b25b507646aab3d0b39528236f6ff05b36 Mon Sep 17 00:00:00 2001
From: zghh <1069308575@qq.com>
Date: Thu, 11 Aug 2022 20:03:06 +0800
Subject: [PATCH] Fix inconsistent state between WAL and saved Snapshot, and
 add the unit test to reproduce the problem.

---
 orderer/consensus/etcdraft/storage_test.go | 79 ++++++++++++++++++++++
 1 file changed, 79 insertions(+)

diff --git a/orderer/consensus/etcdraft/storage_test.go b/orderer/consensus/etcdraft/storage_test.go
index 3668714594c..89a0b26d89b 100644
--- a/orderer/consensus/etcdraft/storage_test.go
+++ b/orderer/consensus/etcdraft/storage_test.go
@@ -395,3 +395,82 @@ func TestApplyOutOfDateSnapshot(t *testing.T) {
 		assertFileCount(t, 12, 1)
 	})
 }
+
+func TestAbortWhenWritingSnapshot(t *testing.T) {
+	t.Run("Abort when writing snapshot", func(t *testing.T) {
+		setup(t)
+		defer clean(t)
+
+		// set SegmentSizeBytes to a small value so that
+		// every entry persisted to wal would result in
+		// a new wal being created.
+		oldSegmentSizeBytes := wal.SegmentSizeBytes
+		wal.SegmentSizeBytes = 10
+		defer func() {
+			wal.SegmentSizeBytes = oldSegmentSizeBytes
+		}()
+
+		// create 5 new entry
+		for i := 0; i < 5; i++ {
+			store.Store(
+				[]raftpb.Entry{{Index: uint64(i), Data: make([]byte, 100)}},
+				raftpb.HardState{Commit: uint64(i)},
+				raftpb.Snapshot{},
+			)
+		}
+		assertFileCount(t, 6, 0)
+
+		// Assume an orderer missed some records due to exceptions and receives a new snapshot from other orderers.
+		commit := 10
+		store.Store(
+			[]raftpb.Entry{},
+			raftpb.HardState{Commit: uint64(commit)},
+			raftpb.Snapshot{
+				Metadata: raftpb.SnapshotMetadata{
+					Index: uint64(commit),
+				},
+				Data: make([]byte, 100),
+			},
+		)
+		err = store.Close()
+		assert.NoError(t, err)
+
+		// In old logic, it will use rs.wal.Save(hardstate, entries) to save the state firstly, so we remove the snapshot files.
+		// sd, err := os.Open(snapDir)
+		// assert.NoError(t, err)
+		// defer sd.Close()
+		// names, err := sd.Readdirnames(-1)
+		// assert.NoError(t, err)
+		// sort.Sort(sort.Reverse(sort.StringSlice(names)))
+		// os.Remove(filepath.Join(snapDir, names[0]))
+		// wd, err := os.Open(walDir)
+		// assert.NoError(t, err)
+		// defer wd.Close()
+		// names, err = wd.Readdirnames(-1)
+		// assert.NoError(t, err)
+		// sort.Sort(sort.Reverse(sort.StringSlice(names)))
+		// os.Remove(filepath.Join(walDir, names[0]))
+
+		// But in the new logic, it will use rs.saveSnap(snapshot) to save the snapshot firstly, so we remove the WAL files.
+		wd, err := os.Open(walDir)
+		assert.NoError(t, err)
+		defer wd.Close()
+		names, err := wd.Readdirnames(-1)
+		assert.NoError(t, err)
+		sort.Sort(sort.Reverse(sort.StringSlice(names)))
+		os.Remove(filepath.Join(walDir, names[0]))
+
+		// Then restart the orderer.
+		ram := raft.NewMemoryStorage()
+		store, err = CreateStorage(logger, walDir, snapDir, ram)
+		assert.NoError(t, err)
+
+		// Check the state from go.etcd.io/etcd/raft/raft.go
+		// func (r *raft) loadState(state pb.HardState)
+		hd, _, err := store.ram.InitialState()
+		assert.NoError(t, err)
+		lastIndex, err := store.ram.LastIndex()
+		assert.NoError(t, err)
+		assert.False(t, hd.Commit > lastIndex)
+	})
+}